From 74f5b14d18de3f14278eb6cc1488e9a36b0aa05e Mon Sep 17 00:00:00 2001
From: BFlat <bombayapirat@gmail.com>
Date: Wed, 5 Mar 2025 10:23:48 +0700
Subject: [PATCH 1/8] added loading indicator text (#439)

added signal for loading indicator text change
---
 .../SignaledLoadingIndicator.tsx                  |  4 +++-
 .../src/signals/uiEvents/loadingIndicator.ts      | 15 +++++++++++++++
 2 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/frontend/realtime/src/KonvaRootComponent/SignaledLoadingIndicator.tsx b/frontend/realtime/src/KonvaRootComponent/SignaledLoadingIndicator.tsx
index f67ec2a66d..d2ea3b72c1 100644
--- a/frontend/realtime/src/KonvaRootComponent/SignaledLoadingIndicator.tsx
+++ b/frontend/realtime/src/KonvaRootComponent/SignaledLoadingIndicator.tsx
@@ -3,6 +3,7 @@ import { LoadingBar, LoadingBarStatus } from "~/components/ui";
 import {
   isLoadingVisible,
   loadingProgress,
+  loadingText,
 } from "~/signals/uiEvents/loadingIndicator";
 
 export const SignaledLoadingIndicator = () => {
@@ -10,6 +11,7 @@ export const SignaledLoadingIndicator = () => {
   useSignalEffect(() => {
     isLoadingVisible.value;
     loadingProgress.value;
+    loadingText.value;
   });
 
   if (!isLoadingVisible.value) {
@@ -47,7 +49,7 @@ export const SignaledLoadingIndicator = () => {
             className="text-white/90"
           ></path>
         </svg>
-        Downloading models - {loadingProgress.value}%
+        {loadingText.value} - {loadingProgress.value}%
       </div>
     </div>
   );
diff --git a/frontend/realtime/src/signals/uiEvents/loadingIndicator.ts b/frontend/realtime/src/signals/uiEvents/loadingIndicator.ts
index 2a5ceae646..23b49d59c1 100644
--- a/frontend/realtime/src/signals/uiEvents/loadingIndicator.ts
+++ b/frontend/realtime/src/signals/uiEvents/loadingIndicator.ts
@@ -28,13 +28,28 @@ const onLoadingProgressChanged = (callback: (progress: number) => void) => {
   });
 };
 
+// Loading text signal
+export const DEFAULT_LOADING_TEXT = "Downloading";
+export const loadingText = signal<string>(DEFAULT_LOADING_TEXT);
+const setLoadingText = (text: string) => {
+  loadingText.value = text;
+};
+
+const onLoadingTextChanged = (callback: (text: string) => void) => {
+  effect(() => {
+    callback(loadingText.value);
+  });
+};
+
 // EXPORTS
 export const dispatchers = {
   setIsLoadingVisible,
   setLoadingProgress,
+  setLoadingText,
 };
 
 export const events = {
   onIsLoadingVisibleChanged,
   onLoadingProgressChanged,
+  onLoadingTextChanged,
 };

From 4ab6ca1caaa3b7afce9bfab8a8d41954fff4b34b Mon Sep 17 00:00:00 2001
From: bflatastic <bombayapirat@gmail.com>
Date: Wed, 5 Mar 2025 23:21:11 +0700
Subject: [PATCH 2/8] disabled annoying focus ring

---
 frontend/realtime/src/components/features/ToolbarButton.tsx | 2 +-
 frontend/realtime/src/components/ui/Button.tsx              | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/frontend/realtime/src/components/features/ToolbarButton.tsx b/frontend/realtime/src/components/features/ToolbarButton.tsx
index a55724f2e3..e8694ccf03 100644
--- a/frontend/realtime/src/components/features/ToolbarButton.tsx
+++ b/frontend/realtime/src/components/features/ToolbarButton.tsx
@@ -44,7 +44,7 @@ export const ToolbarButton = ({
   }
 
   const mergedButtonClasses = twMerge(
-    "rounded-lg hover:bg-white/15 transition-all duration-100 border-2 border-transparent text-white/80 px-2",
+    "rounded-lg hover:bg-white/15 transition-all duration-100 border-2 border-transparent text-white/80 px-2 focus-visible:ring-0 focus-visible:ring-offset-0",
     children ? "w-fit flex items-center gap-2.5 text-nowrap" : "size-10",
     active && "bg-primary/30 border-2 border-primary hover:bg-primary/30",
     disabled && "pointer-events-none text-secondary-300",
diff --git a/frontend/realtime/src/components/ui/Button.tsx b/frontend/realtime/src/components/ui/Button.tsx
index 5e3deb69f6..be875daaf2 100644
--- a/frontend/realtime/src/components/ui/Button.tsx
+++ b/frontend/realtime/src/components/ui/Button.tsx
@@ -51,7 +51,7 @@ export const Button = ({
     "text-[15px] font-semibold rounded-lg px-3 py-2 shadow-sm",
     "flex gap-2 items-center justify-center",
     "transition-all duration-150",
-    "focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2",
+    "focus-visible:ring-0 focus-visible:ring-offset-0",
     getVariantClassNames(propsVariant),
     propsClassName,
     disabledClass,

From 3556dcbfd3a33d1b7bd5bf45974183a231801d19 Mon Sep 17 00:00:00 2001
From: bflatastic <bombayapirat@gmail.com>
Date: Wed, 5 Mar 2025 23:40:33 +0700
Subject: [PATCH 3/8] removed annoying focus on buttons 2

---
 frontend/realtime/src/app/global.css              | 15 +++++++++++++++
 .../src/components/features/ToolbarButton.tsx     |  8 +++++++-
 frontend/realtime/src/components/ui/Button.tsx    |  9 ++++-----
 3 files changed, 26 insertions(+), 6 deletions(-)

diff --git a/frontend/realtime/src/app/global.css b/frontend/realtime/src/app/global.css
index 077379a221..3c0abd0ee7 100644
--- a/frontend/realtime/src/app/global.css
+++ b/frontend/realtime/src/app/global.css
@@ -6,6 +6,21 @@
   :root {
     @apply select-none bg-ui-background font-sans text-white;
   }
+
+  /* Remove all focus styles */
+  button {
+    @apply outline-none ring-0;
+    -webkit-tap-highlight-color: transparent !important;
+    outline: none !important;
+    box-shadow: none !important;
+  }
+
+  button:focus {
+    @apply outline-none ring-0;
+    outline: none !important;
+    box-shadow: none !important;
+  }
+
   .pegboard {
     background-image: radial-gradient(#ffffff25 1px, transparent 0);
     background-color: #1b1b1f;
diff --git a/frontend/realtime/src/components/features/ToolbarButton.tsx b/frontend/realtime/src/components/features/ToolbarButton.tsx
index e8694ccf03..fadc860c35 100644
--- a/frontend/realtime/src/components/features/ToolbarButton.tsx
+++ b/frontend/realtime/src/components/features/ToolbarButton.tsx
@@ -44,7 +44,7 @@ export const ToolbarButton = ({
   }
 
   const mergedButtonClasses = twMerge(
-    "rounded-lg hover:bg-white/15 transition-all duration-100 border-2 border-transparent text-white/80 px-2 focus-visible:ring-0 focus-visible:ring-offset-0",
+    "rounded-lg hover:bg-white/15 transition-all duration-100 border-2 border-transparent text-white/80 px-2",
     children ? "w-fit flex items-center gap-2.5 text-nowrap" : "size-10",
     active && "bg-primary/30 border-2 border-primary hover:bg-primary/30",
     disabled && "pointer-events-none text-secondary-300",
@@ -58,6 +58,12 @@ export const ToolbarButton = ({
       className={mergedButtonClasses}
       disabled={disabled}
       autoFocus={false}
+      tabIndex={-1}
+      style={{
+        outline: "none",
+        boxShadow: "none",
+        WebkitTapHighlightColor: "transparent",
+      }}
       {...restButtonProps}
       onClick={(e) => {
         e.preventDefault();
diff --git a/frontend/realtime/src/components/ui/Button.tsx b/frontend/realtime/src/components/ui/Button.tsx
index be875daaf2..fe83115ce5 100644
--- a/frontend/realtime/src/components/ui/Button.tsx
+++ b/frontend/realtime/src/components/ui/Button.tsx
@@ -28,17 +28,17 @@ export const Button = ({
   function getVariantClassNames(variant: string) {
     switch (variant) {
       case "tertiary": {
-        return "bg-tertiary hover:bg-tertiary-400 text-white focus-visible:outline-tertiary-600";
+        return "bg-tertiary hover:bg-tertiary-400 text-white";
       }
       case "secondary": {
-        return "bg-ui-controls hover:bg-ui-controls/80 text-white focus-visible:outline-secondary";
+        return "bg-ui-controls hover:bg-ui-controls/80 text-white";
       }
       case "action": {
-        return " bg-action hover:bg-action-500 text-white focus-visible:outline-action";
+        return "bg-action hover:bg-action-500 text-white";
       }
       case "primary":
       default: {
-        return "bg-primary hover:bg-primary-600 text-white focus-visible:outline-primary-600";
+        return "bg-primary hover:bg-primary-600 text-white";
       }
     }
   }
@@ -51,7 +51,6 @@ export const Button = ({
     "text-[15px] font-semibold rounded-lg px-3 py-2 shadow-sm",
     "flex gap-2 items-center justify-center",
     "transition-all duration-150",
-    "focus-visible:ring-0 focus-visible:ring-offset-0",
     getVariantClassNames(propsVariant),
     propsClassName,
     disabledClass,

From ada1f46e626b0d062d153ac9c8b8576bd1fe96ab Mon Sep 17 00:00:00 2001
From: Brandon Thomas <echelon@gmail.com>
Date: Wed, 5 Mar 2025 18:24:05 -0500
Subject: [PATCH 4/8] model downloading (wip)

---
 Cargo.lock                                    | 90 +++++++++++--------
 crates/desktop/src-tauri/Cargo.toml           |  1 +
 .../src/endpoints/download_models.rs          | 17 ++++
 .../src-tauri/src/endpoints/image_endpoint.rs | 18 ++--
 crates/desktop/src-tauri/src/endpoints/mod.rs |  1 +
 crates/desktop/src-tauri/src/lib.rs           |  2 +
 .../src/ml/downloads/download_all_models.rs   |  8 ++
 .../downloads/download_model_registry_file.rs | 21 +++++
 .../desktop/src-tauri/src/ml/downloads/mod.rs |  2 +
 crates/desktop/src-tauri/src/ml/mod.rs        |  4 +-
 .../src-tauri/src/ml/model_registry.rs        | 25 ++++++
 .../infer_clip_text_embeddings.rs             | 13 ++-
 .../stable_diffusion_pipeline.rs              |  5 +-
 13 files changed, 160 insertions(+), 47 deletions(-)
 create mode 100644 crates/desktop/src-tauri/src/endpoints/download_models.rs
 create mode 100644 crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
 create mode 100644 crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs
 create mode 100644 crates/desktop/src-tauri/src/ml/downloads/mod.rs
 create mode 100644 crates/desktop/src-tauri/src/ml/model_registry.rs

diff --git a/Cargo.lock b/Cargo.lock
index 820da85465..9702fccefc 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -672,6 +672,7 @@ dependencies = [
  "memory_caching",
  "once_cell",
  "rand 0.8.5",
+ "reqwest 0.12.12",
  "serde",
  "serde_json",
  "serde_yaml",
@@ -1079,7 +1080,7 @@ dependencies = [
  "rustversion",
  "serde",
  "sync_wrapper 0.1.2",
- "tower",
+ "tower 0.4.13",
  "tower-layer",
  "tower-service",
 ]
@@ -3316,7 +3317,7 @@ dependencies = [
  "google_drive_common",
  "hashing",
  "hostname 0.4.0",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "jobs_common",
  "log",
  "mimetypes",
@@ -4762,7 +4763,7 @@ dependencies = [
  "errors",
  "jwt-simple",
  "my-workspace-hack",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde",
  "serde_derive",
  "serde_json",
@@ -4988,7 +4989,7 @@ dependencies = [
  "native-tls",
  "num_cpus",
  "rand 0.8.5",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde",
  "serde_json",
  "thiserror 2.0.11",
@@ -5192,9 +5193,9 @@ dependencies = [
 
 [[package]]
 name = "httparse"
-version = "1.8.0"
+version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d897f394bad6a705d5f4104762e116a75639e470d80901eed05a860a95cb1904"
+checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
 
 [[package]]
 name = "httpdate"
@@ -5234,9 +5235,9 @@ dependencies = [
 
 [[package]]
 name = "hyper"
-version = "1.3.1"
+version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fe575dd17d0862a9a33781c8c4696a55c320909004a67a00fb286ba8b1bc496d"
+checksum = "cc2b571658e38e0c01b1fdca3bbbe93c00d3d71693ff2770043f8c29bc7d6f80"
 dependencies = [
  "bytes 1.10.0",
  "futures-channel",
@@ -5260,7 +5261,7 @@ checksum = "08afdbb5c31130e3034af566421053ab03787c640246a446327f550d11bcb333"
 dependencies = [
  "futures-util",
  "http 1.1.0",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "hyper-util",
  "rustls 0.23.13",
  "rustls-pki-types",
@@ -5303,7 +5304,7 @@ checksum = "70206fc6890eaca9fde8a0bf71caa2ddfc9fe045ac9e5c70df101a7dbde866e0"
 dependencies = [
  "bytes 1.10.0",
  "http-body-util",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "hyper-util",
  "native-tls",
  "tokio",
@@ -5313,20 +5314,19 @@ dependencies = [
 
 [[package]]
 name = "hyper-util"
-version = "0.1.5"
+version = "0.1.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7b875924a60b96e5d7b9ae7b066540b1dd1cbd90d1828f54c92e02a283351c56"
+checksum = "df2dcfbe0677734ab2f3ffa7fa7bfd4706bfdc1ef393f2ee30184aed67e631b4"
 dependencies = [
  "bytes 1.10.0",
  "futures-channel",
  "futures-util",
  "http 1.1.0",
  "http-body 1.0.0",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "pin-project-lite",
  "socket2",
  "tokio",
- "tower",
  "tower-service",
  "tracing",
 ]
@@ -5563,7 +5563,7 @@ dependencies = [
  "rand 0.8.5",
  "redis_common",
  "redis_schema",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde",
  "serde_derive",
  "serde_json",
@@ -6852,7 +6852,7 @@ dependencies = [
  "heck 0.4.1",
  "hmac 0.12.1",
  "hyper 0.14.28",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "hyper-util",
  "indexmap 2.2.6",
  "itertools 0.12.1",
@@ -6878,7 +6878,7 @@ dependencies = [
  "regex",
  "regex-automata",
  "regex-syntax",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "rsa",
  "rustls-pki-types",
  "semver 1.0.23",
@@ -6904,7 +6904,7 @@ dependencies = [
  "tokio",
  "tokio-stream",
  "tokio-util",
- "tower",
+ "tower 0.4.13",
  "tracing",
  "unicode-bidi",
  "unicode-normalization",
@@ -9269,9 +9269,9 @@ dependencies = [
 
 [[package]]
 name = "reqwest"
-version = "0.12.7"
+version = "0.12.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f8f4955649ef5c38cc7f9e8aa41761d48fb9677197daea9984dc54f56aad5e63"
+checksum = "43e734407157c3c2034e0258f5e4473ddb361b1e85f95a66690d67264d7cd1da"
 dependencies = [
  "base64 0.22.1",
  "bytes 1.10.0",
@@ -9283,7 +9283,7 @@ dependencies = [
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "hyper-rustls",
  "hyper-tls 0.6.0",
  "hyper-util",
@@ -9308,6 +9308,7 @@ dependencies = [
  "tokio-native-tls",
  "tokio-rustls",
  "tokio-util",
+ "tower 0.5.2",
  "tower-service",
  "url",
  "wasm-bindgen",
@@ -9502,7 +9503,7 @@ dependencies = [
  "minidom",
  "percent-encoding",
  "quick-xml 0.36.2",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde",
  "serde_derive",
  "serde_json",
@@ -10833,7 +10834,7 @@ dependencies = [
  "redis_common",
  "redis_schema",
  "regex",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "reusable_types",
  "rust-s3",
  "serde",
@@ -11597,7 +11598,7 @@ dependencies = [
  "percent-encoding",
  "plist",
  "raw-window-handle",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde",
  "serde_json",
  "serde_repr",
@@ -11945,7 +11946,7 @@ dependencies = [
  "easyenv",
  "log",
  "my-workspace-hack",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde_json",
 ]
 
@@ -12399,7 +12400,7 @@ dependencies = [
  "prost",
  "tokio",
  "tokio-stream",
- "tower",
+ "tower 0.4.13",
  "tower-layer",
  "tower-service",
  "tracing",
@@ -12435,16 +12436,31 @@ dependencies = [
 ]
 
 [[package]]
-name = "tower-layer"
-version = "0.3.2"
+name = "tower"
+version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c20c8dbed6283a09604c3e69b4b7eeb54e298b8a600d4d5ecb5ad39de609f1d0"
+checksum = "d039ad9159c98b70ecfd540b2573b97f7f52c3e8d9f8ad57a24b916a536975f9"
+dependencies = [
+ "futures-core",
+ "futures-util",
+ "pin-project-lite",
+ "sync_wrapper 1.0.1",
+ "tokio",
+ "tower-layer",
+ "tower-service",
+]
+
+[[package]]
+name = "tower-layer"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
 
 [[package]]
 name = "tower-service"
-version = "0.3.2"
+version = "0.3.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6bc1c9ce2b5135ac7f93c72918fc37feb872bdc6a5533a8b85eb4b86bfdae52"
+checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
 
 [[package]]
 name = "tracing"
@@ -12581,7 +12597,7 @@ dependencies = [
  "google_drive_common",
  "hashing",
  "hostname 0.4.0",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "jobs_common",
  "log",
  "my-workspace-hack",
@@ -12619,7 +12635,7 @@ dependencies = [
  "my-workspace-hack",
  "mysql_queries",
  "r2d2_redis",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "serde",
  "serde_derive",
  "serde_json",
@@ -12979,7 +12995,7 @@ dependencies = [
  "actix-web",
  "mime_guess",
  "regex",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "rust-embed",
  "serde",
  "serde_json",
@@ -13108,7 +13124,7 @@ dependencies = [
  "google_drive_common",
  "hashing",
  "hostname 0.4.0",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "jobs_common",
  "log",
  "my-workspace-hack",
@@ -13138,7 +13154,7 @@ dependencies = [
  "filesys",
  "google_drive_common",
  "hostname 0.4.0",
- "hyper 1.3.1",
+ "hyper 1.6.0",
  "jobs_common",
  "log",
  "my-workspace-hack",
@@ -13315,7 +13331,7 @@ dependencies = [
  "log",
  "my-workspace-hack",
  "once_cell",
- "reqwest 0.12.7",
+ "reqwest 0.12.12",
  "rss",
  "scraper",
  "serde",
diff --git a/crates/desktop/src-tauri/Cargo.toml b/crates/desktop/src-tauri/Cargo.toml
index 23a42e61ed..c06ccc1160 100644
--- a/crates/desktop/src-tauri/Cargo.toml
+++ b/crates/desktop/src-tauri/Cargo.toml
@@ -39,6 +39,7 @@ serde_yaml = "0.9"
 log = "0.4"
 once_cell = "1.20.3"
 rand = "0.8"
+reqwest = "0.12.12"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 tauri = { version = "2.2.4", features = [] }
diff --git a/crates/desktop/src-tauri/src/endpoints/download_models.rs b/crates/desktop/src-tauri/src/endpoints/download_models.rs
new file mode 100644
index 0000000000..ae85380048
--- /dev/null
+++ b/crates/desktop/src-tauri/src/endpoints/download_models.rs
@@ -0,0 +1,17 @@
+use crate::ml::downloads::download_all_models::download_all_models;
+use crate::state::app_config::AppConfig;
+use tauri::{AppHandle, State};
+
+#[tauri::command]
+pub async fn download_models(
+  model_config: State<'_, AppConfig>,
+  app: AppHandle,
+) -> Result<String, String> {
+  
+  download_all_models().await
+    .map_err(|err| {
+      "error downloading models".to_string()
+    })?;
+  
+  Ok(("downloaded".to_string()))
+}
diff --git a/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs b/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs
index a202e396ba..3336a54896 100644
--- a/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs
+++ b/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs
@@ -10,6 +10,7 @@ use log::{error, info};
 use std::io::Cursor;
 use std::path::PathBuf;
 use tauri::{AppHandle, State};
+use crate::ml::downloads::download_all_models::download_all_models;
 
 const PROMPT_FILENAME : &str = "prompt.txt";
 const PROMPT: &str = "A beautiful landscape with mountains and a lake";
@@ -40,7 +41,14 @@ pub async fn infer_image(
   let image = hydrate_base64_image(image)
     .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
 
-  let result = do_infer_image(&prompt, image, strength, &model_config, &model_cache, prompt_cache, app);
+  // TODO(bt): Move this to another endpoint
+  download_all_models().await
+    .map_err(|err| {
+      error!("couldn't download models: {:?}", err);
+      "Couldn't download models".to_string()
+    })?;
+
+  let result = do_infer_image(&prompt, image, strength, &model_config, &model_cache, prompt_cache, app).await;
   
   if let Err(err) = result.as_deref() {
     error!("There was an error: {:?}", err);
@@ -49,22 +57,20 @@ pub async fn infer_image(
   result
 }
 
-fn do_infer_image(
+async fn do_infer_image(
   prompt: &str,
   image: DynamicImage,
   strength: Option<u8>,
   config: &AppConfig,
   model_cache: &ModelCache,
-  prompt_cache: State<PromptCache>,
+  prompt_cache: State<'_, PromptCache>,
   app: AppHandle,
 ) -> Result<String, String> {
-  println!("infer_image called; generating image with SDXL Turbo...");
-  
+
   let args = Args {
     image: &image,
     prompt: prompt.to_string(),
     uncond_prompt: NEGATIVE_PROMPT.to_string(),
-    //guidance_scale: Some(0.0),
     model_cache,
     configs: config,
     prompt_cache: &prompt_cache,
diff --git a/crates/desktop/src-tauri/src/endpoints/mod.rs b/crates/desktop/src-tauri/src/endpoints/mod.rs
index 4872b3088f..fc4c6dac2a 100644
--- a/crates/desktop/src-tauri/src/endpoints/mod.rs
+++ b/crates/desktop/src-tauri/src/endpoints/mod.rs
@@ -1,3 +1,4 @@
+pub mod download_models;
 pub mod flip_image;
 pub mod image_endpoint;
 pub mod test_counter;
diff --git a/crates/desktop/src-tauri/src/lib.rs b/crates/desktop/src-tauri/src/lib.rs
index 7f656b907b..524a98e151 100644
--- a/crates/desktop/src-tauri/src/lib.rs
+++ b/crates/desktop/src-tauri/src/lib.rs
@@ -6,6 +6,7 @@ pub mod state;
 use crate::ml::model_cache::ModelCache;
 use crate::ml::prompt_cache::PromptCache;
 use crate::state::app_config::AppConfig;
+use endpoints::download_models::download_models;
 use endpoints::image_endpoint::infer_image;
 use endpoints::test_counter::test_counter;
 use tauri_plugin_log::Target;
@@ -50,6 +51,7 @@ pub fn run() {
     .invoke_handler(tauri::generate_handler![
       test_counter, 
       infer_image,
+      download_models,
     ])
     .run(tauri::generate_context!())
     .expect("error while running tauri application");
diff --git a/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs b/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
new file mode 100644
index 0000000000..3a916b6122
--- /dev/null
+++ b/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
@@ -0,0 +1,8 @@
+use crate::ml::downloads::download_model_registry_file::download_model_registry_file;
+use crate::ml::model_registry::ModelRegistry;
+
+pub async fn download_all_models() -> anyhow::Result<()> {
+  download_model_registry_file(ModelRegistry::Clip).await?;
+
+  Ok(())
+}
diff --git a/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs b/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs
new file mode 100644
index 0000000000..375e88352d
--- /dev/null
+++ b/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs
@@ -0,0 +1,21 @@
+use std::fs::File;
+use std::path::{Path, PathBuf};
+use crate::ml::model_registry::ModelRegistry;
+
+pub async fn download_model_registry_file(model_type: ModelRegistry) -> anyhow::Result<PathBuf> {
+  let path = PathBuf::from(model_type.get_filename());
+  if path.exists() {
+    return Ok(path);
+  }
+
+  let body = reqwest::get(model_type.get_download_url())
+    .await?
+    .bytes()
+    .await?;
+  
+  let _file = File::create(&path)?;
+  
+  std::fs::write(&path, body.as_ref())?;
+  
+  Ok(path)
+}
\ No newline at end of file
diff --git a/crates/desktop/src-tauri/src/ml/downloads/mod.rs b/crates/desktop/src-tauri/src/ml/downloads/mod.rs
new file mode 100644
index 0000000000..0ca6ed312a
--- /dev/null
+++ b/crates/desktop/src-tauri/src/ml/downloads/mod.rs
@@ -0,0 +1,2 @@
+pub mod download_model_registry_file;
+pub mod download_all_models;
diff --git a/crates/desktop/src-tauri/src/ml/mod.rs b/crates/desktop/src-tauri/src/ml/mod.rs
index 87a2bc1e4c..42ca9d128b 100644
--- a/crates/desktop/src-tauri/src/ml/mod.rs
+++ b/crates/desktop/src-tauri/src/ml/mod.rs
@@ -3,4 +3,6 @@ pub mod model_cache;
 pub mod model_file;
 pub mod models;
 pub mod prompt_cache;
-pub mod stable_diffusion;
\ No newline at end of file
+pub mod stable_diffusion;
+pub mod model_registry;
+pub mod downloads;
\ No newline at end of file
diff --git a/crates/desktop/src-tauri/src/ml/model_registry.rs b/crates/desktop/src-tauri/src/ml/model_registry.rs
new file mode 100644
index 0000000000..5204bad425
--- /dev/null
+++ b/crates/desktop/src-tauri/src/ml/model_registry.rs
@@ -0,0 +1,25 @@
+
+#[derive(Copy,Clone)]
+pub enum ModelRegistry {
+  Clip
+}
+
+impl ModelRegistry {
+  pub fn get_hf_id(&self) -> Option<&'static str> {
+    match self {
+      Self::Clip => Some("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"),
+    }
+  }
+  
+  pub fn get_download_url(&self) -> &'static str {
+    match self {
+      Self::Clip => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/tokenizer.json",
+    }
+  }
+  
+  pub fn get_filename(&self) -> &'static str {
+    match self {
+      Self::Clip => "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
+    }
+  }
+}
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs
index 8a4836f729..80cade0ff3 100644
--- a/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs
@@ -3,6 +3,7 @@ extern crate accelerate_src;
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
 
+use std::path::PathBuf;
 use candle_transformers::models::stable_diffusion;
 
 use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
@@ -10,6 +11,7 @@ use anyhow::Error as E;
 use candle_core::{DType, Device, Module, Tensor, D};
 use log::info;
 use tokenizers::Tokenizer;
+use crate::ml::model_registry::ModelRegistry;
 
 #[allow(clippy::too_many_arguments)]
 pub fn infer_clip_text_embeddings(
@@ -76,12 +78,21 @@ fn do_infer_clip_text_embeddings(
   ) -> anyhow::Result<Tensor> {
   
   let tokenizer_file = if first {
+    info!("ModelFile::Tokenizer");
     ModelFile::Tokenizer
   } else {
+    info!("ModelFile::Tokenizer2");
     ModelFile::Tokenizer2
   };
-  let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
+  
+  info!("Loading clip from download");
+  //let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
+  let tokenizer = PathBuf::from(ModelRegistry::Clip.get_filename());
+  
+  info!("Tokenizer path: {:?}", tokenizer);
+  
   let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
+  
   let pad_id = match &sd_config.clip.pad_with {
     Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
     None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
index 16f45200a7..ea8f624bd7 100644
--- a/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
@@ -1,9 +1,12 @@
 use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
 
+use crate::events::notification_event::{ModelType, NotificationEvent};
 use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
 use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
 use crate::ml::model_cache::ModelCache;
 use crate::ml::model_file::StableDiffusionVersion;
+use crate::ml::model_registry::ModelRegistry;
+use crate::ml::models::unet_model::UNetModel;
 use crate::ml::prompt_cache::PromptCache;
 use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
 use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
@@ -15,8 +18,6 @@ use image::DynamicImage;
 use log::info;
 use rand::Rng;
 use tauri::{AppHandle, Emitter};
-use crate::events::notification_event::{ModelType, NotificationEvent};
-use crate::ml::models::unet_model::UNetModel;
 
 pub struct Args<'a> {
     pub image: &'a DynamicImage,

From 706becfd7e978b292835a3815d27731f6e0c0b30 Mon Sep 17 00:00:00 2001
From: Brandon Thomas <echelon@gmail.com>
Date: Wed, 5 Mar 2025 19:32:41 -0500
Subject: [PATCH 5/8] download unet

---
 .../src/ml/downloads/download_all_models.rs        |  1 +
 .../ml/downloads/download_model_registry_file.rs   |  3 +++
 crates/desktop/src-tauri/src/ml/model_registry.rs  | 12 ++++++++----
 .../stable_diffusion/stable_diffusion_pipeline.rs  | 14 ++++----------
 4 files changed, 16 insertions(+), 14 deletions(-)

diff --git a/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs b/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
index 3a916b6122..6699daf7a5 100644
--- a/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
+++ b/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
@@ -3,6 +3,7 @@ use crate::ml::model_registry::ModelRegistry;
 
 pub async fn download_all_models() -> anyhow::Result<()> {
   download_model_registry_file(ModelRegistry::Clip).await?;
+  download_model_registry_file(ModelRegistry::SdxlTurboUnet).await?;
 
   Ok(())
 }
diff --git a/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs b/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs
index 375e88352d..0e8cc0fb14 100644
--- a/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs
+++ b/crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs
@@ -1,10 +1,13 @@
 use std::fs::File;
 use std::path::{Path, PathBuf};
+use log::info;
 use crate::ml::model_registry::ModelRegistry;
 
 pub async fn download_model_registry_file(model_type: ModelRegistry) -> anyhow::Result<PathBuf> {
+  info!("downloading model: {:?}", model_type);
   let path = PathBuf::from(model_type.get_filename());
   if path.exists() {
+    info!("model already exists: {:?}", path);
     return Ok(path);
   }
 
diff --git a/crates/desktop/src-tauri/src/ml/model_registry.rs b/crates/desktop/src-tauri/src/ml/model_registry.rs
index 5204bad425..144431f4fd 100644
--- a/crates/desktop/src-tauri/src/ml/model_registry.rs
+++ b/crates/desktop/src-tauri/src/ml/model_registry.rs
@@ -1,25 +1,29 @@
 
-#[derive(Copy,Clone)]
+#[derive(Copy, Clone, Debug)]
 pub enum ModelRegistry {
-  Clip
+  Clip,
+  SdxlTurboUnet,
 }
 
 impl ModelRegistry {
   pub fn get_hf_id(&self) -> Option<&'static str> {
     match self {
       Self::Clip => Some("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"),
+      Self::SdxlTurboUnet => Some("stabilityai/sdxl-turbo"),
     }
   }
-  
+
   pub fn get_download_url(&self) -> &'static str {
     match self {
       Self::Clip => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/tokenizer.json",
+      Self::SdxlTurboUnet => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.unet.safetensors",
     }
   }
-  
+
   pub fn get_filename(&self) -> &'static str {
     match self {
       Self::Clip => "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
+      Self::SdxlTurboUnet => "diffusion_pytorch_model.unet.safetensors",
     }
   }
 }
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
index ea8f624bd7..d5bb1955d5 100644
--- a/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
@@ -168,23 +168,17 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No unet found in cache; loading...");
 
-            let repo = configs.sd_version.repo();
-
-            info!("Downloading UNET model files from: {} ...", repo);
-
-            let unet_file = configs.hf_api.model(repo.to_string())
-              .get("unet/diffusion_pytorch_model.safetensors")
-              .map_err(|err| anyhow!("error fetching model: {:?}", err))?;
-
             let mut notify_download_complete = false;
             //if !unet_file.exists() {
             if true {
                 notify_download_complete = true;
                 app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: repo,
+                    model_name: "sdxl-turbo-unet",
                     model_type: ModelType::Unet,
                 })?;
             }
+            
+            let unet_file = ModelRegistry::SdxlTurboUnet.get_filename();
 
             let unet = UNetModel::new(&configs.sd_config, unet_file, &configs.device, configs.dtype)
               .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
@@ -195,7 +189,7 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
             if notify_download_complete {
                 app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: repo,
+                    model_name: "sdxl-turbo-unet",
                     model_type: ModelType::Unet,
                 })?;
             }

From a859bd427bf4d5eda205afd2e38b07305f8377e5 Mon Sep 17 00:00:00 2001
From: Brandon Thomas <echelon@gmail.com>
Date: Thu, 6 Mar 2025 08:42:44 -0500
Subject: [PATCH 6/8] download clip models (not using them yet)

---
 .../src/ml/downloads/download_all_models.rs   |  6 +++--
 .../src-tauri/src/ml/model_registry.rs        | 22 ++++++++++++++-----
 .../infer_clip_text_embeddings.rs             |  2 +-
 .../stable_diffusion_pipeline.rs              | 15 ++++---------
 4 files changed, 26 insertions(+), 19 deletions(-)

diff --git a/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs b/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
index 6699daf7a5..953e084345 100644
--- a/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
+++ b/crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs
@@ -2,8 +2,10 @@ use crate::ml::downloads::download_model_registry_file::download_model_registry_
 use crate::ml::model_registry::ModelRegistry;
 
 pub async fn download_all_models() -> anyhow::Result<()> {
-  download_model_registry_file(ModelRegistry::Clip).await?;
+  download_model_registry_file(ModelRegistry::ClipJson).await?;
   download_model_registry_file(ModelRegistry::SdxlTurboUnet).await?;
-
+  download_model_registry_file(ModelRegistry::SdxlTurboVae).await?;
+  download_model_registry_file(ModelRegistry::SdxlTurboClipEncoder).await?;
+  download_model_registry_file(ModelRegistry::SdxlTurboClipEncoder2).await?;
   Ok(())
 }
diff --git a/crates/desktop/src-tauri/src/ml/model_registry.rs b/crates/desktop/src-tauri/src/ml/model_registry.rs
index 144431f4fd..e3a4e933b7 100644
--- a/crates/desktop/src-tauri/src/ml/model_registry.rs
+++ b/crates/desktop/src-tauri/src/ml/model_registry.rs
@@ -1,29 +1,41 @@
 
 #[derive(Copy, Clone, Debug)]
 pub enum ModelRegistry {
-  Clip,
+  ClipJson,
   SdxlTurboUnet,
+  SdxlTurboVae,
+  SdxlTurboClipEncoder,
+  SdxlTurboClipEncoder2,
 }
 
 impl ModelRegistry {
   pub fn get_hf_id(&self) -> Option<&'static str> {
     match self {
-      Self::Clip => Some("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"),
-      Self::SdxlTurboUnet => Some("stabilityai/sdxl-turbo"),
+      Self::ClipJson => Some("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"),
+      Self::SdxlTurboUnet 
+      | Self::SdxlTurboVae
+      | Self::SdxlTurboClipEncoder
+      | Self::SdxlTurboClipEncoder2 => Some("stabilityai/sdxl-turbo"),
     }
   }
 
   pub fn get_download_url(&self) -> &'static str {
     match self {
-      Self::Clip => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/tokenizer.json",
+      Self::ClipJson => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/tokenizer.json",
       Self::SdxlTurboUnet => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.unet.safetensors",
+      Self::SdxlTurboVae => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.vae.safetensors",
+      Self::SdxlTurboClipEncoder => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/clip_text_encoder.safetensors",
+      Self::SdxlTurboClipEncoder2 => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/clip_text_encoder_2.safetensors",
     }
   }
 
   pub fn get_filename(&self) -> &'static str {
     match self {
-      Self::Clip => "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
+      Self::ClipJson => "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
       Self::SdxlTurboUnet => "diffusion_pytorch_model.unet.safetensors",
+      Self::SdxlTurboVae => "diffusion_pytorch_model.vae.safetensors",
+      Self::SdxlTurboClipEncoder => "clip_text_encoder.safetensors",
+      Self::SdxlTurboClipEncoder2 => "clip_text_encoder_2.safetensors",
     }
   }
 }
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs
index 80cade0ff3..11a56137e1 100644
--- a/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs
@@ -87,7 +87,7 @@ fn do_infer_clip_text_embeddings(
   
   info!("Loading clip from download");
   //let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
-  let tokenizer = PathBuf::from(ModelRegistry::Clip.get_filename());
+  let tokenizer = PathBuf::from(ModelRegistry::ClipJson.get_filename());
   
   info!("Tokenizer path: {:?}", tokenizer);
   
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
index d5bb1955d5..60e3361078 100644
--- a/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs
@@ -123,25 +123,18 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No vae found in cache; loading...");
             
-            let repo = configs.sd_version.repo();
-
-            println!("Building VAE model from : {:?} ... (3)", repo);
-
-            let vae_file = configs.hf_api.model(repo.to_string())
-              .get("vae/diffusion_pytorch_model.safetensors")?;
-
-            println!("Building VAE model from file {:?}...", &vae_file);
-
             let mut notify_download_complete = false;
             //if !vae_file.exists() {
             if true {
                 notify_download_complete = true;
                 app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: repo,
+                    model_name: "sdxl-turbo-vae",
                     model_type: ModelType::Vae,
                 })?;
             }
             
+            let vae_file = ModelRegistry::SdxlTurboVae.get_filename();
+            
             let vae = configs
               .sd_config
               .build_vae(vae_file, &configs.device, configs.dtype)?;
@@ -152,7 +145,7 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
             
             if notify_download_complete {
                 app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: repo,
+                    model_name: "sdxl-turbo-vae",
                     model_type: ModelType::Vae,
                 })?;
             }

From c2929cfdb6748f585be4cb91cfcfbcad222e926d Mon Sep 17 00:00:00 2001
From: BFlat <bombayapirat@gmail.com>
Date: Thu, 6 Mar 2025 20:52:25 +0700
Subject: [PATCH 7/8] color picker ux and make paint cursor visible on both
 dark/light bg (#441)

ux updates to color picker and paint cursor
---
 .../RenderingPrimitives/RealTimeDrawEngine.ts | 34 ++++----
 frontend/realtime/src/app/global.css          | 27 +++++++
 .../src/components/ui/HexEyedropPicker.tsx    | 56 +++++++------
 .../components/ui/TextEditor/ColorPicker.tsx  |  7 +-
 .../ui/ToolbarMain/PaintModeMenu.tsx          | 78 ++++++++++++++++---
 5 files changed, 145 insertions(+), 57 deletions(-)

diff --git a/frontend/realtime/src/KonvaApp/RenderingPrimitives/RealTimeDrawEngine.ts b/frontend/realtime/src/KonvaApp/RenderingPrimitives/RealTimeDrawEngine.ts
index 0efc9e5a29..134e444fdb 100644
--- a/frontend/realtime/src/KonvaApp/RenderingPrimitives/RealTimeDrawEngine.ts
+++ b/frontend/realtime/src/KonvaApp/RenderingPrimitives/RealTimeDrawEngine.ts
@@ -4,7 +4,7 @@ import { Shape } from "konva/lib/Shape";
 import { Group } from "konva/lib/Group";
 
 import { invoke } from "@tauri-apps/api/core";
-import { listen } from '@tauri-apps/api/event';
+import { listen } from "@tauri-apps/api/event";
 
 import { FileUtilities } from "../FileUtilities/FileUtilities";
 import { ImageNode, VideoNode, TextNode, ShapeNode, ShapeType } from "../Nodes";
@@ -31,7 +31,6 @@ interface ServerSettings {
   lora_path?: string;
 }
 
-
 interface GenerateImageParams {
   image: string;
   prompt: string;
@@ -210,16 +209,18 @@ export class RealTimeDrawEngine {
       console.log(event);
       // Handle model download events
       const payload = event.payload as any;
-      
+
       // Model download started
       if (payload.model_download_started) {
         const modelInfo = payload.model_download_started;
-        console.log(`Model download started: ${modelInfo.model_name} (${modelInfo.model_type})`);
-        
+        console.log(
+          `Model download started: ${modelInfo.model_name} (${modelInfo.model_type})`,
+        );
+
         // Set up loading indicator
         isLoadingVisible.value = true;
         loadingProgress.value = 0;
-        
+
         // Create fake progress updates
         const downloadTimer = setInterval(() => {
           loadingProgress.value += 2;
@@ -230,23 +231,23 @@ export class RealTimeDrawEngine {
           }
         }, 500);
       }
-      
+
       // Model download completed
       if (payload.model_download_complete) {
         const modelInfo = payload.model_download_complete;
-        console.log(`Model download completed: ${modelInfo.model_name} (${modelInfo.model_type})`);
-        
+        console.log(
+          `Model download completed: ${modelInfo.model_name} (${modelInfo.model_type})`,
+        );
+
         // Complete the loading progress
         loadingProgress.value = 100;
-        
+
         // Hide loading indicator after a short delay
         setTimeout(() => {
           isLoadingVisible.value = false;
         }, 1000);
       }
-  
     });
-    
   }
 
   private isEnabled: boolean = false;
@@ -320,9 +321,16 @@ export class RealTimeDrawEngine {
     const ctx = cursorCanvas.getContext("2d");
     if (!ctx) return;
 
-    // Draw the circle
+    // Draw the outer circle with a light stroke
     ctx.beginPath();
     ctx.arc(size, size, size / 2, 0, Math.PI * 2);
+    ctx.strokeStyle = "white";
+    ctx.lineWidth = 3;
+    ctx.stroke();
+
+    // Draw the inner circle with a dark stroke
+    ctx.beginPath();
+    ctx.arc(size, size, size / 2 - 1, 0, Math.PI * 2);
     ctx.strokeStyle = "black";
     ctx.lineWidth = 1;
     ctx.stroke();
diff --git a/frontend/realtime/src/app/global.css b/frontend/realtime/src/app/global.css
index 3c0abd0ee7..2a70999845 100644
--- a/frontend/realtime/src/app/global.css
+++ b/frontend/realtime/src/app/global.css
@@ -202,3 +202,30 @@
     background-position: -200% 0;
   }
 }
+
+.react-colorful__saturation {
+  margin: 16px 0;
+  border-radius: 10px !important;
+  border-bottom: none !important;
+}
+
+.react-colorful__saturation-pointer {
+  width: 24px !important;
+  height: 24px !important;
+}
+
+.react-colorful__hue,
+.react-colorful__alpha {
+  height: 20px !important;
+  border-radius: 50px !important;
+}
+
+.react-colorful__hue {
+  margin-bottom: 12px;
+}
+
+.react-colorful__hue-pointer,
+.react-colorful__alpha-pointer {
+  width: 20px !important;
+  height: 20px !important;
+}
diff --git a/frontend/realtime/src/components/ui/HexEyedropPicker.tsx b/frontend/realtime/src/components/ui/HexEyedropPicker.tsx
index d2e1791b0b..9c0af09a8e 100644
--- a/frontend/realtime/src/components/ui/HexEyedropPicker.tsx
+++ b/frontend/realtime/src/components/ui/HexEyedropPicker.tsx
@@ -1,12 +1,13 @@
-import { MouseEvent, MouseEventHandler, useState } from "react";
-import { HexColorPicker } from "react-colorful";
-import { EyeDropper } from "react-eyedrop";
 import { Button } from "./Button";
 import { faEyedropper } from "@fortawesome/pro-solid-svg-icons";
 import { twMerge } from "tailwind-merge";
 
-type ColorPickerHTMLAttributes = Omit<React.HTMLAttributes<HTMLDivElement>, "color" | "onChange" | "onChangeCapture">;
-interface ColorPickerBaseProps<T extends string> extends ColorPickerHTMLAttributes {
+type ColorPickerHTMLAttributes = Omit<
+  React.HTMLAttributes<HTMLDivElement>,
+  "color" | "onChange" | "onChangeCapture"
+>;
+interface ColorPickerBaseProps<T extends string>
+  extends ColorPickerHTMLAttributes {
   color: T;
   onChange: (newColor: T) => void;
 }
@@ -17,7 +18,7 @@ export const HexEyedropPicker = ({
   dropperClassName,
   pickerClassName,
   onPickerChange,
-  Picker
+  Picker,
 }: {
   color: string;
   onDropperChange?: (newColor: string) => void;
@@ -26,9 +27,8 @@ export const HexEyedropPicker = ({
   onPickerChange: (newColor: string) => void;
   Picker: (props: Partial<ColorPickerBaseProps<string>>) => JSX.Element;
 }) => {
-
   const handleDropperClick = () => {
-    if (!('EyeDropper' in window)) {
+    if (!("EyeDropper" in window)) {
       // NOTE: This shouldn't happen cuz we're using Chromium
       console.error("EyeDropper not found in window");
       return;
@@ -46,42 +46,40 @@ export const HexEyedropPicker = ({
         // User cancelled the picker
         return null;
       }
-    }
+    };
 
     startEyeDrop().then((newColor) => {
       if (newColor) {
         onDropperChange?.(newColor);
       }
     });
-  }
+  };
 
-  const [isDragging, setIsDragging] = useState(false);
-  const handleOnMouseDown = () => {
-    setIsDragging(true);
-  }
-  const handleOnMouseUp = () => {
-    setIsDragging(false);
-  }
+  // const [isDragging, setIsDragging] = useState(false);
+  // const handleOnMouseDown = () => {
+  //   setIsDragging(true);
+  // };
+  // const handleOnMouseUp = () => {
+  //   setIsDragging(false);
+  // };
 
   return (
     <div className={"relative"}>
-      <Picker
-        className={pickerClassName}
-        color={color}
-        onChange={onPickerChange}
-        onMouseDown={handleOnMouseDown}
-        onMouseUp={handleOnMouseUp}
-      />
       <Button
         className={twMerge([
-          "rounded-full m-1 size-8 text-ui-controls bg-transparent hover:bg-black/70 hover:text-white z-20 shadow-none hover:shadow-xl duration-200",
+          "z-20 size-8 rounded-xl bg-ui-background/30 text-white shadow-none duration-200 hover:bg-white/10 hover:text-white hover:shadow-xl",
           dropperClassName,
-          "absolute left-0 top-0",
-          isDragging ? "opacity-0 -z-10" : ""
         ])}
         icon={faEyedropper}
         onClick={handleDropperClick}
       />
+      <Picker
+        className={pickerClassName}
+        color={color}
+        onChange={onPickerChange}
+        // onMouseDown={handleOnMouseDown}
+        // onMouseUp={handleOnMouseUp}
+      />
     </div>
-  )
-}
+  );
+};
diff --git a/frontend/realtime/src/components/ui/TextEditor/ColorPicker.tsx b/frontend/realtime/src/components/ui/TextEditor/ColorPicker.tsx
index fc55f775b5..97cdaef38d 100644
--- a/frontend/realtime/src/components/ui/TextEditor/ColorPicker.tsx
+++ b/frontend/realtime/src/components/ui/TextEditor/ColorPicker.tsx
@@ -91,7 +91,7 @@ export const ColorPicker = ({
     if (streamChanges) {
       onChange(color);
     }
-  }
+  };
 
   return (
     <Popover className="relative">
@@ -124,7 +124,7 @@ export const ColorPicker = ({
                 anchor={anchor}
                 className={twMerge(
                   paperWrapperStyles,
-                  "flex flex-col items-center gap-2 overflow-hidden",
+                  "flex flex-col items-center gap-2 overflow-hidden p-4",
                 )}
                 style={
                   {
@@ -134,13 +134,14 @@ export const ColorPicker = ({
                 static
               >
                 <HexEyedropPicker
-                  pickerClassName="overflow-hidden"
                   color={currColor}
                   onPickerChange={handleColorChange}
                   onDropperChange={handleColorChange}
                   Picker={HexAlphaColorPicker}
                 />
                 <Input
+                  className="mt-2"
+                  inputClassName="bg-ui-background/30"
                   style={{ width: "198px" }}
                   icon={faHashtag}
                   value={textInput}
diff --git a/frontend/realtime/src/components/ui/ToolbarMain/PaintModeMenu.tsx b/frontend/realtime/src/components/ui/ToolbarMain/PaintModeMenu.tsx
index e42f0a903e..5dee0f52c1 100644
--- a/frontend/realtime/src/components/ui/ToolbarMain/PaintModeMenu.tsx
+++ b/frontend/realtime/src/components/ui/ToolbarMain/PaintModeMenu.tsx
@@ -5,13 +5,18 @@ import {
 } from "@fortawesome/pro-solid-svg-icons";
 import { FontAwesomeIcon } from "@fortawesome/react-fontawesome";
 import { Popover, PopoverButton, PopoverPanel } from "@headlessui/react";
-import { ChangeEvent, useEffect, useState } from "react";
+import { ChangeEvent, useEffect, useState, useRef } from "react";
 import { HexAlphaColorPicker } from "react-colorful";
 import { twMerge } from "tailwind-merge";
 
 import { useSignals } from "@preact/signals-react/runtime";
 import { paperWrapperStyles } from "~/components/styles";
-import { BRUSH_MAX_SIZE, BRUSH_MIN_SIZE, paintBrushSize, setPaintBrushSize } from "~/signals/uiEvents/toolbarMain/paintMode";
+import {
+  BRUSH_MAX_SIZE,
+  BRUSH_MIN_SIZE,
+  paintBrushSize,
+  setPaintBrushSize,
+} from "~/signals/uiEvents/toolbarMain/paintMode";
 import { Button } from "../Button";
 import { Input } from "../Input";
 import { Slider } from "../Slider";
@@ -52,16 +57,58 @@ export const PaintModeMenu = ({
     currColor: prevColor,
     textInput: prevColor.substring(1),
   });
+  const [isMouseDown, setIsMouseDown] = useState(false);
+  const [isMouseOutside, setIsMouseOutside] = useState(false);
+  const closeRef = useRef<(() => void) | null>(null);
+  const closeTimeoutRef = useRef<NodeJS.Timeout | null>(null);
 
   useEffect(() => {
     setIsOpen(defaultOpen);
   }, [defaultOpen]);
 
+  // Global mouse up handler
+  useEffect(() => {
+    const handleGlobalMouseUp = () => {
+      if (isMouseDown) {
+        setIsMouseDown(false);
+
+        // If mouse is outside when released, close the panel after a short delay
+        if (isMouseOutside && closeOnMouseLeave && closeRef.current) {
+          closeTimeoutRef.current = setTimeout(() => {
+            onChange(currColor);
+            setIsOpen(false);
+            closeRef.current?.();
+          }, 500);
+        }
+      }
+    };
+
+    window.addEventListener("mouseup", handleGlobalMouseUp);
+
+    return () => {
+      window.removeEventListener("mouseup", handleGlobalMouseUp);
+    };
+  }, [isMouseDown, isMouseOutside, closeOnMouseLeave, currColor, onChange]);
+
   const handleMouseLeave = (close: () => void) => {
-    if (closeOnMouseLeave) {
-      onChange(currColor);
-      setIsOpen(false);
-      close();
+    closeRef.current = close;
+    setIsMouseOutside(true);
+
+    // Only close if mouse is not down and closeOnMouseLeave is true
+    if (closeOnMouseLeave && !isMouseDown) {
+      closeTimeoutRef.current = setTimeout(() => {
+        onChange(currColor);
+        setIsOpen(false);
+        close();
+      }, 200);
+    }
+  };
+
+  const handleMouseEnter = () => {
+    setIsMouseOutside(false);
+    if (closeTimeoutRef.current) {
+      clearTimeout(closeTimeoutRef.current);
+      closeTimeoutRef.current = null;
     }
   };
 
@@ -92,7 +139,7 @@ export const PaintModeMenu = ({
     if (streamChanges) {
       onChange(color);
     }
-  }
+  };
 
   // Rerender the component when these signals change:
   useSignals();
@@ -123,12 +170,16 @@ export const PaintModeMenu = ({
             )}
           </PopoverButton>
           {(open || isOpen) && (
-            <div onMouseLeave={() => handleMouseLeave(close)}>
+            <div
+              onMouseLeave={() => handleMouseLeave(close)}
+              onMouseEnter={handleMouseEnter}
+              onMouseDown={() => setIsMouseDown(true)}
+            >
               <PopoverPanel
                 anchor={anchor}
                 className={twMerge(
                   paperWrapperStyles,
-                  "flex flex-col items-center gap-2 overflow-hidden",
+                  "flex flex-col items-center gap-2 overflow-hidden p-4",
                 )}
                 style={
                   {
@@ -138,20 +189,23 @@ export const PaintModeMenu = ({
                 static
               >
                 <HexEyedropPicker
-                  pickerClassName="overflow-hidden"
                   color={currColor}
                   onPickerChange={handleColorChange}
                   onDropperChange={handleColorChange}
                   Picker={HexAlphaColorPicker}
                 />
                 <Input
+                  className="mt-2"
+                  inputClassName="bg-ui-background/30"
                   style={{ width: "198px" }}
                   icon={faHashtag}
                   value={textInput}
                   onChange={handleTextInput}
                 />
-                <div className="flex flex-col w-full mt-2 gap-2">
-                  <p className="text-white font-medium text-sm w-full justify-start">Brush Size:</p>
+                <div className="mt-2 flex w-full flex-col gap-2">
+                  <p className="w-full justify-start text-sm font-medium text-white">
+                    Brush Size:
+                  </p>
                   <Slider
                     min={BRUSH_MIN_SIZE}
                     max={BRUSH_MAX_SIZE}

From 3f25aefc3f819484414b49727f7a6ee29bb1a031 Mon Sep 17 00:00:00 2001
From: Kasisnu <5044120+kasisnu@users.noreply.github.com>
Date: Thu, 6 Mar 2025 21:11:08 +0530
Subject: [PATCH 8/8] starting point (#440)

---
 app_config.yaml                               |    2 +-
 crates/desktop/src-tauri/Cargo.toml           |    2 +-
 .../src-tauri/src/endpoints/image_endpoint.rs |    4 +-
 .../src-tauri/src/ml/models/unet_model.rs     |   29 +-
 .../src/ml/stable_diffusion/lcm_pipeline.rs   |  294 +++
 .../src-tauri/src/ml/stable_diffusion/mod.rs  |    3 +-
 .../desktop/src-tauri/src/state/app_config.rs |    2 +-
 .../candle-examples/examples/lcm/main.rs      | 1670 +++++++++--------
 .../vendor/candle/candle-flash-attn/build.rs  |    2 +
 crates/vendor/candle/candle-nn/Cargo.toml     |    1 +
 .../src/models/stable_diffusion/lcm.rs        | 1148 +++++------
 .../src/models/stable_diffusion/mod.rs        |   10 +
 .../src/models/stable_diffusion/unet_2d.rs    |  154 +-
 13 files changed, 1853 insertions(+), 1468 deletions(-)
 create mode 100644 crates/desktop/src-tauri/src/ml/stable_diffusion/lcm_pipeline.rs

diff --git a/app_config.yaml b/app_config.yaml
index fca2ac4c14..2ddf4fd258 100644
--- a/app_config.yaml
+++ b/app_config.yaml
@@ -2,6 +2,6 @@ image_width: 512
 image_height: 512
 #scheduler_steps: 15
 # NB(bt): It looks like we can get down to 2 sampling steps, perhaps ?
-scheduler_steps: 2
+scheduler_steps: 4
 scheduler_samples: 1
 cfg_scale: 1.1
diff --git a/crates/desktop/src-tauri/Cargo.toml b/crates/desktop/src-tauri/Cargo.toml
index c06ccc1160..aaa0304add 100644
--- a/crates/desktop/src-tauri/Cargo.toml
+++ b/crates/desktop/src-tauri/Cargo.toml
@@ -48,7 +48,7 @@ tokenizers = "0.15.2"
 
 [features]
 default = ["cuda"]
-cuda = ["candle-core/cuda", "candle-nn/cuda", "candle-transformers/cuda"]
+cuda = ["candle-core/cuda", "candle-nn/cuda", "candle-transformers/cuda", "candle-transformers/flash-attn", "candle-core/cudnn"]
 
 [dev-dependencies]
 libloading = "0.8.6"
diff --git a/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs b/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs
index 3336a54896..6343743efb 100644
--- a/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs
+++ b/crates/desktop/src-tauri/src/endpoints/image_endpoint.rs
@@ -1,6 +1,6 @@
 use crate::ml::model_cache::ModelCache;
 use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::stable_diffusion_pipeline::{stable_diffusion_pipeline, Args};
+use crate::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
 use crate::state::app_config::AppConfig;
 use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine};
 use bytes::BytesMut;
@@ -79,7 +79,7 @@ async fn do_infer_image(
     app: &app,
   };
 
-  match stable_diffusion_pipeline(args) {
+  match lcm_pipeline(args) {
     Ok(image) => {
       let mut bytes = Vec::with_capacity(1024*1024);
       
diff --git a/crates/desktop/src-tauri/src/ml/models/unet_model.rs b/crates/desktop/src-tauri/src/ml/models/unet_model.rs
index 2483525c96..4a8416c02f 100644
--- a/crates/desktop/src-tauri/src/ml/models/unet_model.rs
+++ b/crates/desktop/src-tauri/src/ml/models/unet_model.rs
@@ -8,28 +8,19 @@ pub struct UNetModel {
 }
 
 impl UNetModel {
-  pub fn new<P: AsRef<Path>>(
-    sd_config: &StableDiffusionConfig,
-    unet_weights_file: P,
-    device: &Device,
-    dtype: DType,
-  ) -> anyhow::Result<Self> {
+  pub fn new<P: AsRef<Path>>(sd_config: &StableDiffusionConfig, unet_weights_file: P, device: &Device, dtype: DType) -> anyhow::Result<Self> {
     println!("building unet model... (1)");
-    
-    let model = sd_config.build_unet(
-      unet_weights_file, 
-      &device, 
-      4, 
-      false, 
-      dtype
-    )?;
-    
-    Ok(Self {
-      model,
-    })
+
+    let model = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)?;
+
+    Ok(Self { model })
   }
-  
+
   pub fn inference(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor) -> anyhow::Result<Tensor> {
     Ok(self.model.forward(&latent_model_input, timestep, &text_embeddings)?)
   }
+
+  pub fn forward_with_guidance(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor, guidance_scale_embedding: &Tensor) -> anyhow::Result<Tensor> {
+    Ok(self.model.forward_with_guidance(&latent_model_input, timestep, &text_embeddings, Some(guidance_scale_embedding))?)
+  }
 }
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/lcm_pipeline.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/lcm_pipeline.rs
new file mode 100644
index 0000000000..ed294220d5
--- /dev/null
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/lcm_pipeline.rs
@@ -0,0 +1,294 @@
+use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
+
+use crate::events::notification_event::{ModelType, NotificationEvent};
+use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
+use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
+use crate::ml::model_cache::ModelCache;
+use crate::ml::model_file::StableDiffusionVersion;
+use crate::ml::model_registry::ModelRegistry;
+use crate::ml::models::unet_model::UNetModel;
+use crate::ml::prompt_cache::PromptCache;
+use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
+use crate::state::app_config::AppConfig;
+use anyhow::{anyhow, Error as E, Result};
+use candle_core::{DType, IndexOp, Tensor, D};
+use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
+use image::DynamicImage;
+use log::info;
+use rand::Rng;
+use tauri::{AppHandle, Emitter};
+
+pub struct Args<'a> {
+    pub image: &'a DynamicImage,
+    pub prompt: String,
+    pub uncond_prompt: String,
+    pub cfg_scale: Option<f64>,
+    pub i2i_strength: Option<u8>,
+    pub configs: &'a AppConfig,
+    pub model_cache: &'a ModelCache,
+    pub prompt_cache: &'a PromptCache,
+    pub app: &'a AppHandle,
+}
+
+pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
+    let Args { 
+        prompt, 
+        uncond_prompt, 
+        cfg_scale, 
+        i2i_strength, 
+        configs, 
+        model_cache, 
+        prompt_cache, 
+        app, 
+        image,  
+    } = args;
+    
+    println!("Starting image generation with the following configuration:");
+    println!("  Model: {:?}", configs.sd_version);
+    println!("  Prompt: {}", prompt);
+    println!("  Steps: {}", configs.scheduler_steps);
+    println!("  Device: {:?}", configs.device);
+
+    println!("Model dimensions: {}x{}", configs.sd_config.width, configs.sd_config.height);
+
+    // Use LCM Scheduler instead of Euler Ancestral for better speed and quality
+    let mut scheduler = LCMScheduler::new(
+        configs.scheduler_steps,
+        candle_transformers::models::stable_diffusion::lcm::LCMSchedulerConfig::default(),
+    )?;
+
+    let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
+    configs.device.set_seed(seed)?;
+
+    info!("Using seed: {}", seed);
+
+    let guidance_scale = match cfg_scale {
+        Some(guidance_scale) => guidance_scale,
+        None => match configs.sd_version {
+            StableDiffusionVersion::V1_5
+            | StableDiffusionVersion::V2_1
+            | StableDiffusionVersion::Xl => 7.5,
+            StableDiffusionVersion::Turbo => 0.,
+            _ => 0., // NB: Not sure what the other model families should use, so sticking with "0"
+        },
+    };
+
+    let use_guide_scale = guidance_scale > 1.0;
+
+    info!("Using guide scale: {}", use_guide_scale);
+
+    info!("Checking if prompt is cached");
+    let maybe_cached = prompt_cache.get_copy(&prompt)?;
+
+    let mut text_embeddings = if let Some(tensor) = maybe_cached {
+        tensor
+    } else {
+        
+        info!("Prompt is NOT cached! Calculating embedding...");
+        let tensor = infer_clip_text_embeddings(
+            &prompt,
+            &uncond_prompt,
+            None, // tokenizer
+            None, // clip_weights
+            None, // clip2_weights
+            configs.sd_version,
+            &configs.sd_config,
+            false, // use_f16
+            &configs.device,
+            configs.dtype,
+            use_guide_scale,
+        )?;
+        prompt_cache.store_copy(&prompt, &tensor)?;
+        tensor
+    };
+
+    println!("Text embeddings shape: {:?}", text_embeddings.shape());
+
+    println!("Loading input image into tensor...");
+
+    let input_image = dynamic_image_to_tensor(
+        image,
+        &configs.device,
+        configs.dtype)?;
+
+    println!("Reference image shape: {:?}", input_image.shape());
+
+    let vae_scale = get_vae_scale(configs.sd_version);
+
+    let maybe_vae = model_cache.get_vae()?;
+    
+    let vae = match maybe_vae {
+        Some(vae) => vae,
+        None => {
+            info!("No vae found in cache; loading...");
+            
+            let repo = configs.sd_version.repo();
+
+            println!("Building VAE model from : {:?} ... (3)", repo);
+
+            let vae_file = configs.hf_api.model(repo.to_string())
+              .get("vae/diffusion_pytorch_model.safetensors")?;
+
+            println!("Building VAE model from file {:?}...", &vae_file);
+
+            let mut notify_download_complete = false;
+            //if !vae_file.exists() {
+            if true {
+                notify_download_complete = true;
+                app.emit("notification", NotificationEvent::ModelDownloadStarted {
+                    model_name: repo,
+                    model_type: ModelType::Vae,
+                })?;
+            }
+            
+            let vae = configs
+              .sd_config
+              .build_vae(vae_file, &configs.device, configs.dtype)?;
+
+            let vae = Arc::new(vae);
+            
+            model_cache.set_vae(vae.clone())?;
+            
+            if notify_download_complete {
+                app.emit("notification", NotificationEvent::ModelDownloadComplete {
+                    model_name: repo,
+                    model_type: ModelType::Vae,
+                })?;
+            }
+            
+            vae
+        }
+    };
+
+    let maybe_unet = model_cache.get_unet()?;
+
+    let unet = match maybe_unet {
+        Some(unet) => unet,
+        None => {
+            info!("No unet found in cache; loading...");
+
+            let mut notify_download_complete = false;
+            //if !unet_file.exists() {
+            if true {
+                notify_download_complete = true;
+                app.emit("notification", NotificationEvent::ModelDownloadStarted {
+                    model_name: "sdxl-turbo-unet",
+                    model_type: ModelType::Unet,
+                })?;
+            }
+            
+            let unet_file = ModelRegistry::SdxlTurboUnet.get_filename();
+
+            let unet = UNetModel::new(&configs.sd_config, unet_file, &configs.device, configs.dtype)
+              .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
+            
+            let unet = Arc::new(unet);
+
+            model_cache.set_unet(unet.clone())?;
+
+            if notify_download_complete {
+                app.emit("notification", NotificationEvent::ModelDownloadComplete {
+                    model_name: "sdxl-turbo-unet",
+                    model_type: ModelType::Unet,
+                })?;
+            }
+
+            unet
+        }
+    };
+
+    let init_latent_dist : DiagonalGaussianDistribution = vae.encode(&input_image)?;
+
+    // Parse and calculate the img2img strength
+    let img2img_strength = match i2i_strength {
+        None => 0.75f64,
+        Some(strength) => (strength as f64) / 100.0f64,
+    };
+
+    // Create LCM guidance scale embeddings
+    let embedding_dim = 256;
+    let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(
+        guidance_scale, 
+        embedding_dim, 
+        &configs.device, 
+        configs.dtype
+    )?;
+
+    // Generate latents from input image using the LCM approach
+    println!("Generating latents from input image...");
+    let init_latents = (init_latent_dist.sample()? * vae_scale)?;
+    println!("Initial latents shape: {:?}", init_latents.shape());
+
+    // Calculate timesteps for LCM with proper img2img handling
+    let timesteps = LCMScheduler::get_timesteps_for_steps(configs.scheduler_steps, img2img_strength);
+    let t_start = timesteps[0]; // First denoising step
+
+    // Add noise at the appropriate timestep for img2img
+    println!("Adding noise to latents for t_start={}", t_start);
+    let noise = init_latents.randn_like(0f64, 1f64)?;
+    let latents = scheduler.add_noise(&init_latents, noise, t_start)?;
+
+    println!("Latents initialized successfully");
+    println!("Initial noise shape: {:?}", latents.shape());
+
+    println!("Starting diffusion process...");
+    
+    let mut latents = latents;
+    let timesteps: Vec<_> = scheduler.timesteps().iter().copied().collect();
+
+    for (timestep_index, &timestep) in timesteps.iter().enumerate() {
+        println!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
+
+        let latent_model_input = if use_guide_scale {
+            Tensor::cat(&[&latents, &latents], 0)?
+        } else {
+            latents.clone()
+        };
+
+        let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
+
+        // Use the guidance scale embedding in the UNet inference with the proper method name
+        let mut noise_pred = match unet.forward_with_guidance(
+            &latent_model_input, 
+            timestep as f64, 
+            &text_embeddings,
+            &guidance_scale_embedding) {
+            Ok(pred) => pred,
+            Err(e) => {
+                println!("UNet inference failed with error: {}", e);
+                return Err(anyhow::anyhow!("UNet inference failed: {}", e));
+            },
+        };
+
+        if use_guide_scale {
+            let chunks = noise_pred.chunk(2, 0)?;
+            let (noise_pred_uncond, noise_pred_text) = (&chunks[0], &chunks[1]);
+            noise_pred = (noise_pred_uncond + (guidance_scale * (noise_pred_text - noise_pred_uncond)?)?)?;
+        }
+
+        // Apply the LCM scheduler step
+        latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, configs.scheduler_steps)?;
+    }
+
+    println!("Diffusion process completed, decoding image...");
+    let image = vae.decode(&(latents / vae_scale)?)?;
+    
+    println!("VAE decode completed");
+
+    println!("Scaling image back");
+    let image = ((image / 2.)? + 0.5)?;
+    
+    println!("Normalized image values");
+    let image = (image.clamp(0f32, 1.)? * 255.)?;
+
+    println!("Convert to int8");
+    let image = image.to_dtype(DType::U8)?;
+    
+    println!("Converted to 8-bit format");
+
+    let image = tensor_to_image_buffer(&image.i(0)?)?;
+
+    Ok(image)
+}
diff --git a/crates/desktop/src-tauri/src/ml/stable_diffusion/mod.rs b/crates/desktop/src-tauri/src/ml/stable_diffusion/mod.rs
index d07f90d6e6..67e6759bd0 100644
--- a/crates/desktop/src-tauri/src/ml/stable_diffusion/mod.rs
+++ b/crates/desktop/src-tauri/src/ml/stable_diffusion/mod.rs
@@ -1,3 +1,4 @@
 pub mod get_vae_scale;
 pub mod infer_clip_text_embeddings;
-pub mod stable_diffusion_pipeline;
\ No newline at end of file
+pub mod stable_diffusion_pipeline;
+pub mod lcm_pipeline;
\ No newline at end of file
diff --git a/crates/desktop/src-tauri/src/state/app_config.rs b/crates/desktop/src-tauri/src/state/app_config.rs
index 0d22f1c5fd..27a9d65f95 100644
--- a/crates/desktop/src-tauri/src/state/app_config.rs
+++ b/crates/desktop/src-tauri/src/state/app_config.rs
@@ -71,7 +71,7 @@ impl AppConfig {
       sd_config,
       image_height: yaml_configs.image_height.unwrap_or(DEFAULT_SD_IMAGE_HEIGHT),
       image_width: yaml_configs.image_width.unwrap_or(DEFAULT_SD_IMAGE_WIDTH),
-      scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(1),
+      scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(4),
       scheduler_samples: yaml_configs.scheduler_samples.unwrap_or(15),
       seed: yaml_configs.seed,
       cfg_scale: yaml_configs.cfg_scale,
diff --git a/crates/vendor/candle/candle-examples/examples/lcm/main.rs b/crates/vendor/candle/candle-examples/examples/lcm/main.rs
index 7319f7d584..746198475f 100644
--- a/crates/vendor/candle/candle-examples/examples/lcm/main.rs
+++ b/crates/vendor/candle/candle-examples/examples/lcm/main.rs
@@ -5,6 +5,7 @@ extern crate accelerate_src;
 extern crate intel_mkl_src;
 
 use candle_transformers::models::stable_diffusion;
+use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
 use std::ops::Div;
 
 use anyhow::{Error as E, Result};
@@ -18,855 +19,1048 @@ use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel
 use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModelConfig;
 use candle_transformers::models::stable_diffusion::unet_2d::BlockConfig;
 
-const PRINT_DEBUG_TENSORS: bool = false;
-
 #[derive(Parser)]
 #[command(author, version, about, long_about = None)]
 struct Args {
-  /// The prompt to be used for image generation.
-  #[arg(long, default_value = "A very realistic photo of a rusty robot walking on a sandy beach")]
-  prompt: String,
+    /// The prompt to be used for image generation.
+    #[arg(
+        long,
+        default_value = "A very realistic photo of a rusty robot walking on a sandy beach"
+    )]
+    prompt: String,
 
-  #[arg(long, default_value = "")]
-  uncond_prompt: String,
+    #[arg(long, default_value = "")]
+    uncond_prompt: String,
 
-  /// Run on CPU rather than on GPU.
-  #[arg(long)]
-  cpu: bool,
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
 
-  /// Enable tracing (generates a trace-timestamp.json file).
-  #[arg(long)]
-  tracing: bool,
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
 
-  /// The height in pixels of the generated image.
-  #[arg(long)]
-  height: Option<usize>,
+    /// The height in pixels of the generated image.
+    #[arg(long)]
+    height: Option<usize>,
 
-  /// The width in pixels of the generated image.
-  #[arg(long)]
-  width: Option<usize>,
+    /// The width in pixels of the generated image.
+    #[arg(long)]
+    width: Option<usize>,
 
-  /// The UNet weight file, in .safetensors format.
-  #[arg(long, value_name = "FILE")]
-  unet_weights: Option<String>,
+    /// The UNet weight file, in .safetensors format.
+    #[arg(long, value_name = "FILE")]
+    unet_weights: Option<String>,
 
-  /// The CLIP weight file, in .safetensors format.
-  #[arg(long, value_name = "FILE")]
-  clip_weights: Option<String>,
+    /// The CLIP weight file, in .safetensors format.
+    #[arg(long, value_name = "FILE")]
+    clip_weights: Option<String>,
 
-  /// The CLIP2 weight file, in .safetensors format.
-  #[arg(long, value_name = "FILE")]
-  clip2_weights: Option<String>,
+    /// The CLIP2 weight file, in .safetensors format.
+    #[arg(long, value_name = "FILE")]
+    clip2_weights: Option<String>,
 
-  /// The VAE weight file, in .safetensors format.
-  #[arg(long, value_name = "FILE")]
-  vae_weights: Option<String>,
+    /// The VAE weight file, in .safetensors format.
+    #[arg(long, value_name = "FILE")]
+    vae_weights: Option<String>,
 
-  #[arg(long, value_name = "FILE")]
-  /// The file specifying the tokenizer to used for tokenization.
-  tokenizer: Option<String>,
+    #[arg(long, value_name = "FILE")]
+    /// The file specifying the tokenizer to used for tokenization.
+    tokenizer: Option<String>,
 
-  /// The size of the sliced attention or 0 for automatic slicing (disabled by default)
-  #[arg(long)]
-  sliced_attention_size: Option<usize>,
+    /// The size of the sliced attention or 0 for automatic slicing (disabled by default)
+    #[arg(long)]
+    sliced_attention_size: Option<usize>,
 
-  /// The number of inference steps to run.
-  #[arg(long, default_value_t = 4)]
-  n_steps: usize,
+    /// The number of inference steps to run.
+    #[arg(long, default_value_t = 4)]
+    n_steps: usize,
 
-  /// The number of samples to generate iteratively.
-  #[arg(long, default_value_t = 1)]
-  num_samples: usize,
+    /// The number of samples to generate iteratively.
+    #[arg(long, default_value_t = 1)]
+    num_samples: usize,
 
-  /// The numbers of samples to generate simultaneously.
-  #[arg[long, default_value_t = 1]]
-  bsize: usize,
+    /// The numbers of samples to generate simultaneously.
+    #[arg[long, default_value_t = 1]]
+    bsize: usize,
 
-  /// The name of the final image to generate.
-  #[arg(long, value_name = "FILE", default_value = "sd_final.png")]
-  final_image: String,
+    /// The name of the final image to generate.
+    #[arg(long, value_name = "FILE", default_value = "sd_final.png")]
+    final_image: String,
 
-  #[arg(long, value_enum, default_value = "v2-1")]
-  sd_version: StableDiffusionVersion,
+    #[arg(long, value_enum, default_value = "v2-1")]
+    sd_version: StableDiffusionVersion,
 
-  /// Generate intermediary images at each step.
-  #[arg(long, action)]
-  intermediary_images: bool,
+    /// Generate intermediary images at each step.
+    #[arg(long, action)]
+    intermediary_images: bool,
 
-  #[arg(long)]
-  use_flash_attn: bool,
+    #[arg(long)]
+    use_flash_attn: bool,
 
-  #[arg(long)]
-  use_f16: bool,
+    #[arg(long)]
+    use_f16: bool,
 
-  /// The guidance scale for classifier-free guidance.
-  #[arg(long, default_value_t = 2.0)]
-  guidance_scale: f64,
+    /// The guidance scale for classifier-free guidance.
+    #[arg(long, default_value_t = 7.5)]
+    guidance_scale: f64,
 
-  /// Path to the mask image for inpainting.
-  #[arg(long, value_name = "FILE")]
-  mask_path: Option<String>,
+    /// Path to the mask image for inpainting.
+    #[arg(long, value_name = "FILE")]
+    mask_path: Option<String>,
 
-  /// Path to the image used to initialize the latents. For inpainting, this is the image to be masked.
-  #[arg(long, value_name = "FILE")]
-  img2img: Option<String>,
+    /// Path to the image used to initialize the latents. For inpainting, this is the image to be masked.
+    #[arg(long, value_name = "FILE")]
+    img2img: Option<String>,
 
-  /// The strength parameter for img2img, between 0 and 1.
-  #[arg(long, default_value_t = 0.7)]
-  strength: f64,
+    /// The strength parameter for img2img, between 0 and 1.
+    #[arg(long, default_value_t = 0.7)]
+    strength: f64,
 
-  /// The seed to use when generating random samples.
-  #[arg(long)]
-  seed: Option<u64>,
+    /// The seed to use when generating random samples.
+    #[arg(long)]
+    seed: Option<u64>,
 
-  /// Force the saved image to update only the masked region
-  #[arg(long)]
-  only_update_masked: bool,
-
-  #[arg(long)]
-  save_tensors: bool,
-
-  /// Use half precision (F16) rather than F32
-  #[arg(long)]
-  half_precision: bool,
+    /// Force the saved image to update only the masked region
+    #[arg(long)]
+    only_update_masked: bool,
 }
 
 #[derive(Debug, Clone, Copy, clap::ValueEnum, PartialEq, Eq)]
 enum StableDiffusionVersion {
-  V1_5,
-  V1_5Inpaint,
-  V2_1,
-  V2Inpaint,
-  Xl,
-  XlInpaint,
-  Turbo,
+    V1_5,
+    V1_5Inpaint,
+    V2_1,
+    V2Inpaint,
+    Xl,
+    XlInpaint,
+    Turbo,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 enum ModelFile {
-  Tokenizer,
-  Tokenizer2,
-  Clip,
-  Clip2,
-  Unet,
-  UnetLcm,
-  Vae,
+    Tokenizer,
+    Tokenizer2,
+    Clip,
+    Clip2,
+    Unet,
+    UnetLcm,
+    Vae,
 }
 
 impl StableDiffusionVersion {
-  fn repo(&self) -> &'static str {
-    match self {
-      Self::XlInpaint => "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
-      Self::Xl => "stabilityai/stable-diffusion-xl-base-1.0",
-      Self::V2Inpaint => "stabilityai/stable-diffusion-2-inpainting",
-      Self::V2_1 => "stabilityai/stable-diffusion-2-1",
-      Self::V1_5 => "runwayml/stable-diffusion-v1-5",
-      Self::V1_5Inpaint => "stable-diffusion-v1-5/stable-diffusion-inpainting",
-      Self::Turbo => "stabilityai/sdxl-turbo",
-    }
-  }
-
-  fn unet_file(&self, use_f16: bool) -> &'static str {
-    match self {
-      Self::V1_5 | Self::V1_5Inpaint | Self::V2_1 | Self::V2Inpaint | Self::Xl | Self::XlInpaint | Self::Turbo => {
-        if use_f16 {
-          "unet/diffusion_pytorch_model.fp16.safetensors"
-        } else {
-          "unet/diffusion_pytorch_model.safetensors"
+    fn repo(&self) -> &'static str {
+        match self {
+            Self::XlInpaint => "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
+            Self::Xl => "stabilityai/stable-diffusion-xl-base-1.0",
+            Self::V2Inpaint => "stabilityai/stable-diffusion-2-inpainting",
+            Self::V2_1 => "stabilityai/stable-diffusion-2-1",
+            Self::V1_5 => "runwayml/stable-diffusion-v1-5",
+            Self::V1_5Inpaint => "stable-diffusion-v1-5/stable-diffusion-inpainting",
+            Self::Turbo => "stabilityai/sdxl-turbo",
         }
-      },
     }
-  }
 
-  fn vae_file(&self, use_f16: bool) -> &'static str {
-    match self {
-      Self::V1_5 | Self::V1_5Inpaint | Self::V2_1 | Self::V2Inpaint | Self::Xl | Self::XlInpaint | Self::Turbo => {
-        if use_f16 {
-          "vae/diffusion_pytorch_model.fp16.safetensors"
-        } else {
-          "vae/diffusion_pytorch_model.safetensors"
+    fn unet_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "unet/diffusion_pytorch_model.fp16.safetensors"
+                } else {
+                    "unet/diffusion_pytorch_model.safetensors"
+                }
+            }
         }
-      },
     }
-  }
 
-  fn clip_file(&self, use_f16: bool) -> &'static str {
-    match self {
-      Self::V1_5 | Self::V1_5Inpaint | Self::V2_1 | Self::V2Inpaint | Self::Xl | Self::XlInpaint | Self::Turbo => {
-        if use_f16 {
-          "text_encoder/model.fp16.safetensors"
-        } else {
-          "text_encoder/model.safetensors"
+    fn vae_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "vae/diffusion_pytorch_model.fp16.safetensors"
+                } else {
+                    "vae/diffusion_pytorch_model.safetensors"
+                }
+            }
         }
-      },
     }
-  }
 
-  fn clip2_file(&self, use_f16: bool) -> &'static str {
-    match self {
-      Self::V1_5 | Self::V1_5Inpaint | Self::V2_1 | Self::V2Inpaint | Self::Xl | Self::XlInpaint | Self::Turbo => {
-        if use_f16 {
-          "text_encoder_2/model.fp16.safetensors"
-        } else {
-          "text_encoder_2/model.safetensors"
+    fn clip_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "text_encoder/model.fp16.safetensors"
+                } else {
+                    "text_encoder/model.safetensors"
+                }
+            }
+        }
+    }
+
+    fn clip2_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "text_encoder_2/model.fp16.safetensors"
+                } else {
+                    "text_encoder_2/model.safetensors"
+                }
+            }
         }
-      },
     }
-  }
 }
 
 impl ModelFile {
-  fn get(&self, filename: Option<String>, version: StableDiffusionVersion, use_f16: bool) -> Result<std::path::PathBuf> {
-    use hf_hub::api::sync::Api;
-    match filename {
-      Some(filename) => Ok(std::path::PathBuf::from(filename)),
-      None => {
-        let (repo, path) = match self {
-          Self::Tokenizer => {
-            let tokenizer_repo = match version {
-              StableDiffusionVersion::V1_5 | StableDiffusionVersion::V2_1 | StableDiffusionVersion::V1_5Inpaint | StableDiffusionVersion::V2Inpaint => "openai/clip-vit-base-patch32",
-              StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint | StableDiffusionVersion::Turbo => {
-                // This seems similar to the patch32 version except some very small
-                // difference in the split regex.
-                "openai/clip-vit-large-patch14"
-              },
-            };
-            (tokenizer_repo, "tokenizer.json")
-          },
-          Self::Tokenizer2 => ("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", "tokenizer.json"),
-          Self::Clip => (version.repo(), version.clip_file(use_f16)),
-          Self::Clip2 => (version.repo(), version.clip2_file(use_f16)),
-          Self::UnetLcm => ("SimianLuo/LCM_Dreamshaper_v7", if use_f16 { "unet/diffusion_pytorch_model.fp16.safetensors" } else { "unet/diffusion_pytorch_model.safetensors" }),
-          Self::Unet => (version.repo(), version.unet_file(use_f16)),
-          Self::Vae => {
-            // Override for SDXL when using f16 weights.
-            // See https://github.com/huggingface/candle/issues/1060
-            if matches!(version, StableDiffusionVersion::Xl | StableDiffusionVersion::Turbo,) && use_f16 {
-              ("madebyollin/sdxl-vae-fp16-fix", "diffusion_pytorch_model.safetensors")
-            } else {
-              (version.repo(), version.vae_file(use_f16))
+    fn get(
+        &self,
+        filename: Option<String>,
+        version: StableDiffusionVersion,
+        use_f16: bool,
+    ) -> Result<std::path::PathBuf> {
+        use hf_hub::api::sync::Api;
+        match filename {
+            Some(filename) => Ok(std::path::PathBuf::from(filename)),
+            None => {
+                let (repo, path) = match self {
+                    Self::Tokenizer => {
+                        let tokenizer_repo = match version {
+                            StableDiffusionVersion::V1_5
+                            | StableDiffusionVersion::V2_1
+                            | StableDiffusionVersion::V1_5Inpaint
+                            | StableDiffusionVersion::V2Inpaint => "openai/clip-vit-base-patch32",
+                            StableDiffusionVersion::Xl
+                            | StableDiffusionVersion::XlInpaint
+                            | StableDiffusionVersion::Turbo => {
+                                // This seems similar to the patch32 version except some very small
+                                // difference in the split regex.
+                                "openai/clip-vit-large-patch14"
+                            }
+                        };
+                        (tokenizer_repo, "tokenizer.json")
+                    }
+                    Self::Tokenizer2 => {
+                        ("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", "tokenizer.json")
+                    }
+                    Self::Clip => (version.repo(), version.clip_file(use_f16)),
+                    Self::Clip2 => (version.repo(), version.clip2_file(use_f16)),
+                    Self::UnetLcm => (
+                        "SimianLuo/LCM_Dreamshaper_v7",
+                        if use_f16 {
+                            "unet/diffusion_pytorch_model.fp16.safetensors"
+                        } else {
+                            "unet/diffusion_pytorch_model.safetensors"
+                        },
+                    ),
+                    Self::Unet => (version.repo(), version.unet_file(use_f16)),
+                    Self::Vae => {
+                        // Override for SDXL when using f16 weights.
+                        // See https://github.com/huggingface/candle/issues/1060
+                        if matches!(
+                            version,
+                            StableDiffusionVersion::Xl | StableDiffusionVersion::Turbo,
+                        ) && use_f16
+                        {
+                            (
+                                "madebyollin/sdxl-vae-fp16-fix",
+                                "diffusion_pytorch_model.safetensors",
+                            )
+                        } else {
+                            (version.repo(), version.vae_file(use_f16))
+                        }
+                    }
+                };
+                let filename = Api::new()?.model(repo.to_string()).get(path)?;
+                Ok(filename)
             }
-          },
-        };
-        let filename = Api::new()?.model(repo.to_string()).get(path)?;
-        Ok(filename)
-      },
+        }
     }
-  }
 }
 
-fn output_filename(basename: &str, sample_idx: usize, num_samples: usize, timestep_idx: Option<usize>) -> String {
-  let filename = if num_samples > 1 {
-    match basename.rsplit_once('.') {
-      None => format!("{basename}.{sample_idx}.png"),
-      Some((filename_no_extension, extension)) => {
-        format!("{filename_no_extension}.{sample_idx}.{extension}")
-      },
+fn output_filename(
+    basename: &str,
+    sample_idx: usize,
+    num_samples: usize,
+    timestep_idx: Option<usize>,
+) -> String {
+    let filename = if num_samples > 1 {
+        match basename.rsplit_once('.') {
+            None => format!("{basename}.{sample_idx}.png"),
+            Some((filename_no_extension, extension)) => {
+                format!("{filename_no_extension}.{sample_idx}.{extension}")
+            }
+        }
+    } else {
+        basename.to_string()
+    };
+    match timestep_idx {
+        None => filename,
+        Some(timestep_idx) => match filename.rsplit_once('.') {
+            None => format!("{filename}-{timestep_idx}.png"),
+            Some((filename_no_extension, extension)) => {
+                format!("{filename_no_extension}-{timestep_idx}.{extension}")
+            }
+        },
     }
-  } else {
-    basename.to_string()
-  };
-  match timestep_idx {
-    None => filename,
-    Some(timestep_idx) => match filename.rsplit_once('.') {
-      None => format!("{filename}-{timestep_idx}.png"),
-      Some((filename_no_extension, extension)) => {
-        format!("{filename_no_extension}-{timestep_idx}.{extension}")
-      },
-    },
-  }
 }
 
 #[allow(clippy::too_many_arguments)]
-fn save_image(vae: &AutoEncoderKL, latents: &Tensor, vae_scale: f64, bsize: usize, idx: usize, final_image: &str, num_samples: usize, timestep_ids: Option<usize>) -> Result<()> {
-  let images = vae.decode(&(latents / vae_scale)?)?;
-  let images = ((images / 2.)? + 0.5)?.to_device(&Device::Cpu)?;
-  let images = (images.clamp(0f32, 1.)? * 255.)?.to_dtype(DType::U8)?;
-  for batch in 0..bsize {
-    let image = images.i(batch)?;
-    let image_filename = output_filename(final_image, (bsize * idx) + batch + 1, batch + num_samples, timestep_ids);
-    candle_examples::save_image(&image, image_filename)?;
-  }
-  Ok(())
+fn save_image(
+    vae: &AutoEncoderKL,
+    latents: &Tensor,
+    vae_scale: f64,
+    bsize: usize,
+    idx: usize,
+    final_image: &str,
+    num_samples: usize,
+    timestep_ids: Option<usize>,
+) -> Result<()> {
+    let images = vae.decode(&(latents / vae_scale)?)?;
+    let images = ((images / 2.)? + 0.5)?.to_device(&Device::Cpu)?;
+    let images = (images.clamp(0f32, 1.)? * 255.)?.to_dtype(DType::U8)?;
+    for batch in 0..bsize {
+        let image = images.i(batch)?;
+        let image_filename = output_filename(
+            final_image,
+            (bsize * idx) + batch + 1,
+            batch + num_samples,
+            timestep_ids,
+        );
+        candle_examples::save_image(&image, image_filename)?;
+    }
+    Ok(())
 }
 
 #[allow(clippy::too_many_arguments)]
-fn text_embeddings(prompt: &str, uncond_prompt: &str, tokenizer: Option<String>, clip_weights: Option<String>, clip2_weights: Option<String>, sd_version: StableDiffusionVersion, sd_config: &stable_diffusion::StableDiffusionConfig, use_f16: bool, device: &Device, dtype: DType, use_guide_scale: bool, first: bool) -> Result<Tensor> {
-  let tokenizer_file = if first { ModelFile::Tokenizer } else { ModelFile::Tokenizer2 };
-  let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
-  let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
-  let pad_id = match &sd_config.clip.pad_with {
-    Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
-    None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
-  };
-  println!("Running with prompt \"{prompt}\".");
-  let mut tokens = tokenizer.encode(prompt, true).map_err(E::msg)?.get_ids().to_vec();
-  if tokens.len() > sd_config.clip.max_position_embeddings {
-    anyhow::bail!("the prompt is too long, {} > max-tokens ({})", tokens.len(), sd_config.clip.max_position_embeddings)
-  }
-  while tokens.len() < sd_config.clip.max_position_embeddings {
-    tokens.push(pad_id)
-  }
-  let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
-
-  println!("Building the Clip transformer.");
-  let clip_weights_file = if first { ModelFile::Clip } else { ModelFile::Clip2 };
-  let clip_weights = if first { clip_weights_file.get(clip_weights, sd_version, use_f16)? } else { clip_weights_file.get(clip2_weights, sd_version, use_f16)? };
-  let clip_config = if first { &sd_config.clip } else { sd_config.clip2.as_ref().unwrap() };
-  let text_model = stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
-  let text_embeddings = text_model.forward(&tokens)?;
-
-  let text_embeddings = if use_guide_scale {
-    let mut uncond_tokens = tokenizer.encode(uncond_prompt, true).map_err(E::msg)?.get_ids().to_vec();
-    if uncond_tokens.len() > sd_config.clip.max_position_embeddings {
-      anyhow::bail!("the negative prompt is too long, {} > max-tokens ({})", uncond_tokens.len(), sd_config.clip.max_position_embeddings)
+fn text_embeddings(
+    prompt: &str,
+    uncond_prompt: &str,
+    tokenizer: Option<String>,
+    clip_weights: Option<String>,
+    clip2_weights: Option<String>,
+    sd_version: StableDiffusionVersion,
+    sd_config: &stable_diffusion::StableDiffusionConfig,
+    use_f16: bool,
+    device: &Device,
+    dtype: DType,
+    use_guide_scale: bool,
+    first: bool,
+) -> Result<Tensor> {
+    let tokenizer_file = if first {
+        ModelFile::Tokenizer
+    } else {
+        ModelFile::Tokenizer2
+    };
+    let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
+    let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
+    let pad_id = match &sd_config.clip.pad_with {
+        Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
+        None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
+    };
+    println!("Running with prompt \"{prompt}\".");
+    let mut tokens = tokenizer
+        .encode(prompt, true)
+        .map_err(E::msg)?
+        .get_ids()
+        .to_vec();
+    if tokens.len() > sd_config.clip.max_position_embeddings {
+        anyhow::bail!(
+            "the prompt is too long, {} > max-tokens ({})",
+            tokens.len(),
+            sd_config.clip.max_position_embeddings
+        )
     }
-    while uncond_tokens.len() < sd_config.clip.max_position_embeddings {
-      uncond_tokens.push(pad_id)
+    while tokens.len() < sd_config.clip.max_position_embeddings {
+        tokens.push(pad_id)
     }
+    let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
 
-    let uncond_tokens = Tensor::new(uncond_tokens.as_slice(), device)?.unsqueeze(0)?;
-    let uncond_embeddings = text_model.forward(&uncond_tokens)?;
+    println!("Building the Clip transformer.");
+    let clip_weights_file = if first {
+        ModelFile::Clip
+    } else {
+        ModelFile::Clip2
+    };
+    let clip_weights = if first {
+        clip_weights_file.get(clip_weights, sd_version, use_f16)?
+    } else {
+        clip_weights_file.get(clip2_weights, sd_version, use_f16)?
+    };
+    let clip_config = if first {
+        &sd_config.clip
+    } else {
+        sd_config.clip2.as_ref().unwrap()
+    };
+    let text_model =
+        stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
+    let text_embeddings = text_model.forward(&tokens)?;
 
-    Tensor::cat(&[uncond_embeddings, text_embeddings], 0)?.to_dtype(dtype)?
-  } else {
-    text_embeddings.to_dtype(dtype)?
-  };
-  Ok(text_embeddings)
+    let text_embeddings = if use_guide_scale {
+        let mut uncond_tokens = tokenizer
+            .encode(uncond_prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        if uncond_tokens.len() > sd_config.clip.max_position_embeddings {
+            anyhow::bail!(
+                "the negative prompt is too long, {} > max-tokens ({})",
+                uncond_tokens.len(),
+                sd_config.clip.max_position_embeddings
+            )
+        }
+        while uncond_tokens.len() < sd_config.clip.max_position_embeddings {
+            uncond_tokens.push(pad_id)
+        }
+
+        let uncond_tokens = Tensor::new(uncond_tokens.as_slice(), device)?.unsqueeze(0)?;
+        let uncond_embeddings = text_model.forward(&uncond_tokens)?;
+
+        Tensor::cat(&[uncond_embeddings, text_embeddings], 0)?.to_dtype(dtype)?
+    } else {
+        text_embeddings.to_dtype(dtype)?
+    };
+    Ok(text_embeddings)
 }
 
 fn image_preprocess<T: AsRef<std::path::Path>>(path: T) -> anyhow::Result<Tensor> {
-  let img = image::ImageReader::open(path)?.decode()?;
-  let (height, width) = (img.height() as usize, img.width() as usize);
-  let height = height - height % 32;
-  let width = width - width % 32;
-  let img = img.resize_to_fill(width as u32, height as u32, image::imageops::FilterType::CatmullRom);
-  let img = img.to_rgb8();
-  let img = img.into_raw();
-  let img = Tensor::from_vec(img, (height, width, 3), &Device::Cpu)?.permute((2, 0, 1))?.to_dtype(DType::F32)?.affine(2. / 255., -1.)?.unsqueeze(0)?;
-  Ok(img)
+    let img = image::ImageReader::open(path)?.decode()?;
+    let (height, width) = (img.height() as usize, img.width() as usize);
+    let height = height - height % 32;
+    let width = width - width % 32;
+    let img = img.resize_to_fill(
+        width as u32,
+        height as u32,
+        image::imageops::FilterType::CatmullRom,
+    );
+    let img = img.to_rgb8();
+    let img = img.into_raw();
+    let img = Tensor::from_vec(img, (height, width, 3), &Device::Cpu)?
+        .permute((2, 0, 1))?
+        .to_dtype(DType::F32)?
+        .affine(2. / 255., -1.)?
+        .unsqueeze(0)?;
+    Ok(img)
 }
 
 /// Convert the mask image to a single channel tensor. Also ensure the image is a multiple of 32 in both dimensions.
 fn mask_preprocess<T: AsRef<std::path::Path>>(path: T) -> anyhow::Result<Tensor> {
-  let img = image::open(path)?.to_luma8();
-  let (new_width, new_height) = {
-    let (width, height) = img.dimensions();
-    (width - width % 32, height - height % 32)
-  };
-  let img = image::imageops::resize(&img, new_width, new_height, image::imageops::FilterType::CatmullRom).into_raw();
-  let mask = Tensor::from_vec(img, (new_height as usize, new_width as usize), &Device::Cpu)?.unsqueeze(0)?.to_dtype(DType::F32)?.div(255.0)?.unsqueeze(0)?;
-  Ok(mask)
+    let img = image::open(path)?.to_luma8();
+    let (new_width, new_height) = {
+        let (width, height) = img.dimensions();
+        (width - width % 32, height - height % 32)
+    };
+    let img = image::imageops::resize(
+        &img,
+        new_width,
+        new_height,
+        image::imageops::FilterType::CatmullRom,
+    )
+    .into_raw();
+    let mask = Tensor::from_vec(img, (new_height as usize, new_width as usize), &Device::Cpu)?
+        .unsqueeze(0)?
+        .to_dtype(DType::F32)?
+        .div(255.0)?
+        .unsqueeze(0)?;
+    Ok(mask)
 }
 
 /// Generates the mask latents, scaled mask and mask_4 for inpainting. Returns a tuple of None if inpainting is not
 /// being used.
 #[allow(clippy::too_many_arguments)]
-fn inpainting_tensors(sd_version: StableDiffusionVersion, mask_path: Option<String>, dtype: DType, device: &Device, use_guide_scale: bool, vae: &AutoEncoderKL, image: Option<Tensor>, vae_scale: f64) -> Result<(Option<Tensor>, Option<Tensor>, Option<Tensor>)> {
-  match sd_version {
-    StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => {
-      let inpaint_mask = mask_path.ok_or_else(|| anyhow::anyhow!("An inpainting model was requested but mask-path is not provided."))?;
-      // Get the mask image with shape [1, 1, 128, 128]
-      let mask = mask_preprocess(inpaint_mask)?.to_device(device)?.to_dtype(dtype)?;
-      // Generate the masked image from the image and the mask with shape [1, 3, 1024, 1024]
-      let xmask = mask.le(0.5)?.repeat(&[1, 3, 1, 1])?.to_dtype(dtype)?;
-      let image = &image.ok_or_else(|| anyhow::anyhow!("An inpainting model was requested but img2img which is used as the input image is not provided."))?;
-      let masked_img = (image * xmask)?;
-      // Scale down the mask
-      let shape = masked_img.shape();
-      let (w, h) = (shape.dims()[3] / 8, shape.dims()[2] / 8);
-      let mask = mask.interpolate2d(w, h)?;
-      // shape: [1, 4, 128, 128]
-      let mask_latents = vae.encode(&masked_img)?;
-      let mask_latents = (mask_latents.sample()? * vae_scale)?.to_device(device)?;
+fn inpainting_tensors(
+    sd_version: StableDiffusionVersion,
+    mask_path: Option<String>,
+    dtype: DType,
+    device: &Device,
+    use_guide_scale: bool,
+    vae: &AutoEncoderKL,
+    image: Option<Tensor>,
+    vae_scale: f64,
+) -> Result<(Option<Tensor>, Option<Tensor>, Option<Tensor>)> {
+    match sd_version {
+        StableDiffusionVersion::XlInpaint
+        | StableDiffusionVersion::V2Inpaint
+        | StableDiffusionVersion::V1_5Inpaint => {
+            let inpaint_mask = mask_path.ok_or_else(|| {
+                anyhow::anyhow!("An inpainting model was requested but mask-path is not provided.")
+            })?;
+            // Get the mask image with shape [1, 1, 128, 128]
+            let mask = mask_preprocess(inpaint_mask)?
+                .to_device(device)?
+                .to_dtype(dtype)?;
+            // Generate the masked image from the image and the mask with shape [1, 3, 1024, 1024]
+            let xmask = mask.le(0.5)?.repeat(&[1, 3, 1, 1])?.to_dtype(dtype)?;
+            let image = &image
+                .ok_or_else(|| anyhow::anyhow!(
+                    "An inpainting model was requested but img2img which is used as the input image is not provided."
+                ))?;
+            let masked_img = (image * xmask)?;
+            // Scale down the mask
+            let shape = masked_img.shape();
+            let (w, h) = (shape.dims()[3] / 8, shape.dims()[2] / 8);
+            let mask = mask.interpolate2d(w, h)?;
+            // shape: [1, 4, 128, 128]
+            let mask_latents = vae.encode(&masked_img)?;
+            let mask_latents = (mask_latents.sample()? * vae_scale)?.to_device(device)?;
 
-      let mask_4 = mask.as_ref().repeat(&[1, 4, 1, 1])?;
-      let (mask_latents, mask) = if use_guide_scale { (Tensor::cat(&[&mask_latents, &mask_latents], 0)?, Tensor::cat(&[&mask, &mask], 0)?) } else { (mask_latents, mask) };
-      Ok((Some(mask_latents), Some(mask), Some(mask_4)))
-    },
-    _ => Ok((None, None, None)),
-  }
+            let mask_4 = mask.as_ref().repeat(&[1, 4, 1, 1])?;
+            let (mask_latents, mask) = if use_guide_scale {
+                (
+                    Tensor::cat(&[&mask_latents, &mask_latents], 0)?,
+                    Tensor::cat(&[&mask, &mask], 0)?,
+                )
+            } else {
+                (mask_latents, mask)
+            };
+            Ok((Some(mask_latents), Some(mask), Some(mask_4)))
+        }
+        _ => Ok((None, None, None)),
+    }
 }
 
 fn get_tensor_stats(tensor: &Tensor) -> Result<(f32, f32, f32, f32)> {
-  // Convert tensor to F32 for stats calculation - makes statistics more accurate anyway
-  let tensor_f32 = tensor.to_dtype(DType::F32)?;
-  let flat = tensor_f32.flatten_all()?;
-
-  let min = flat.min(0)?.to_scalar::<f32>()?;
-  let max = flat.max(0)?.to_scalar::<f32>()?;
-  let mean = flat.mean_all()?.to_scalar::<f32>()?;
-
-  // Create a tensor with the same shape as flat but filled with the mean value
-  let mean_tensor = Tensor::new(&[mean], flat.device())?;
-  let mean_tensor = mean_tensor.broadcast_as(flat.shape().dims())?;
-  let diff = (flat - mean_tensor)?;
-  let squared = diff.sqr()?;
-  let variance = squared.mean_all()?;
-  let std = variance.sqrt()?.to_scalar::<f32>()?;
-
-  Ok((min, max, mean, std))
-}
-
-// Add this function to track statistics at key points
-fn log_tensor_stats(name: &str, tensor: &Tensor) -> Result<()> {
-  let (min, max, mean, std) = get_tensor_stats(tensor)?;
-  println!("{} stats: min={:.5}, max={:.5}, mean={:.5}, std={:.5}", name, min, max, mean, std);
-  Ok(())
-}
-
-// Add this simpler alternative function
-fn save_tensor_as_numpy(tensor: &Tensor, filename: &str) -> Result<()> {
-  use std::fs::File;
-  use std::io::Write;
-
-  // Move tensor to CPU and convert to f32 for consistent serialization
-  let tensor_cpu = tensor.to_device(&Device::Cpu)?.to_dtype(DType::F32)?;
-  let data = tensor_cpu.to_vec1::<f32>()?;
-
-  // Create a simple binary file with the tensor data
-  let mut file = File::create(filename)?;
-
-  // Write shape information first (as u32 values)
-  let shape = tensor.shape().dims();
-  file.write_all(&(shape.len() as u32).to_le_bytes())?;
-  for &dim in shape {
-    file.write_all(&(dim as u32).to_le_bytes())?;
-  }
-
-  // Write all float values
-  for val in data {
-    file.write_all(&val.to_le_bytes())?;
-  }
-
-  Ok(())
+    let flat = tensor.flatten_all()?;
+    let min = flat.min(0)?.to_scalar::<f32>()?;
+    let max = flat.max(0)?.to_scalar::<f32>()?;
+    let mean = flat.mean_all()?.to_scalar::<f32>()?;
+    
+    // Calculate standard deviation manually
+    // Create a tensor with the same shape as flat but filled with the mean value
+    let mean_tensor = Tensor::new(&[mean], flat.device())?;
+    let mean_tensor = mean_tensor.broadcast_as(flat.shape().dims())?;
+    let diff = (flat - mean_tensor)?;
+    let squared = diff.sqr()?;
+    let variance = squared.mean_all()?;
+    let std = variance.sqrt()?.to_scalar::<f32>()?;
+    
+    Ok((min, max, mean, std))
 }
 
 fn run(args: Args) -> Result<()> {
-  use tracing_chrome::ChromeLayerBuilder;
-  use tracing_subscriber::prelude::*;
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
 
-  let Args { prompt, uncond_prompt, cpu, height, width, n_steps, tokenizer, final_image, sliced_attention_size, num_samples, bsize, sd_version, clip_weights, clip2_weights, vae_weights, unet_weights, tracing, use_f16, guidance_scale, use_flash_attn, mask_path, img2img, strength, seed, only_update_masked, save_tensors, intermediary_images, half_precision } = args;
+    let Args {
+        prompt,
+        uncond_prompt,
+        cpu,
+        height,
+        width,
+        n_steps,
+        tokenizer,
+        final_image,
+        sliced_attention_size,
+        num_samples,
+        bsize,
+        sd_version,
+        clip_weights,
+        clip2_weights,
+        vae_weights,
+        unet_weights,
+        tracing,
+        use_f16,
+        guidance_scale,
+        use_flash_attn,
+        mask_path,
+        img2img,
+        strength,
+        seed,
+        ..
+    } = args;
 
-  if !(0. ..=1.).contains(&strength) {
-    anyhow::bail!("strength should be between 0 and 1, got {strength}")
-  }
-
-  let _guard = if tracing {
-    let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
-    tracing_subscriber::registry().with(chrome_layer).init();
-    Some(guard)
-  } else {
-    None
-  };
-
-  println!("=== LCM-Dreamshaper F16 DEBUG MODE ===");
-  println!("Using dtype: {:?}", if half_precision { DType::F16 } else { DType::F32 });
-
-  let dtype = if half_precision { DType::F16 } else { DType::F32 };
-  let sd_config = match sd_version {
-    StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => stable_diffusion::StableDiffusionConfig::v1_5(sliced_attention_size, height, width),
-    StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint => stable_diffusion::StableDiffusionConfig::v2_1(sliced_attention_size, height, width),
-    StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint => stable_diffusion::StableDiffusionConfig::sdxl(sliced_attention_size, height, width),
-    StableDiffusionVersion::Turbo => stable_diffusion::StableDiffusionConfig::sdxl_turbo(sliced_attention_size, height, width),
-  };
-
-  let mut scheduler = stable_diffusion::lcm::LCMScheduler::new(n_steps, stable_diffusion::lcm::LCMSchedulerConfig::default())?;
-  println!("Scheduler created with dtype={:?}", scheduler.dtype());
-
-  let device = candle_examples::device(cpu)?;
-  // If a seed is not given, generate a random seed and print it
-  let seed = seed.unwrap_or(rand::thread_rng().gen_range(0u64..u64::MAX));
-  println!("Using seed {seed}");
-  device.set_seed(seed)?;
-  let use_guide_scale = guidance_scale > 1.0;
-
-  let which = match sd_version {
-    StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint | StableDiffusionVersion::Turbo => vec![true, false],
-    _ => vec![true],
-  };
-
-  // Define the base model and LCM model repositories
-  let base_model_repo = match sd_version {
-    StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
-      // Dreamshaper is based on SD 1.5, so use it for those versions
-      "Lykon/dreamshaper-7"
-    },
-    _ => {
-      // For other versions, use the standard repo
-      sd_version.repo()
-    },
-  };
-  let lcm_model_repo = "SimianLuo/LCM_Dreamshaper_v7"; // LCM UNet
-
-  // Create a custom ModelFile enum variant for the base model
-  let clip_weights = if clip_weights.is_none() {
-    use hf_hub::api::sync::Api;
-    let api = Api::new()?;
-    let clip_path = api.model(base_model_repo.to_string()).get(if use_f16 { "text_encoder/model.fp16.safetensors" } else { "text_encoder/model.safetensors" })?;
-    Some(clip_path.to_string_lossy().to_string())
-  } else {
-    clip_weights
-  };
-
-  let vae_weights = if vae_weights.is_none() {
-    use hf_hub::api::sync::Api;
-    let api = Api::new()?;
-    let vae_path = api.model(base_model_repo.to_string()).get(if use_f16 { "vae/diffusion_pytorch_model.fp16.safetensors" } else { "vae/diffusion_pytorch_model.safetensors" })?;
-    Some(vae_path.to_string_lossy().to_string())
-  } else {
-    vae_weights
-  };
-
-  // For the UNet, always use the LCM model
-  let unet_weights = {
-    use hf_hub::api::sync::Api;
-    let api = Api::new()?;
-    let unet_path = api.model(lcm_model_repo.to_string()).get(if use_f16 {
-      if lcm_model_repo == "SimianLuo/LCM_Dreamshaper_v7" {
-        "unet/diffusion_pytorch_model.safetensors"
-      } else {
-        "unet/diffusion_pytorch_model.fp16.safetensors"
-      }
-    } else {
-      "unet/diffusion_pytorch_model.safetensors"
-    })?;
-    Some(unet_path.to_string_lossy().to_string())
-  };
-
-  // Load text embeddings from the base model
-  let text_embeddings = which.iter().map(|first| text_embeddings(&prompt, &uncond_prompt, tokenizer.clone(), clip_weights.clone(), clip2_weights.clone(), sd_version, &sd_config, use_f16, &device, dtype, use_guide_scale, *first)).collect::<Result<Vec<_>>>()?;
-
-  let text_embeddings = Tensor::cat(&text_embeddings, D::Minus1)?;
-  let text_embeddings = text_embeddings.repeat((bsize, 1, 1))?;
-  println!("Text embeddings shape: {:?}", text_embeddings.shape());
-  if PRINT_DEBUG_TENSORS {
-    println!("Text embeddings first few values: {}", text_embeddings.i(0)?.i(0)?.narrow(0, 0, 10)?);
-  }
-
-  // Replace the random projection with a fixed one to match diffusers
-  let text_embeddings = if text_embeddings.dim(2)? == 2048 {
-    println!("Projecting text embeddings from 2048 to 768 dimensions");
-
-    // Load the projection matrix from a file (you'll need to extract this from the diffusers model)
-    // For now, we'll still use a random one but with a fixed seed
-    device.set_seed(42)?; // Use a fixed seed for the projection
-    let projection = Tensor::randn(0.0, 0.02, (2048, 768), &device)?.to_dtype(dtype)?;
-
-    let proj_embeddings = text_embeddings.matmul(&projection)?;
-    proj_embeddings
-  } else {
-    text_embeddings
-  };
-  if PRINT_DEBUG_TENSORS {
-    println!("Projected text embeddings: {:?}", text_embeddings);
-  }
-
-  println!("Building the autoencoder.");
-  let vae_weights = ModelFile::Vae.get(vae_weights, sd_version, use_f16)?;
-  let vae = sd_config.build_vae(vae_weights, &device, dtype)?;
-
-  let (image, init_latent_dist) = match &img2img {
-    None => (None, None),
-    Some(image) => {
-      let image = image_preprocess(image)?.to_device(&device)?.to_dtype(dtype)?;
-      println!("Input image shape: {:?}", image.shape());
-      (Some(image.clone()), Some(vae.encode(&image)?))
-    },
-  };
-
-  println!(
-    "Latent shape after VAE encoding: {:?}",
-    init_latent_dist.as_ref().map(|dist| {
-      // Try to get shape from a temporary sample
-      match dist.sample() {
-        Ok(sample) => format!("{:?}", sample.shape()),
-        Err(_) => "Unable to determine shape".to_string(),
-      }
-    })
-  );
-
-  println!("Building the unet.");
-  let unet_weights = ModelFile::UnetLcm.get(unet_weights, sd_version, use_f16)?;
-  let in_channels = match sd_version {
-    StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => 9,
-    _ => 4,
-  };
-  println!("Loading UNet weights from: SimianLuo/LCM_Dreamshaper_v7/unet");
-
-  // Ensure we're loading the correct UNet configuration
-  println!("Building UNet for LCM (SimianLuo/LCM_Dreamshaper_v7)");
-
-  // For LCM models, ensure that:
-  // 1. We're using the correct prediction type (epsilon)
-  // 2. The cross-attention dimension is correct (768 for Dreamshaper)
-  // 3. The in_channels matches the expected input (4 for standard, 9 for inpainting)
-  let unet_config = UNet2DConditionModelConfig {
-    // LCM specific config values
-    cross_attention_dim: 768, // Dreamshaper v7 uses 768
-    // Rest of configuration...
-    center_input_sample: false,
-    flip_sin_to_cos: true,
-    freq_shift: 0.0,
-    blocks: vec![BlockConfig { out_channels: 320, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 640, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 1280, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 1280, use_cross_attn: None, attention_head_dim: 8 }],
-    layers_per_block: 2,
-    downsample_padding: 1,
-    mid_block_scale_factor: 1.0,
-    norm_num_groups: 32,
-    norm_eps: 1e-5,
-    sliced_attention_size: None,
-    use_linear_projection: false,
-  };
-
-  // Load the model directly from the safetensors file
-  let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_weights], dtype, &device)? };
-
-  let unet = UNet2DConditionModel::new(vs_unet, in_channels, 4, use_flash_attn, unet_config)?;
-
-  // Ensure strength is at least 0.4 for img2img to work well
-  let effective_strength = if img2img.is_some() {
-    strength.max(0.4).min(0.99) // Keep between 0.4 and 0.99
-  } else {
-    strength
-  };
-
-  // Make sure we're doing a proper image-to-image transform
-  // with the right amount of noise
-  println!("Using img2img with strength={:.4}", effective_strength);
-  let mut raw_t_start = n_steps - (n_steps as f64 * effective_strength).ceil() as usize;
-
-  let vae_scale = match sd_version {
-    StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint | StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::XlInpaint | StableDiffusionVersion::Xl => 0.18215,
-    StableDiffusionVersion::Turbo => 0.13025,
-  };
-
-  let (mask_latents, mask, mask_4) = inpainting_tensors(sd_version, mask_path, dtype, &device, use_guide_scale, &vae, image, vae_scale)?;
-
-  println!("Timesteps: {:?}", scheduler.timesteps());
-  // println!("Alpha cumprod: {:?}", scheduler.alphas_cumprod());
-
-  // Enable debug dumps to compare with diffusers
-  let enable_debug_dumps = true;
-  let debug_dir = "lcm_rust_debug";
-  if enable_debug_dumps {
-    std::fs::create_dir_all(debug_dir).expect("Failed to create debug directory");
-
-    // Log the scheduler configuration
-    let mut config_file = std::fs::File::create(format!("{}/scheduler_config.txt", debug_dir)).expect("Failed to create config file");
-    use std::io::Write;
-    writeln!(config_file, "Rust LCM Configuration:").unwrap();
-    writeln!(config_file, "  beta_start: {}", scheduler.config.beta_start).unwrap();
-    writeln!(config_file, "  beta_end: {}", scheduler.config.beta_end).unwrap();
-    // More config params...
-    writeln!(config_file, "\nTimesteps: {:?}", scheduler.timesteps()).unwrap();
-    writeln!(config_file, "First 10 alphas_cumprod: {:?}", scheduler.alphas_cumprod().iter().take(10).collect::<Vec<_>>()).unwrap();
-  }
-
-  // Add this code before the loop where we iterate through samples
-  // This will generate and log the guidance scale embedding
-
-  // Create and log guidance scale embedding for comparison with Python
-  println!("\n==== Generating LCM Guidance Scale Embedding ====");
-  let guidance_scale = args.guidance_scale;
-  println!("Guidance scale: {}", guidance_scale);
-
-  // The embedding dimension should match the UNet time embedding dimension
-  // For LCM Dreamshaper, this is typically 1280 (not 320)
-  // Use a hardcoded value since we don't have direct access to the config
-  let embedding_dim = 1280; // Standard for SD models based on UNet config
-  println!("Embedding dimension: {}", embedding_dim);
-
-  // Generate the embedding
-  let guidance_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &device, dtype)?;
-
-  println!("Guidance embedding shape: {:?}", guidance_embedding.shape());
-
-  // Save the embedding to a numpy file for comparison with Python
-  if save_tensors {
-    let tensor_dir = "lcm_rust_tensors";
-    std::fs::create_dir_all(tensor_dir)?;
-
-    save_tensor_as_numpy(&guidance_embedding, &format!("{}/guidance_embedding.npy", tensor_dir))?;
-    println!("Saved guidance embedding to {}/guidance_embedding.npy", tensor_dir);
-  }
-  println!("===============================================\n");
-
-  for idx in 0..num_samples {
-    let timesteps = scheduler.timesteps().to_vec();
-
-    // Initialize latents
-    let latents = if let Some(init_latent_dist) = &init_latent_dist {
-      println!("Using img2img with strength={:.4}", effective_strength);
-
-      // LCM can be more sensitive to the starting timestep
-      // Get the latents from the VAE
-      let latents = (init_latent_dist.sample()? * vae_scale)?.to_device(&device)?;
-
-      // Calculate the actual start timestep based on strength
-      // In LCM, starting too late can cause artifacts, so ensure we start early enough
-      let t_start_index = (timesteps.len() as f64 * (1.0 - effective_strength.max(0.5))).floor() as usize;
-      let t_start = timesteps[t_start_index.min(timesteps.len() - 1)];
-
-      println!("Starting from timestep {}/{} (index: {})", t_start, timesteps[0], t_start_index);
-
-      // Use a fixed seed for noise to improve reproducibility
-      device.set_seed(seed)?;
-      let noise = (Tensor::randn(0f32, 1f32, latents.shape(), &device)?).to_dtype(dtype)?;
-
-      // Clone the noise when passing it to add_noise
-      let noised_latents = scheduler.add_noise(&latents, noise.clone(), t_start)?;
-
-      // Now we can use noise for logging since we cloned it above
-      log_tensor_stats("initial_latents", &latents)?;
-      log_tensor_stats("noise", &noise)?;
-      log_tensor_stats("noised_latents", &noised_latents)?;
-
-      // Check the dtype of noised_latents
-      debug_tensor("final_img2img_latents", &noised_latents)?;
-
-      // Explicitly convert to expected dtype to avoid errors
-      noised_latents.to_dtype(dtype)?
-    } else {
-      // Regular text-to-image generation with random noise
-      let latents = Tensor::randn(0f32, 1f32, (bsize, 4, sd_config.height / 8, sd_config.width / 8), &device)?;
-
-      // Use affine for multiplication
-      let scaled_latents = latents.affine(scheduler.init_noise_sigma(), 0.0)?;
-      debug_tensor("scaled_initial_latents", &scaled_latents)?;
-
-      // Explicitly convert to expected dtype
-      scaled_latents.to_dtype(dtype)?
-    };
-    let mut latents = latents.to_dtype(dtype)?;
-
-    println!("Setting seed to {}", seed);
-    device.set_seed(seed)?;
-
-    if PRINT_DEBUG_TENSORS {
-      println!("Initial latents shape: {:?}", latents.shape());
-      println!("Initial latents first few values: {}", latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+    if !(0. ..=1.).contains(&strength) {
+        anyhow::bail!("strength should be between 0 and 1, got {strength}")
     }
 
-    println!("Starting LCM sampling");
-    let start_time = std::time::Instant::now();
-
-    for (timestep_index, &timestep) in timesteps.iter().enumerate() {
-      if timestep_index < raw_t_start {
-        continue;
-      }
-
-      println!("\n--- Timestep {}/{} (value: {}) ---", timestep_index + 1, timesteps.len(), timestep);
-
-      let latent_model_input = if use_guide_scale {
-        println!("Using classifier-free guidance with scale: {}", guidance_scale);
-        Tensor::cat(&[&latents, &latents], 0)?
-      } else {
-        latents.clone()
-      };
-
-      if PRINT_DEBUG_TENSORS {
-        println!("Latent input shape before scaling: {:?}", latent_model_input.shape());
-        println!("Latent input first few values: {}", latent_model_input.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
-      }
-
-      let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
-
-      if PRINT_DEBUG_TENSORS {
-        println!("Latent input shape after scaling: {:?}", latent_model_input.shape());
-        println!("Scaled latent input first few values: {}", latent_model_input.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
-      }
-
-      // Generate guidance scale embedding
-      let guidance_emb = if use_guide_scale {
-        // For LCM/Dreamshaper, the time embedding dimension is 1280
-        let time_embed_dim = 1280; // Standard for SD models (320 base channels * 4)
-
-        println!("Generating guidance scale embedding (scale={})", guidance_scale);
-        Some(scheduler.get_guidance_scale_embedding(guidance_scale, time_embed_dim, &device, dtype)?)
-      } else {
+    let _guard = if tracing {
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
         None
-      };
+    };
 
-      // Before UNet forward call
-      debug_tensor("latent_input_to_unet", &latent_model_input)?;
-      debug_tensor("text_embeddings_to_unet", &text_embeddings)?;
+    let dtype = if use_f16 { DType::F16 } else { DType::F32 };
+    let sd_config = match sd_version {
+        StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
+            stable_diffusion::StableDiffusionConfig::v1_5(sliced_attention_size, height, width)
+        }
+        StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint => {
+            stable_diffusion::StableDiffusionConfig::v2_1(sliced_attention_size, height, width)
+        }
+        StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint => {
+            stable_diffusion::StableDiffusionConfig::sdxl(sliced_attention_size, height, width)
+        }
+        StableDiffusionVersion::Turbo => stable_diffusion::StableDiffusionConfig::sdxl_turbo(
+            sliced_attention_size,
+            height,
+            width,
+        ),
+    };
 
-      // Pass the guidance embedding to the UNet
-      // This would require modifying the UNet forward method to accept this parameter
-      println!("Running UNet forward pass with guidance embedding...");
-      let noise_pred = unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, guidance_emb.as_ref())?;
+    let mut scheduler = stable_diffusion::lcm::LCMScheduler::new(
+        n_steps,
+        stable_diffusion::lcm::LCMSchedulerConfig::default(),
+    )?;
 
-      if PRINT_DEBUG_TENSORS {
-        println!("Noise prediction shape: {:?}", noise_pred.shape());
-        println!("Noise prediction first few values: {}", noise_pred.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
-      }
+    let device = candle_examples::device(cpu)?;
+    // If a seed is not given, generate a random seed and print it
+    let seed = seed.unwrap_or(rand::thread_rng().gen_range(0u64..u64::MAX));
+    println!("Using seed {seed}");
+    device.set_seed(seed)?;
+    let use_guide_scale = guidance_scale > 1.0;
 
-      let noise_pred = if use_guide_scale {
-        let noise_pred = noise_pred.chunk(2, 0)?;
-        let (noise_pred_uncond, noise_pred_text) = (&noise_pred[0], &noise_pred[1]);
+    let which = match sd_version {
+        StableDiffusionVersion::Xl
+        | StableDiffusionVersion::XlInpaint
+        | StableDiffusionVersion::Turbo => vec![true, false],
+        _ => vec![true],
+    };
 
-        // LCM requires a different guidance approach than standard diffusion
-        // Linear interpolation with scale factor
-        let diff = (noise_pred_text - noise_pred_uncond)?;
-        let scaled_diff = diff.affine(guidance_scale, 0.0)?;
-        (noise_pred_uncond + &scaled_diff)?
-      } else {
-        noise_pred
-      };
+    // Define the base model and LCM model repositories
+    let base_model_repo = match sd_version {
+        StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
+            // Dreamshaper is based on SD 1.5, so use it for those versions
+            "Lykon/dreamshaper-7"
+        },
+        _ => {
+            // For other versions, use the standard repo
+            sd_version.repo()
+        }
+    };
+    let lcm_model_repo = "SimianLuo/LCM_Dreamshaper_v7"; // LCM UNet
 
-      if PRINT_DEBUG_TENSORS {
-        println!("Latents before step first few values: {}", latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
-      }
+    // Create a custom ModelFile enum variant for the base model
+    let clip_weights = if clip_weights.is_none() {
+        use hf_hub::api::sync::Api;
+        let api = Api::new()?;
+        let clip_path = api.model(base_model_repo.to_string())
+            .get(if use_f16 { "text_encoder/model.fp16.safetensors" } 
+                 else { "text_encoder/model.safetensors" })?;
+        Some(clip_path.to_string_lossy().to_string())
+    } else {
+        clip_weights
+    };
 
-      println!("Applying scheduler step...");
-      latents = scheduler.step(&noise_pred, timestep, &latents)?;
+    let vae_weights = if vae_weights.is_none() {
+        use hf_hub::api::sync::Api;
+        let api = Api::new()?;
+        let vae_path = api.model(base_model_repo.to_string())
+            .get(if use_f16 { "vae/diffusion_pytorch_model.fp16.safetensors" } 
+                 else { "vae/diffusion_pytorch_model.safetensors" })?;
+        Some(vae_path.to_string_lossy().to_string())
+    } else {
+        vae_weights
+    };
 
-      if PRINT_DEBUG_TENSORS {
-        println!("Latents after step first few values: {}", latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
-      }
+    // For the UNet, always use the LCM model
+    let unet_weights = {
+        use hf_hub::api::sync::Api;
+        let api = Api::new()?;
+        let unet_path = api.model(lcm_model_repo.to_string())
+            .get(if use_f16 { "unet/diffusion_pytorch_model.fp16.safetensors" } 
+                 else { "unet/diffusion_pytorch_model.safetensors" })?;
+        Some(unet_path.to_string_lossy().to_string())
+    };
 
-      println!("Step {}/{} completed", timestep_index + 1, n_steps);
+    // Load text embeddings from the base model
+    let text_embeddings = which
+        .iter()
+        .map(|first| {
+            text_embeddings(
+                &prompt,
+                &uncond_prompt,
+                tokenizer.clone(),
+                clip_weights.clone(),
+                clip2_weights.clone(),
+                sd_version,
+                &sd_config,
+                use_f16,
+                &device,
+                dtype,
+                use_guide_scale,
+                *first,
+            )
+        })
+        .collect::<Result<Vec<_>>>()?;
 
-      // Then use it at key points
-      log_tensor_stats("latents_before_step", &latents)?;
-      log_tensor_stats("noise_pred", &noise_pred)?;
-      log_tensor_stats("latents_after_step", &latents)?;
+    let text_embeddings = Tensor::cat(&text_embeddings, D::Minus1)?;
+    let text_embeddings = text_embeddings.repeat((bsize, 1, 1))?;
+    println!("Text embeddings shape: {:?}", text_embeddings.shape());
+    println!("Text embeddings first few values: {}", 
+             text_embeddings.i(0)?.i(0)?.narrow(0, 0, 10)?);
 
-      if save_tensors {
-        let tensor_dir = "lcm_rust_tensors";
-        std::fs::create_dir_all(tensor_dir)?;
+    // Replace the random projection with a fixed one to match diffusers
+    let text_embeddings = if text_embeddings.dim(2)? == 2048 {
+        println!("Projecting text embeddings from 2048 to 768 dimensions");
+        
+        // Load the projection matrix from a file (you'll need to extract this from the diffusers model)
+        // For now, we'll still use a random one but with a fixed seed
+        device.set_seed(42)?; // Use a fixed seed for the projection
+        let projection = Tensor::randn(
+            0.0, 
+            0.02, 
+            (2048, 768), 
+            &device,
+        )?.to_dtype(dtype)?;
+        
+        let proj_embeddings = text_embeddings.matmul(&projection)?;
+        proj_embeddings
+    } else {
+        text_embeddings
+    };
+    println!("Projected text embeddings: {:?}", text_embeddings);
 
-        // Save key tensors at each step
-        let step_dir = format!("{}/step_{}", tensor_dir, timestep_index);
-        std::fs::create_dir_all(&step_dir)?;
+    println!("Building the autoencoder.");
+    let vae_weights = ModelFile::Vae.get(vae_weights, sd_version, use_f16)?;
+    let vae = sd_config.build_vae(vae_weights, &device, dtype)?;
 
-        // Save various tensors as numpy arrays for later comparison
-        save_tensor_as_numpy(&latent_model_input, &format!("{}/input.npy", step_dir))?;
-        save_tensor_as_numpy(&noise_pred, &format!("{}/noise_pred.npy", step_dir))?;
-        save_tensor_as_numpy(&latents, &format!("{}/output.npy", step_dir))?;
-      }
+    let (image, init_latent_dist) = match &img2img {
+        None => (None, None),
+        Some(image) => {
+            let image = image_preprocess(image)?
+                .to_device(&device)?
+                .to_dtype(dtype)?;
+            println!("Input image shape: {:?}", image.shape());
+            (Some(image.clone()), Some(vae.encode(&image)?))
+        }
+    };
 
-      // After UNet forward call
-      debug_tensor("noise_pred_from_unet", &noise_pred)?;
+    println!("Latent shape after VAE encoding: {:?}", 
+             init_latent_dist.as_ref().map(|dist| {
+                 // Try to get shape from a temporary sample
+                 match dist.sample() {
+                     Ok(sample) => format!("{:?}", sample.shape()),
+                     Err(_) => "Unable to determine shape".to_string()
+                 }
+             }));
+
+    println!("Building the unet.");
+    let unet_weights = ModelFile::UnetLcm.get(unet_weights, sd_version, use_f16)?;
+    let in_channels = match sd_version {
+        StableDiffusionVersion::XlInpaint
+        | StableDiffusionVersion::V2Inpaint
+        | StableDiffusionVersion::V1_5Inpaint => 9,
+        _ => 4,
+    };
+    println!("Loading UNet weights from: {:?}", unet_weights);
+    
+    // Load tensors using candle's safetensors loader
+    let tensors = candle::safetensors::load(&unet_weights, &device)?;
+    
+    // Create a new HashMap to store the processed tensors
+    let mut processed_tensors = std::collections::HashMap::new();
+    
+    // Process and print shapes of tensors
+    for (name, tensor) in tensors.iter() {
+        let processed_tensor = if name.contains("proj_in.weight") || name.contains("proj_out.weight") {
+            let shape = tensor.shape();
+            if shape.dims().len() == 2 {
+                // If it's a 2D tensor, reshape it to 4D with trailing 1s
+                println!("Expanding tensor {} from {:?}", name, shape);
+                tensor.reshape((shape.dims()[0], shape.dims()[1], 1, 1))?
+            } else {
+                tensor.clone()
+            }
+        } else {
+            tensor.clone()
+        };
+        
+        if name.contains("down_blocks.1.attentions.0.proj_in") {
+            println!("Found tensor: {} with shape: {:?} -> {:?}", 
+                    name, tensor.shape(), processed_tensor.shape());
+        }
+        
+        processed_tensors.insert(name.clone(), processed_tensor);
     }
+    
+    println!("Building UNet with in_channels: {}", in_channels);
 
-    let dt = start_time.elapsed().as_secs_f32();
-    println!("Sampling completed in {:.2}s", dt);
+    // Make sure the UNet config exactly matches the diffusers model
+    let unet_config = UNet2DConditionModelConfig {
+        // LCM specific config values
+        cross_attention_dim: 768,  // Dreamshaper v7 uses 768
+        time_cond_proj_dim: Some(256),  // LCM models use 256-dimensional guidance embeddings
+        // Rest of configuration...
+        center_input_sample: false,
+        flip_sin_to_cos: true,
+        freq_shift: 0.0,
+        blocks: vec![
+            BlockConfig {
+                out_channels: 320,
+                use_cross_attn: Some(1),
+                attention_head_dim: 8,
+            },
+            BlockConfig {
+                out_channels: 640,
+                use_cross_attn: Some(1),
+                attention_head_dim: 8,
+            },
+            BlockConfig {
+                out_channels: 1280,
+                use_cross_attn: Some(1),
+                attention_head_dim: 8,
+            },
+            BlockConfig {
+                out_channels: 1280,
+                use_cross_attn: None,
+                attention_head_dim: 8,
+            },
+        ],
+        layers_per_block: 2,
+        downsample_padding: 1,
+        mid_block_scale_factor: 1.0,
+        norm_num_groups: 32,
+        norm_eps: 1e-5,
+        sliced_attention_size: None,
+        use_linear_projection: false,
+        time_embed_dim: Some(1280),
+        // time_embed_dim: Some(256),
+    };
+    
+    
+    // Let's try a different approach - load the model directly from the safetensors file
+    // and modify the weights after loading
+    let vs_unet = unsafe { 
+        VarBuilder::from_mmaped_safetensors(&[unet_weights], dtype, &device)? 
+    };
+    
+    let unet = UNet2DConditionModel::new(
+        vs_unet,
+        in_channels,
+        4,
+        use_flash_attn,
+        unet_config,
+    )?;
 
-    // Generate final image
-    println!("Generating final image for sample {}/{}", idx + 1, num_samples);
-    save_image(&vae, &latents, vae_scale, bsize, idx, &final_image, num_samples, None)?;
-  }
-  Ok(())
-}
+    // Run with a much higher strength to ensure sufficient transformation
+    let effective_strength = if img2img.is_some() {
+        // Force at least 0.8 strength for good results 
+        strength.max(0.8)
+    } else {
+        strength
+    };
 
-// Add this debug function to check dtype at critical points
-fn debug_tensor(name: &str, tensor: &Tensor) -> Result<()> {
-  println!("DEBUG {}: shape={:?}, dtype={:?}", name, tensor.shape(), tensor.dtype());
-  Ok(())
+    // Use this effective_strength in the t_start calculation
+    // let raw_t_start = n_steps - (n_steps as f64 * effective_strength) as usize;
+    // let raw_t_start = (n_steps as f64 * (1.0 - effective_strength)).round() as usize;
+    let raw_t_start = ((n_steps as f64 * (1.0 - effective_strength)).round() as usize).min(n_steps - 1);
+
+    println!("Using img2img with strength={}, raw t_start={}", effective_strength, raw_t_start);
+    
+    // Ensure we have at least 2 steps for img2img (more steps = more transformation)
+    if raw_t_start >= n_steps - 1 {
+        println!("WARNING: strength too low, forcing more steps");
+        n_steps - 2 // Ensure we do at least 2 steps
+    } else {
+        raw_t_start
+    };
+
+    let vae_scale = match sd_version {
+        StableDiffusionVersion::V1_5
+        | StableDiffusionVersion::V1_5Inpaint
+        | StableDiffusionVersion::V2_1
+        | StableDiffusionVersion::V2Inpaint
+        | StableDiffusionVersion::XlInpaint
+        | StableDiffusionVersion::Xl => 0.18215,
+        StableDiffusionVersion::Turbo => 0.13025,
+    };
+
+    let (mask_latents, mask, mask_4) = inpainting_tensors(
+        sd_version,
+        mask_path,
+        dtype,
+        &device,
+        use_guide_scale,
+        &vae,
+        image,
+        vae_scale,
+    )?;
+
+    println!("Timesteps: {:?}", scheduler.timesteps());
+    println!("Alpha cumprod: {:?}", scheduler.alphas_cumprod());
+
+    for idx in 0..num_samples {
+        let timesteps = scheduler.timesteps().to_vec();
+        
+        // Initialize latents
+        // let latents = match &init_latent_dist {
+        //     Some(init_latent_dist) => {
+        //         let latents = (init_latent_dist.sample()? * vae_scale)?.to_device(&device)?;
+        //         if raw_t_start < timesteps.len() {
+        //             println!("Adding noise to image latents for img2img");
+        //             let noise = latents.randn_like(0f64, 1f64)?;
+                    
+        //             // Print the latents before and after adding noise
+        //             // println!("Latents before noise: {}", 
+        //                     // latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+                    
+        //             let noised_latents = scheduler.add_noise(&latents, noise, timesteps[raw_t_start])?;
+                    
+        //             let scaled_noised_latents = ((latents.clone() * (1.0 - strength))? + (noised_latents.clone() * strength)?)?;
+        //             // println!("Latents after noise: {}", 
+        //                     // noised_latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+                    
+        //             // Print the amount of noise added - clone tensors to avoid moves
+        //             let latents_clone = latents.clone();
+        //             let noised_clone = noised_latents.clone();
+        //             println!("Amount of noise added: {}", 
+        //                     ((noised_clone - latents_clone)?.div(&latents)?).i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+                    
+        //             scaled_noised_latents
+        //         } else {
+        //             println!("No noise added (t_start >= timesteps.len())");
+        //             latents
+        //         }
+        //     }
+        //     None => {
+        //         let latents = Tensor::randn(
+        //             0f32,
+        //             1f32,
+        //             (bsize, 4, sd_config.height / 8, sd_config.width / 8),
+        //             &device,
+        //         )?;
+        //         (latents * scheduler.init_noise_sigma())?
+        //     }
+        // };
+        let latents = match &init_latent_dist {
+            Some(init_latent_dist) => {
+                let latents = (init_latent_dist.sample()? * vae_scale)?.to_device(&device)?;
+        
+                // Compute `t_start`
+                let timesteps = LCMScheduler::get_timesteps_for_steps(n_steps, strength);
+                let t_start = timesteps[0]; // First denoising step
+        
+                // Add noise at `t_start`
+                let noise = latents.randn_like(0f64, 1f64)?;
+                let noisy_latents = scheduler.add_noise(&latents, noise, t_start)?;
+        
+                println!("Latents after adding noise at {}: {:?}", t_start, noisy_latents.shape());
+                noisy_latents
+            }
+            None => {
+                let latents = Tensor::randn(
+                    0f32,
+                    1f32,
+                    (bsize, 4, sd_config.height / 8, sd_config.width / 8),
+                    &device,
+                )?;
+                (latents * scheduler.init_noise_sigma())?
+            }
+        };
+
+        let mut latents = latents.to_dtype(dtype)?;
+
+        println!("Setting seed to {}", seed);
+        device.set_seed(seed)?;
+
+        println!("Initial latents shape: {:?}", latents.shape());
+        println!("Initial latents first few values: {}", 
+                 latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+
+        println!("Starting LCM sampling");
+        let start_time = std::time::Instant::now();
+
+        let embedding_dim = 256;
+        let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &device, dtype)?;
+
+        for (timestep_index, &timestep) in timesteps.iter().enumerate() {
+            if timestep_index < raw_t_start  {
+                println!("Skipping timestep {} because it's before t_start", timestep);
+                // continue;
+            }
+
+            println!("\n--- Timestep {}/{} (value: {}) ---", 
+                     timestep_index + 1, timesteps.len(), timestep);
+            
+            let latent_model_input = if use_guide_scale {
+                println!("Using classifier-free guidance with scale: {}", guidance_scale);
+                Tensor::cat(&[&latents, &latents], 0)?
+            } else {
+                latents.clone()
+            };
+            
+            println!("Latent input shape before scaling: {:?}", latent_model_input.shape());
+            println!("Latent input first few values: {}", 
+                     latent_model_input.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+
+            let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
+            
+            println!("Latent input shape after scaling: {:?}", latent_model_input.shape());
+            println!("Scaled latent input first few values: {}", 
+                     latent_model_input.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+
+            println!("Running UNet forward pass...");
+            let noise_pred = unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, Some(&guidance_scale_embedding))?;
+            
+            println!("Noise prediction shape: {:?}", noise_pred.shape());
+            println!("Noise prediction first few values: {}", 
+                     noise_pred.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+
+            let noise_pred = if use_guide_scale {
+                let noise_pred = noise_pred.chunk(2, 0)?;
+                let (noise_pred_uncond, noise_pred_text) = (&noise_pred[0], &noise_pred[1]);
+                
+                println!("Uncond noise pred first few values: {}", 
+                         noise_pred_uncond.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+                println!("Text noise pred first few values: {}", 
+                         noise_pred_text.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+                
+                let guided = (noise_pred_uncond + ((noise_pred_text - noise_pred_uncond)? * guidance_scale)?)?;
+                
+                println!("Guided noise pred first few values: {}", 
+                         guided.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+                guided
+            } else {
+                noise_pred
+            };
+
+            println!("Latents before step first few values: {}", 
+                     latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+            
+            println!("Applying scheduler step...");
+            latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, n_steps)?;
+            
+            println!("Latents after step first few values: {}", 
+                     latents.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
+            
+            println!("Step {}/{} completed", timestep_index + 1, n_steps);
+        }
+
+        let dt = start_time.elapsed().as_secs_f32();
+        println!("Sampling completed in {:.2}s", dt);
+
+        // Generate final image
+        println!("Generating final image for sample {}/{}", idx + 1, num_samples);
+        save_image(&vae, &latents, vae_scale, bsize, idx, &final_image, num_samples, None)?;
+    }
+    Ok(())
 }
 
 fn main() -> Result<()> {
-  let args = Args::parse();
-  run(args)
+    let args = Args::parse();
+    run(args)
 }
diff --git a/crates/vendor/candle/candle-flash-attn/build.rs b/crates/vendor/candle/candle-flash-attn/build.rs
index e6cefb92c4..b8db9fcd17 100644
--- a/crates/vendor/candle/candle-flash-attn/build.rs
+++ b/crates/vendor/candle/candle-flash-attn/build.rs
@@ -86,6 +86,8 @@ fn main() -> Result<()> {
         .arg("--expt-relaxed-constexpr")
         .arg("--expt-extended-lambda")
         .arg("--use_fast_math")
+        .arg("--compiler-options")
+        .arg("-fPIC")
         .arg("--verbose");
 
     if let Ok(target) = std::env::var("TARGET") {
diff --git a/crates/vendor/candle/candle-nn/Cargo.toml b/crates/vendor/candle/candle-nn/Cargo.toml
index 38711f1df7..4acb624b6d 100644
--- a/crates/vendor/candle/candle-nn/Cargo.toml
+++ b/crates/vendor/candle/candle-nn/Cargo.toml
@@ -32,3 +32,4 @@ metal = ["candle/metal", "dep:candle-metal-kernels", "dep:metal"]
 [[bench]]
 name = "bench_main"
 harness = false
+
diff --git a/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/lcm.rs b/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/lcm.rs
index 8c0b7d27ef..96fbc2cdfd 100644
--- a/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/lcm.rs
+++ b/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/lcm.rs
@@ -1,762 +1,520 @@
-use super::schedulers::{betas_for_alpha_bar, BetaSchedule, PredictionType, TimestepSpacing};
-use candle::{Result, Tensor, IndexOp, Device, DType, Shape};
+use super::schedulers::{betas_for_alpha_bar, BetaSchedule, PredictionType};
+use candle::{DType, Device, IndexOp, Result, Tensor};
 
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum LCMVarianceType {
-  FixedSmall,
-  FixedSmallLog,
-  FixedLarge,
-  FixedLargeLog,
-  Learned,
+    FixedSmall,
+    FixedSmallLog,
+    FixedLarge,
+    FixedLargeLog,
+    Learned,
 }
 
 impl Default for LCMVarianceType {
-  fn default() -> Self {
-    Self::FixedSmall
-  }
-}
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum LCMSolverType {
-  Default,
-  LogRho,
-}
-
-impl Default for LCMSolverType {
-  fn default() -> Self {
-    Self::Default
-  }
+    fn default() -> Self {
+        Self::FixedSmall
+    }
 }
 
 #[derive(Debug, Clone)]
 pub struct LCMSchedulerConfig {
-  /// The value of beta at the beginning of training.
-  pub beta_start: f64,
-  /// The value of beta at the end of training.
-  pub beta_end: f64,
-  /// How beta evolved during training.
-  pub beta_schedule: BetaSchedule,
-  /// Option to predicted sample between -1 and 1 for numerical stability.
-  pub clip_sample: bool,
-  /// Option to clip the variance used when adding noise to the denoised sample.
-  pub variance_type: LCMVarianceType,
-  /// prediction type of the scheduler function
-  pub prediction_type: PredictionType,
-  /// number of diffusion steps used to train the model.
-  pub train_timesteps: usize,
-  /// time step spacing for the diffusion process
-  pub timestep_spacing: TimestepSpacing,
-  /// Original training steps
-  pub original_inference_steps: usize,
-  /// Timestep scaling factor (default: 10.0 in diffusers)
-  pub timestep_scaling: f64,
-  /// Adjust the indexes of the inference schedule by this value.
-  pub steps_offset: usize,
-  /// The solver type to use for the diffusion process
-  pub solver_type: LCMSolverType,
+    /// The value of beta at the beginning of training.
+    pub beta_start: f64,
+    /// The value of beta at the end of training.
+    pub beta_end: f64,
+    /// How beta evolved during training.
+    pub beta_schedule: BetaSchedule,
+    /// Option to predicted sample between -1 and 1 for numerical stability.
+    pub clip_sample: bool,
+    /// Option to clip the variance used when adding noise to the denoised sample.
+    pub variance_type: LCMVarianceType,
+    /// prediction type of the scheduler function
+    pub prediction_type: PredictionType,
+    /// number of diffusion steps used to train the model.
+    pub train_timesteps: usize,
+    /// timestep scaling factor
+    pub timestep_scaling: f64,
 }
 
 impl Default for LCMSchedulerConfig {
-  fn default() -> Self {
-    Self { beta_start: 0.00085, beta_end: 0.012, beta_schedule: BetaSchedule::ScaledLinear, clip_sample: false, variance_type: LCMVarianceType::FixedSmall, prediction_type: PredictionType::Epsilon, train_timesteps: 1000, timestep_spacing: TimestepSpacing::Leading, original_inference_steps: 50, timestep_scaling: 10.0, steps_offset: 1, solver_type: LCMSolverType::Default }
-  }
+    fn default() -> Self {
+        Self {
+            beta_start: 0.00085,
+            beta_end: 0.012,
+            beta_schedule: BetaSchedule::ScaledLinear,
+            clip_sample: false,
+            variance_type: LCMVarianceType::FixedSmall,
+            prediction_type: PredictionType::Epsilon,
+            train_timesteps: 1000,
+            timestep_scaling: 10.0,
+        }
+    }
 }
 
 pub struct LCMScheduler {
-  alphas_cumprod: Vec<f64>,
-  init_noise_sigma: f64,
-  timesteps: Vec<usize>,
-  step_ratio: usize,
-  pub config: LCMSchedulerConfig,
-  step_index: Option<usize>,
-  begin_index: Option<usize>,
-  dtype: DType,
+    alphas_cumprod: Vec<f64>,
+    init_noise_sigma: f64,
+    timesteps: Vec<usize>,
+    step_ratio: usize,
+    pub config: LCMSchedulerConfig,
 }
 
 impl LCMScheduler {
-  pub fn new(inference_steps: usize, config: LCMSchedulerConfig) -> Result<Self> {
-    // Set up the correct betas
-    let betas = match config.beta_schedule {
-      BetaSchedule::ScaledLinear => {
-        // LCM uses scaled_linear specifically this way
-        let start = config.beta_start.sqrt();
-        let end = config.beta_end.sqrt();
-        let betas = super::utils::linspace(start, end, config.train_timesteps)?;
-        betas.sqr()?
-      },
-      // Other schedules...
-      _ => unreachable!("LCM needs ScaledLinear"),
-    };
+    pub fn new(inference_steps: usize, config: LCMSchedulerConfig) -> Result<Self> {
+        let betas = match config.beta_schedule {
+            BetaSchedule::ScaledLinear => super::utils::linspace(
+                config.beta_start.sqrt(),
+                config.beta_end.sqrt(),
+                config.train_timesteps,
+            )?
+            .sqr()?,
+            BetaSchedule::Linear => {
+                super::utils::linspace(config.beta_start, config.beta_end, config.train_timesteps)?
+            }
+            BetaSchedule::SquaredcosCapV2 => betas_for_alpha_bar(config.train_timesteps, 0.999)?,
+        };
 
-    // Set up alphas exactly as in diffusers
-    let betas = betas.to_vec1::<f64>()?;
-    let alphas = betas.iter().map(|beta| 1.0 - beta).collect::<Vec<_>>();
+        let betas = betas.to_vec1::<f64>()?;
+        let mut alphas_cumprod = Vec::with_capacity(betas.len());
+        let mut alpha_prod = 1f64;
+        for &beta in betas.iter() {
+            let alpha = 1.0 - beta;
+            alpha_prod *= alpha;
+            alphas_cumprod.push(alpha_prod);
+        }
 
-    let mut alphas_cumprod = Vec::with_capacity(alphas.len());
-    let mut alpha_prod = 1.0;
-    for alpha in alphas.iter() {
-      alpha_prod *= alpha;
-      alphas_cumprod.push(alpha_prod);
+        // Create proper LCM timesteps to match diffusers
+        let strength = 0.8;
+        let timesteps = LCMScheduler::get_timesteps_for_steps(inference_steps, strength);
+        let step_ratio = config.train_timesteps / inference_steps;
+
+        println!("Using LCM timesteps: {:?}", timesteps);
+
+        Ok(Self {
+            alphas_cumprod,
+            init_noise_sigma: 1.0,
+            timesteps,
+            step_ratio,
+            config,
+        })
     }
 
-    // Create timesteps based on TimestepSpacing, similar to DDIM
-    let step_ratio = config.train_timesteps / inference_steps;
-    let timesteps: Vec<usize> = match config.timestep_spacing {
-      TimestepSpacing::Leading => (0..(inference_steps)).map(|s| s * step_ratio + config.steps_offset).rev().collect(),
-      TimestepSpacing::Trailing => std::iter::successors(Some(config.train_timesteps), |n| if *n > step_ratio { Some(n - step_ratio) } else { None }).map(|n| n - 1).collect(),
-      TimestepSpacing::Linspace => super::utils::linspace(0.0, (config.train_timesteps - 1) as f64, inference_steps)?.to_vec1::<f64>()?.iter().map(|&f| f as usize).rev().collect(),
-    };
+    fn get_variance(&self, timestep: usize) -> f64 {
+        let prev_t = timestep as isize - self.step_ratio as isize;
+        let alpha_prod_t = self.alphas_cumprod[timestep];
+        let alpha_prod_t_prev = if prev_t >= 0 {
+            self.alphas_cumprod[prev_t as usize]
+        } else {
+            1.0
+        };
+        let current_beta_t = 1. - alpha_prod_t / alpha_prod_t_prev;
 
-    println!("Using LCM timesteps with {:?} spacing: {:?}", config.timestep_spacing, timesteps);
+        // For t > 0, compute predicted variance βt (see formula (6) and (7) from [the pdf](https://arxiv.org/pdf/2006.11239.pdf))
+        // and sample from it to get previous sample
+        // x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
+        let variance = (1. - alpha_prod_t_prev) / (1. - alpha_prod_t) * current_beta_t;
 
-    let init_noise_sigma = 1.0;
-
-    Ok(Self { alphas_cumprod, init_noise_sigma, timesteps, step_ratio, config, step_index: None, begin_index: None, dtype: DType::F32 })
-  }
-
-  pub fn with_dtype(mut self, dtype: DType) -> Self {
-    self.dtype = dtype;
-    self
-  }
-
-  fn get_variance(&self, timestep: usize) -> f64 {
-    let prev_t = timestep as isize - self.step_ratio as isize;
-    let alpha_prod_t = self.alphas_cumprod[timestep];
-    let alpha_prod_t_prev = if prev_t >= 0 { self.alphas_cumprod[prev_t as usize] } else { 1.0 };
-    let current_beta_t = 1. - alpha_prod_t / alpha_prod_t_prev;
-
-    // For t > 0, compute predicted variance βt (see formula (6) and (7) from [the pdf](https://arxiv.org/pdf/2006.11239.pdf))
-    // and sample from it to get previous sample
-    // x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
-    let variance = (1. - alpha_prod_t_prev) / (1. - alpha_prod_t) * current_beta_t;
-
-    // retrieve variance
-    match self.config.variance_type {
-      LCMVarianceType::FixedSmall => variance.max(1e-20),
-      // for rl-diffuser https://arxiv.org/abs/2205.09991
-      LCMVarianceType::FixedSmallLog => {
-        let variance = variance.max(1e-20).ln();
-        (variance * 0.5).exp()
-      },
-      LCMVarianceType::FixedLarge => current_beta_t,
-      LCMVarianceType::FixedLargeLog => current_beta_t.ln(),
-      LCMVarianceType::Learned => variance,
-    }
-  }
-
-  pub fn timesteps(&self) -> &[usize] {
-    self.timesteps.as_slice()
-  }
-
-  ///  Ensures interchangeability with schedulers that need to scale the denoising model input
-  /// depending on the current timestep.
-  pub fn scale_model_input(&self, sample: Tensor, timestep: usize) -> Tensor {
-    // LCM doesn't apply any special scaling to model inputs
-    println!("LCM: No scaling applied to model input for timestep {}", timestep);
-    sample
-  }
-
-  fn get_scalings_for_boundary_condition_discrete(&self, timestep: usize) -> (f64, f64) {
-    // LCM uses a fixed sigma_data value of 0.5
-    let sigma_data = 0.5f64;
-
-    // Timestep scaling factor from diffusers configuration
-    let timestep_scaling = self.config.timestep_scaling;
-
-    // In diffusers, they don't use the timestep value directly
-    let alpha_cumprod = self.alphas_cumprod[timestep];
-    let sqrt_alpha_prod = alpha_cumprod.sqrt();
-    let sqrt_one_minus_alpha_prod = (1.0 - alpha_cumprod).sqrt();
-
-    // This is the key calculation - convert alpha to sigma space with scaling
-    let scaled_timestep = sqrt_one_minus_alpha_prod / sqrt_alpha_prod * timestep_scaling;
-
-    println!("Boundary condition for timestep={}: scaled_t={} (with scaling={})", timestep, scaled_timestep, timestep_scaling);
-
-    // Calculate boundary condition coefficients with scaled timestep
-    let c_skip = sigma_data.powf(2.0) / (scaled_timestep.powf(2.0) + sigma_data.powf(2.0));
-    let c_out = scaled_timestep / (scaled_timestep.powf(2.0) + sigma_data.powf(2.0)).sqrt();
-
-    (c_skip, c_out)
-  }
-
-  pub fn step(&mut self, model_output: &Tensor, timestep: usize, sample: &Tensor) -> Result<Tensor> {
-    // Initialize step_index if it's not already set
-    if self.step_index.is_none() {
-      self.init_step_index(timestep)?;
-      println!("Initialized step_index to {} for timestep {}", self.step_index.unwrap(), timestep);
+        // retrieve variance
+        match self.config.variance_type {
+            LCMVarianceType::FixedSmall => variance.max(1e-20),
+            // for rl-diffuser https://arxiv.org/abs/2205.09991
+            LCMVarianceType::FixedSmallLog => {
+                let variance = variance.max(1e-20).ln();
+                (variance * 0.5).exp()
+            }
+            LCMVarianceType::FixedLarge => current_beta_t,
+            LCMVarianceType::FixedLargeLog => current_beta_t.ln(),
+            LCMVarianceType::Learned => variance,
+        }
     }
 
-    // Get current timestep index
-    let timestep_index = self.step_index.unwrap();
-
-    // Calculate prev_timestep_index
-    let prev_timestep_index = timestep_index + 1;
-    let prev_timestep = if prev_timestep_index >= self.timesteps.len() {
-      0 // Final step
-    } else {
-      self.timesteps[prev_timestep_index]
-    };
-
-    // Update step_index for next time
-    self.step_index = Some(prev_timestep_index);
-
-    // Choose solver based on configuration
-    match self.config.solver_type {
-      LCMSolverType::Default => self.step_default(model_output, timestep, prev_timestep, sample),
-      LCMSolverType::LogRho => self.step_logrho(model_output, timestep, prev_timestep, sample),
+    pub fn timesteps(&self) -> &[usize] {
+        self.timesteps.as_slice()
     }
-  }
 
-  // Default solver (existing logic moved to a separate method)
-  fn step_default(&self, model_output: &Tensor, timestep: usize, prev_timestep: usize, sample: &Tensor) -> Result<Tensor> {
-    // Get dtype from sample and use it for consistency
-    let target_dtype = sample.dtype();
-    let device = sample.device();
+    ///  Ensures interchangeability with schedulers that need to scale the denoising model input
+    /// depending on the current timestep.
+    pub fn scale_model_input(&self, sample: Tensor, _timestep: usize) -> Tensor {
+        sample
+    }
+    fn get_scalings_for_boundary_condition_discrete(&self, timestep: usize) -> (f64, f64) {
+        let sigma_data: f64 = 0.5;
+        let scaled_timestep = (timestep as f64) * self.config.timestep_scaling;
+        let sigma_squared = sigma_data.powi(2);
+        let scaled_timestep_squared = scaled_timestep.powi(2);
+        let denominator = scaled_timestep_squared + sigma_squared;
 
-    println!("Step (Default): current_t={}, prev_t={}", timestep, prev_timestep);
+        println!(
+            "Boundary condition for timestep {}, scaled timestep {}",
+            timestep, scaled_timestep
+        );
+        println!(
+            "  Intermediate values: sigma_data^2: {:.20}, scaled_timestep^2: {:.20}, denominator: {:.20}",
+            sigma_squared, scaled_timestep_squared, denominator
+        );
 
-    // Convert model output to match sample dtype
-    let model_output = model_output.to_dtype(target_dtype)?;
+        let c_skip = sigma_squared / denominator;
+        let c_out = scaled_timestep / (scaled_timestep_squared + sigma_squared).sqrt();
 
-    // Get the exact alpha values for this timestep
-    let alpha_prod_t = if timestep < self.alphas_cumprod.len() {
-      self.alphas_cumprod[timestep]
-    } else {
-      println!("Warning: Timestep {} out of range, using last alpha", timestep);
-      *self.alphas_cumprod.last().unwrap()
-    };
+        println!(
+            "  Final computed values: c_skip: {:.20}, c_out: {:.20}",
+            c_skip, c_out
+        );
 
-    // Get the previous alpha value
-    let alpha_prod_t_prev = if prev_timestep == 0 {
-      1.0 // Final step goes to alpha=1
-    } else if prev_timestep < self.alphas_cumprod.len() {
-      self.alphas_cumprod[prev_timestep]
-    } else {
-      println!("Warning: Prev timestep {} out of range", prev_timestep);
-      *self.alphas_cumprod.last().unwrap()
-    };
+        (c_skip, c_out)
+    }
 
-    // Calculate coefficients for boundary conditions
-    let (c_skip, c_out) = self.get_scalings_for_boundary_condition_discrete(timestep);
+    pub fn get_guidance_scale_embedding(
+        &self,
+        guidance_scale: f64,
+        embedding_dim: usize,
+        device: &Device,
+        dtype: DType,
+    ) -> Result<Tensor> {
+        // Scale by exactly 1000.0 to match diffusers
+        let w = guidance_scale * 1000.0;
+        println!(
+            "Creating guidance embedding with dim={}, scaled value={}",
+            embedding_dim, w
+        );
+
+        // Calculate the half dimension
+        let half_dim = embedding_dim / 2;
+        println!("Half dimension: {}", half_dim);
+
+        let log_10000 = (10000.0f32).ln();
+        let freq_factor = log_10000 / ((half_dim - 1) as f32); // pre-compute division
+
+        let mut freqs = Vec::with_capacity(half_dim);
+        for i in 0..half_dim {
+            // This matches torch.exp(torch.arange(half_dim) * -emb)
+            let freq = (-freq_factor * i as f32).exp();
+            freqs.push(freq);
+        }
+
+        println!(
+            "First few frequencies: {:?}",
+            &freqs.iter().take(5).collect::<Vec<_>>()
+        );
+
+        // To match PyTorch's w[:, None] * emb[None, :], we need to broadcast manually
+        // First, create a vector with w multiplied by each frequency
+        let mut emb_data = Vec::with_capacity(half_dim);
+        let w_val = w as f32;
+        for &freq in freqs.iter() {
+            emb_data.push(w_val * freq);
+        }
+
+        println!(
+            "First few scaled frequencies: {:?}",
+            &emb_data.iter().take(5).collect::<Vec<_>>()
+        );
+
+        // Create the embedding tensor
+        let emb = Tensor::new(emb_data.as_slice(), device)?.to_dtype(dtype)?;
+        println!("Embedding tensor shape: {:?}", emb.shape());
+
+        // Reshape for correct sin/cos application
+        let emb = emb.reshape((1, half_dim))?;
+        println!("Reshaped embedding tensor: {:?}", emb.shape());
+
+        // Compute sin and cos components
+        let sin_emb = emb.sin()?;
+        println!("Sin embedding shape: {:?}", sin_emb.shape());
+        let cos_emb = emb.cos()?;
+        println!("Cos embedding shape: {:?}", cos_emb.shape());
+
+        // Concatenate sin and cos (along dimension 1)
+        let result = Tensor::cat(&[sin_emb, cos_emb], 1)?;
+        println!("After concatenation shape: {:?}", result.shape());
+
+        // If odd dimension, pad with zero
+        let result = if embedding_dim % 2 == 1 {
+            println!("Adding zero padding for odd dimension");
+            let zero = Tensor::zeros((1, 1), dtype, device)?;
+            Tensor::cat(&[result, zero], 1)?
+        } else {
+            result
+        };
+
+        println!("Final guidance embedding shape: {:?}", result.shape());
+
+        Ok(result)
+    }
+
+    pub fn get_timesteps_for_steps(inference_steps: usize, strength: f64) -> Vec<usize> {
+        // These are the timesteps for LCM
+        let all_timesteps = match inference_steps {
+            2 => vec![999, 499],
+            3 => vec![999, 679, 339],
+            4 => vec![999, 759, 499, 259],
+            5 => vec![999, 799, 599, 399, 199],
+            6 => vec![999, 839, 679, 499, 339, 179],
+            7 => vec![999, 859, 719, 579, 439, 299, 159],
+            8 => vec![999, 879, 759, 639, 499, 379, 259, 139],
+            9 => vec![999, 899, 779, 679, 559, 459, 339, 239, 119],
+            10 => vec![999, 899, 799, 699, 599, 499, 399, 299, 199, 99],
+            11 => vec![999, 919, 819, 739, 639, 559, 459, 379, 279, 199, 99],
+            12 => vec![999, 919, 839, 759, 679, 599, 499, 419, 339, 259, 179, 99],
+            13 => vec![
+                999, 939, 859, 779, 699, 619, 539, 479, 399, 319, 239, 159, 79,
+            ],
+            14 => vec![
+                999, 939, 859, 799, 719, 659, 579, 499, 439, 359, 299, 219, 159, 79,
+            ],
+            15 => vec![
+                999, 939, 879, 799, 739, 679, 599, 539, 479, 399, 339, 279, 199, 139, 79,
+            ],
+            16 => vec![
+                999, 939, 879, 819, 759, 699, 639, 579, 499, 439, 379, 319, 259, 199, 139, 79,
+            ],
+            _ => {
+                // if there's an error return the timesteps for 16 steps
+                let linspace = match super::utils::linspace(0.0, (999) as f64, inference_steps) {
+                    Ok(linspace) => match linspace.to_vec1::<f64>() {
+                        Ok(linspace) => linspace,
+                        Err(_) => {
+                            return vec![
+                                999, 939, 879, 819, 759, 699, 639, 579, 499, 439, 379, 319, 259,
+                                199, 139, 79,
+                            ];
+                        }
+                    },
+                    Err(_) => {
+                        return vec![
+                            999, 939, 879, 819, 759, 699, 639, 579, 499, 439, 379, 319, 259, 199,
+                            139, 79,
+                        ];
+                    }
+                };
+
+                let timesteps = linspace.iter().map(|&f| f as usize).rev().collect();
+                println!("timesteps: {:?}", timesteps);
+                timesteps
+            }
+        };
+
+        println!("all_timesteps: {:?}", all_timesteps);
+
+        // Compute `t_start` based on `strength`
+        let num_train_timesteps = 1000;
+        let init_timestep = (strength * num_train_timesteps as f64).round() as usize;
+        let t_start = all_timesteps
+            .iter()
+            .position(|&t| t <= init_timestep)
+            .unwrap_or(0);
+
+        println!(
+            "get_timesteps: t_start: {} init_timestep: {} strength: {}",
+            t_start, init_timestep, strength
+        );
+
+        // Return only the steps from `t_start` onward
+        all_timesteps[t_start..].to_vec()
+    }
+
+    pub fn step(
+        &self,
+        model_output: &Tensor,
+        timestep: usize,
+        sample: &Tensor,
+        step_index: usize,
+        num_inference_steps: usize,
+    ) -> Result<Tensor> {
+        let step_index = self.timesteps.iter().position(|&t| t == timestep).unwrap();
+        let prev_timestep = if step_index + 1 < self.timesteps.len() {
+            self.timesteps[step_index + 1]
+        } else {
+            0
+        };
+
+        println!(
+            "LCM step: timestep={}, prev_timestep={}, step_index={}",
+            timestep, prev_timestep, step_index
+        );
+
+        let alpha_prod_t = self.alphas_cumprod[timestep];
+        let alpha_prod_t_prev = if prev_timestep == 0 {
+            1.0
+        } else {
+            self.alphas_cumprod[prev_timestep]
+        };
 
-    // Compute predicted original sample from noise prediction
-    let pred_original_sample = match self.config.prediction_type {
-      PredictionType::Epsilon => {
         let beta_prod_t = 1.0 - alpha_prod_t;
-        let beta_prod_t_sqrt = beta_prod_t.sqrt();
+        let beta_prod_t_prev = 1.0 - alpha_prod_t_prev;
 
-        let scaled_noise = model_output.affine(beta_prod_t_sqrt, 0.0)?;
-        let diff = (sample - &scaled_noise)?;
-        diff.affine(1.0 / alpha_prod_t.sqrt(), 0.0)?
-      },
-      PredictionType::Sample => model_output.clone(),
-      PredictionType::VPrediction => {
-        let sqrt_alpha_prod = alpha_prod_t.sqrt();
-        let sqrt_one_minus_alpha_prod = (1.0 - alpha_prod_t).sqrt();
+        let (c_skip, c_out) = self.get_scalings_for_boundary_condition_discrete(timestep);
 
-        let term1 = sample.affine(sqrt_alpha_prod, 0.0)?;
-        let term2 = model_output.affine(sqrt_one_minus_alpha_prod, 0.0)?;
-        (term1 - term2)?
-      },
-      _ => unreachable!(),
-    };
+        let pred_original_sample = match self.config.prediction_type {
+            PredictionType::Epsilon => {
+                ((sample - (beta_prod_t.sqrt() * model_output)?)? / alpha_prod_t.sqrt())?
+            }
+            PredictionType::Sample => model_output.clone(),
+            PredictionType::VPrediction => {
+                ((sample * alpha_prod_t.sqrt())? - (model_output * beta_prod_t.sqrt())?)?
+            }
+        };
 
-    // Apply possible clipping
-    let pred_orig = if self.config.clip_sample { pred_original_sample.clamp(-1f32, 1f32)? } else { pred_original_sample };
+        let denoised = ((pred_original_sample * c_out)? + (sample * c_skip)?)?;
 
-    // Apply boundary condition
-    let term1 = pred_orig.affine(c_out, 0.0)?;
-    let term2 = sample.affine(c_skip, 0.0)?;
-    let denoised = (&term1 + &term2)?;
+        let pred_prev_sample = if step_index != self.timesteps.len() - 1 {
+            println!("Adding noise at intermediate step {}", timestep);
+            let noise = Tensor::randn_like(&denoised, 0.0, 1.0)?;
+            ((alpha_prod_t_prev.sqrt() * denoised)? + (beta_prod_t_prev.sqrt() * noise)?)?
+        } else {
+            println!("No noise for final timestep {}", timestep);
+            denoised
+        };
 
-    // Set up the previous sample
-    let pred_prev_sample = if prev_timestep == 0 {
-      denoised.clone()
-    } else {
-      let prev_sqrt_alpha = alpha_prod_t_prev.sqrt();
-      denoised.affine(prev_sqrt_alpha, 0.0)?
-    };
+        println!(
+            "Denoised sample shape: {:?}, first few values: {}",
+            pred_prev_sample.shape(),
+            pred_prev_sample
+                .i(0)?
+                .i(0)?
+                .narrow(0, 0, 5)?
+                .narrow(1, 0, 5)?
+        );
 
-    Ok(pred_prev_sample)
-  }
-
-  // LogRho solver implementation
-  fn step_logrho(&self, model_output: &Tensor, timestep: usize, prev_timestep: usize, sample: &Tensor) -> Result<Tensor> {
-    // Ensure consistent dtype by getting it from the sample
-    let target_dtype = sample.dtype();
-    let device = sample.device();
-
-    println!("Step (LogRho): current_t={}, prev_t={}", timestep, prev_timestep);
-    println!("  Using target dtype: {:?}", target_dtype);
-
-    // Convert model output to match sample dtype
-    let model_output = model_output.to_dtype(target_dtype)?;
-    println!("  model_output dtype after conversion: {:?}", model_output.dtype());
-
-    // Get alpha values
-    let alpha_prod_t = if timestep < self.alphas_cumprod.len() { self.alphas_cumprod[timestep] } else { *self.alphas_cumprod.last().unwrap() };
-
-    let alpha_prod_t_prev = if prev_timestep == 0 {
-      1.0
-    } else if prev_timestep < self.alphas_cumprod.len() {
-      self.alphas_cumprod[prev_timestep]
-    } else {
-      *self.alphas_cumprod.last().unwrap()
-    };
-
-    // Special case: if we're at the last timestep, return the predicted sample directly
-    if prev_timestep == 0 {
-      // Handle final step differently - exact same implementation as in default solver
-      return self.step_default(&model_output, timestep, prev_timestep, sample);
+        Ok(pred_prev_sample)
     }
 
-    // Calculate sigma values for LogRho solver with numerical stability
-    let sigma_t = ((1.0 - alpha_prod_t).max(1e-12)).sqrt();
-    let sigma_t_prev = ((1.0 - alpha_prod_t_prev).max(1e-12)).sqrt();
+    // pub fn step(&self, model_output: &Tensor, timestep: usize, sample: &Tensor) -> Result<Tensor> {
+    //     let step_index = self.timesteps.iter().position(|&t| t == timestep).unwrap();
 
-    // Calculate key logrho parameters with numerical stability
-    let logrho_t = sigma_t.ln() - alpha_prod_t.sqrt().max(1e-12).ln();
-    let logrho_t_prev = sigma_t_prev.ln() - alpha_prod_t_prev.sqrt().max(1e-12).ln();
+    //     // Get previous timestep value correctly
+    //     let prev_timestep = if step_index == self.timesteps.len() - 1 {
+    //         0 // If this is the last step, use timestep 0
+    //     } else {
+    //         self.timesteps[step_index + 1]
+    //     };
 
-    // Calculate h parameter (step size in log rho space)
-    let h = logrho_t_prev - logrho_t;
-    let exp_h = h.exp().clamp(0.01, 5.0); // Constrain to a reasonable range
-    println!("  LogRho params: h={:.6}, exp(h)={:.6}", h, exp_h);
+    //     println!(
+    //         "LCM step: timestep={}, prev_timestep={}",
+    //         timestep, prev_timestep
+    //     );
 
-    // For epsilon prediction we need to transform the model output
-    let score = match self.config.prediction_type {
-      PredictionType::Epsilon => {
-        // For ε prediction type, score is -σ/√α_t * ε
-        let scale = -sigma_t / alpha_prod_t.sqrt().max(1e-12);
-        model_output.affine(scale, 0.0)?.to_dtype(target_dtype)?
-      },
-      PredictionType::VPrediction => {
-        // model_output is already in the right form
-        model_output.clone()
-      },
-      PredictionType::Sample => {
-        // Score = (x_t - √α_t * x_0) / σ_t
-        let alpha_sqrt = alpha_prod_t.sqrt();
-        let scaled_pred = model_output.affine(alpha_sqrt, 0.0)?.to_dtype(target_dtype)?;
-        let diff = (sample.clone() - scaled_pred)?.to_dtype(target_dtype)?;
-        diff.affine(1.0 / sigma_t, 0.0)?.to_dtype(target_dtype)?
-      },
-      _ => unreachable!(),
-    };
+    //     // Get alpha values for current and previous timestep
+    //     let alpha_prod_t = self.alphas_cumprod[timestep];
+    //     let alpha_prod_t_prev = if prev_timestep == 0 {
+    //         1.0 // Use 1.0 for the final step
+    //     } else {
+    //         self.alphas_cumprod[prev_timestep]
+    //     };
 
-    println!("  score dtype: {:?}", score.dtype());
+    //     println!(
+    //         "LCM scheduler step: timestep={}, step_index={}, sigma={}",
+    //         timestep,
+    //         step_index,
+    //         alpha_prod_t.sqrt()
+    //     );
 
-    // Calculate d term: d = x_t + σ_t^2/√α_t * score
-    let scale_d = sigma_t.powi(2) / alpha_prod_t.sqrt().max(1e-12);
-    let score_scaled = score.affine(scale_d, 0.0)?.to_dtype(target_dtype)?;
-    let d = (sample + score_scaled)?.to_dtype(target_dtype)?;
-    println!("  d dtype: {:?}", d.dtype());
+    //     let beta_prod_t = 1.0 - alpha_prod_t;
+    //     let beta_prod_t_prev = 1.0 - alpha_prod_t_prev;
 
-    // Apply the main update formula directly:
-    // x_prev = exp_h * sample + sigma_prev * (1 - exp_h) * d
+    //     // Get scalings for boundary conditions
+    //     let (c_skip, c_out) = self.get_scalings_for_boundary_condition_discrete(timestep);
 
-    // First part: exp_h * sample
-    let term1 = sample.affine(exp_h, 0.0)?.to_dtype(target_dtype)?;
-    println!("  term1 dtype: {:?}", term1.dtype());
+    //     // Compute predicted original sample x_0
+    //     let mut pred_original_sample = match self.config.prediction_type {
+    //         PredictionType::Epsilon => {
+    //             let predicted_original_sample =
+    //                 ((sample - (beta_prod_t.sqrt() * model_output)?)? / alpha_prod_t.sqrt())?;
+    //             predicted_original_sample
+    //         }
+    //         PredictionType::Sample => model_output.clone(),
+    //         PredictionType::VPrediction => {
+    //             let predicted_original_sample =
+    //                 ((sample * alpha_prod_t.sqrt())? - (model_output * beta_prod_t.sqrt())?)?;
+    //             predicted_original_sample
+    //         }
+    //     };
 
-    // Second part: sigma_prev * (1 - exp_h) * d
-    let term2_factor = sigma_t_prev * (1.0 - exp_h);
-    println!("  term2_factor: {:.6}", term2_factor);
+    //     if self.config.clip_sample {
+    //         pred_original_sample = pred_original_sample.clamp(-1f32, 1f32)?;
+    //     }
 
-    let term2 = d.affine(term2_factor, 0.0)?.to_dtype(target_dtype)?;
-    println!("  term2 dtype: {:?}", term2.dtype());
+    //     println!("boundary: Pred original sample first few values: {}",
+    //              pred_original_sample.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
 
-    // Add the terms together
-    let pred_prev_sample = (&term1 + &term2)?.to_dtype(target_dtype)?;
-    println!("  pred_prev_sample dtype: {:?}", pred_prev_sample.dtype());
+    //     // Denoise model output using boundary conditions
+    //     let denoised = ((pred_original_sample * c_out)? + (sample * c_skip)?)?;
 
-    Ok(pred_prev_sample)
-  }
+    //     println!("boundary: Denoised first few values: {}",
+    //              denoised.i(0)?.i(0)?.narrow(0, 0, 5)?.narrow(1, 0, 5)?);
 
-  pub fn add_noise(&self, original_samples: &Tensor, noise: Tensor, timestep: usize) -> Result<Tensor> {
-    // Get the correct alpha value
-    let alpha_cumprod = self.alphas_cumprod[timestep];
+    //     // LCM doesn't add noise during inference
+    //     let pred_prev_sample = if prev_timestep >= self.alphas_cumprod.len() {
+    //         denoised
+    //     } else {
+    //         let scale = (Tensor::zeros_like(&denoised)? + alpha_prod_t_prev.sqrt())?;
+    //         (&denoised).mul(&scale)?
+    //     };
 
-    println!("Adding noise for timestep={}, alpha={:.6}, sqrt(alpha)={:.6}", timestep, alpha_cumprod, alpha_cumprod.sqrt());
-    println!("  original_samples: shape={:?}, dtype={:?}", original_samples.shape(), original_samples.dtype());
-    println!("  noise: shape={:?}, dtype={:?}", noise.shape(), noise.dtype());
+    //     // let pred_prev_sample = pred_prev_sample.clamp(-1f32, 1f32)?;
 
-    // Keep operations in original_samples' dtype for now
-    let target_dtype = original_samples.dtype();
-    let noise = noise.to_dtype(target_dtype)?;
+    //     println!(
+    //         "LCM scheduler pred_original shape: {:?}",
+    //         pred_prev_sample.shape()
+    //     );
+    //     println!(
+    //         "LCM scheduler pred_original first few values: {}",
+    //         pred_prev_sample
+    //             .i(0)?
+    //             .i(0)?
+    //             .narrow(0, 0, 5)?
+    //             .narrow(1, 0, 5)?
+    //     );
 
-    // Calculate coefficients
-    let sqrt_alpha = alpha_cumprod.sqrt();
-    let sqrt_one_minus_alpha = (1.0 - alpha_cumprod).sqrt();
+    //     Ok(pred_prev_sample)
+    // }
 
-    // Apply noise formula
-    let scaled_samples = original_samples.affine(sqrt_alpha, 0.0)?;
-    let scaled_noise = noise.affine(sqrt_one_minus_alpha, 0.0)?;
+    pub fn add_noise(&self, original: &Tensor, noise: Tensor, timestep: usize) -> Result<Tensor> {
+        let timestep = if timestep >= self.alphas_cumprod.len() {
+            self.alphas_cumprod.len() - 1 // Ensure valid range
+        } else {
+            timestep
+        };
+        let sqrt_alpha_prod = self.alphas_cumprod[timestep].sqrt();
+        let sqrt_one_minus_alpha_prod = (1.0 - self.alphas_cumprod[timestep] + 1e-6).sqrt(); // Prevent zero issue
 
-    println!("  scaled_samples: shape={:?}, dtype={:?}", scaled_samples.shape(), scaled_samples.dtype());
-    println!("  scaled_noise: shape={:?}, dtype={:?}", scaled_noise.shape(), scaled_noise.dtype());
-
-    // Add scaled noise - should maintain dtype
-    let result = scaled_samples.add(&scaled_noise)?;
-
-    println!("  result: shape={:?}, dtype={:?}", result.shape(), result.dtype());
-
-    Ok(result)
-  }
-
-  pub fn init_noise_sigma(&self) -> f64 {
-    self.init_noise_sigma
-  }
-
-  pub fn alphas_cumprod(&self) -> &[f64] {
-    &self.alphas_cumprod
-  }
-
-  pub fn log_config(&self) {
-    println!("LCM Scheduler Configuration:");
-    println!("  beta_start: {}", self.config.beta_start);
-    println!("  beta_end: {}", self.config.beta_end);
-    println!("  beta_schedule: {:?}", self.config.beta_schedule);
-    println!("  train_timesteps: {}", self.config.train_timesteps);
-    println!("  prediction_type: {:?}", self.config.prediction_type);
-    println!("  clip_sample: {}", self.config.clip_sample);
-    println!("  variance_type: {:?}", self.config.variance_type);
-    println!("  timesteps: {:?}", self.timesteps);
-    // Print the first few alphas_cumprod values
-    let num_to_show = std::cmp::min(10, self.alphas_cumprod.len());
-    println!("  First {} alphas_cumprod: {:?}", num_to_show, self.alphas_cumprod.iter().take(num_to_show).collect::<Vec<_>>());
-  }
-
-  /// Creates embeddings from guidance scale values, similar to timestep embeddings
-  /// Based on diffusers implementation
-  pub fn get_guidance_scale_embedding(&self, guidance_scale: f64, embedding_dim: usize, device: &Device, dtype: DType) -> Result<Tensor> {
-    // Scale by exactly 1000.0 to match diffusers
-    let w = guidance_scale * 1000.0;
-    println!("Creating guidance embedding with dim={}, scaled value={}", embedding_dim, w);
-
-    // Create embedding vector directly to avoid tensor operations
-    let mut embedding = Vec::with_capacity(embedding_dim);
-    let half_dim = embedding_dim / 2;
-
-    // Compute the position frequencies
-    for i in 0..half_dim {
-      let freq = (i as f32) * (-((10000.0f32).ln()) / ((half_dim - 1) as f32));
-      let freq = freq.exp();
-
-      // Apply the guidance scale (w value)
-      let value = w as f32 * freq;
-
-      // Add sin and cos values
-      embedding.push(value.sin());
-      embedding.push(value.cos());
+        (original * sqrt_alpha_prod)? + (noise * sqrt_one_minus_alpha_prod)?
     }
 
-    // If odd dimension, pad with zero
-    if embedding.len() < embedding_dim {
-      embedding.push(0.0);
+    // pub fn add_noise(
+    //     &self,
+    //     original_samples: &Tensor,
+    //     noise: Tensor,
+    //     timestep: usize,
+    // ) -> Result<Tensor> {
+    //     // Get the correct alpha value
+    //     let alpha_cumprod = self.alphas_cumprod[timestep];
+
+    //     // Print the alpha value to debug
+    //     println!(
+    //         "Adding noise with timestep={}, alpha={}, sqrt(alpha)={}",
+    //         timestep,
+    //         alpha_cumprod,
+    //         alpha_cumprod.sqrt()
+    //     );
+
+    //     // Correctly scale the original samples and noise using scalar multiplication
+    //     //     let scaled_samples = (original_samples * alpha_cumprod.sqrt())?;
+    //     //     let scaled_noise = (noise * (1.0 - alpha_cumprod).sqrt())?;
+
+    //     //    scaled_samples + scaled_noise
+    //     let clamped_strength = 0.6;
+    //     let scaled_samples = (original_samples * alpha_cumprod.sqrt())?;
+    //     let scaled_noise = ((noise * (1.0 - alpha_cumprod).sqrt())? * clamped_strength)?;
+
+    //     scaled_samples + scaled_noise
+    // }
+
+    pub fn init_noise_sigma(&self) -> f64 {
+        self.init_noise_sigma
     }
 
-    // Ensure we have the correct length
-    assert_eq!(embedding.len(), embedding_dim, "Embedding length mismatch");
-
-    // Create tensor from the embedding vector
-    let tensor = Tensor::new(embedding.as_slice(), device)?.to_dtype(dtype)?;
-
-    // Reshape to match expected [1, embedding_dim] shape
-    let result = tensor.reshape((1, embedding_dim))?;
-
-    println!("Final guidance embedding shape: {:?}", result.shape());
-
-    Ok(result)
-  }
-
-  /// Calculate timesteps based on LCM's specific formula
-  /// This is now only used for backward compatibility or custom timesteps
-  pub fn get_timesteps_for_steps(inference_steps: usize) -> Vec<usize> {
-    match inference_steps {
-      // 4 => vec![999, 499, 259, 0],
-      4 => vec![499, 259],
-      8 => vec![999, 879, 759, 639, 499, 379, 259, 139, 0],
-      _ => {
-        let intervals = (999 / (inference_steps - 1)) as usize;
-        let mut timesteps = Vec::with_capacity(inference_steps);
-
-        // First step always starts at 999
-        timesteps.push(999);
-
-        // Generate intermediate steps
-        for i in 1..inference_steps {
-          // The pattern looks approximately linear with a step size determined by the number of steps
-          let ts = 999 - i * intervals;
-          timesteps.push(ts);
-        }
-
-        // Last step is always 0
-        if let Some(last) = timesteps.last_mut() {
-          if *last != 0 {
-            *last = 0;
-          }
-        }
-
-        timesteps
-      },
+    pub fn alphas_cumprod(&self) -> &[f64] {
+        &self.alphas_cumprod
     }
-  }
-
-  /// Get the appropriate starting timestep based on strength
-  pub fn get_timestep_for_strength(&self, strength: f64) -> (usize, f64) {
-    // Convert strength to a timestep index
-    let num_steps = self.timesteps.len();
-    let strength_adjusted = (strength * num_steps as f64).floor() as usize;
-    let start_idx = num_steps.saturating_sub(strength_adjusted + 1);
-
-    // Get the timestep at this index
-    let timestep = if start_idx < self.timesteps.len() { self.timesteps[start_idx] } else { *self.timesteps.first().unwrap_or(&999) };
-
-    // Get the corresponding alpha
-    let alpha = if timestep < self.alphas_cumprod.len() { self.alphas_cumprod[timestep] } else { *self.alphas_cumprod.last().unwrap_or(&0.0) };
-
-    println!("For strength {:.2}, using timestep {} (index {}), alpha={:.6}", strength, timestep, start_idx, alpha);
-
-    (timestep, alpha)
-  }
-
-  /// Add noise for img2img based on strength
-  pub fn add_noise_for_img2img(&mut self, original_samples: &Tensor, noise: &Tensor, strength: f64) -> Result<(Tensor, usize)> {
-    // Get starting timestep and alpha
-    let (timestep, alpha) = self.get_timestep_for_strength(strength);
-
-    // Set begin_index based on this timestep
-    let start_idx = self.index_for_timestep(timestep)?;
-    self.set_begin_index(start_idx);
-
-    println!("Adding img2img noise with strength={:.4}, timestep={}, alpha={:.6}", strength, timestep, alpha);
-    println!("  original_samples: shape={:?}, dtype={:?}", original_samples.shape(), original_samples.dtype());
-    println!("  noise: shape={:?}, dtype={:?}", noise.shape(), noise.dtype());
-
-    // Keep operations in original_samples' dtype for now
-    let target_dtype = original_samples.dtype();
-    let noise_converted = noise.to_dtype(target_dtype)?;
-
-    // Calculate coefficients
-    let sqrt_alpha = alpha.sqrt();
-    let sqrt_one_minus_alpha = (1.0 - alpha).sqrt();
-
-    println!("  Using scalar multiplication with target dtype={:?}", target_dtype);
-
-    // Apply noise formula
-    let scaled_samples = original_samples.affine(sqrt_alpha, 0.0)?;
-    let scaled_noise = noise_converted.affine(sqrt_one_minus_alpha, 0.0)?;
-
-    println!("  scaled_samples: shape={:?}, dtype={:?}", scaled_samples.shape(), scaled_samples.dtype());
-    println!("  scaled_noise: shape={:?}, dtype={:?}", scaled_noise.shape(), scaled_noise.dtype());
-
-    // Add scaled noise - should maintain dtype
-    let noised = scaled_samples.add(&scaled_noise)?;
-
-    println!("  noised: shape={:?}, dtype={:?}", noised.shape(), noised.dtype());
-
-    // Return noised image and the timestep to start from
-    Ok((noised, timestep))
-  }
-
-  pub fn index_for_timestep(&self, timestep: usize) -> Result<usize> {
-    // Collect all indices where timestep matches
-    let indices: Vec<usize> = self.timesteps.iter().enumerate().filter_map(|(idx, &ts)| if ts == timestep { Some(idx) } else { None }).collect();
-
-    if indices.is_empty() {
-      return Err(candle::Error::Msg(format!("Timestep {} not found in scheduler timesteps", timestep)));
-    }
-
-    // Take the second index if there are multiple matches, otherwise take the first
-    let pos = if indices.len() > 1 { 1 } else { 0 };
-    Ok(indices[pos])
-  }
-
-  pub fn set_begin_index(&mut self, begin_index: usize) {
-    self.begin_index = Some(begin_index);
-    println!("Set begin_index to {}", begin_index);
-  }
-
-  pub fn init_step_index(&mut self, timestep: usize) -> Result<usize> {
-    self.step_index = if let Some(begin_index) = self.begin_index {
-      // If begin_index is set, use it directly
-      Some(begin_index)
-    } else {
-      // Otherwise find the index based on the timestep
-      // Find the position of timestep in self.timesteps
-      let pos = self.timesteps.iter().position(|&t| t == timestep).ok_or_else(|| candle::Error::Msg(format!("Timestep {} not found in timesteps", timestep)))?;
-      Some(pos)
-    };
-
-    Ok(self.step_index.unwrap())
-  }
-
-  // Updated helper function to match diffusers' get_timesteps logic exactly
-  pub fn retrieve_timesteps(scheduler: &mut LCMScheduler, num_inference_steps: usize, strength: Option<f64>, custom_timesteps: Option<Vec<usize>>) -> Result<Vec<usize>> {
-    // 1. When custom timesteps are provided, use them directly
-    if let Some(custom_ts) = custom_timesteps {
-      println!("Using custom timesteps: {:?}", custom_ts);
-      return Ok(custom_ts);
-    }
-
-    // 2. Get timesteps from scheduler
-    let mut timesteps = scheduler.timesteps().to_vec();
-
-    // 3. For img2img, adjust timesteps based on strength
-    if let Some(strength) = strength {
-      if strength < 1.0 {
-        // Match diffusers logic exactly:
-        // init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
-        let init_timestep = ((num_inference_steps as f64 * strength).floor() as usize).min(num_inference_steps);
-
-        // t_start = max(num_inference_steps - init_timestep, 0)
-        let t_start = num_inference_steps.saturating_sub(init_timestep);
-
-        // In diffusers, order is typically 1 for LCM, so we'll simplify here
-        let scheduler_order = 1;
-        let start_idx = t_start * scheduler_order;
-
-        // Slice timesteps from the calculated start position
-        if start_idx < timesteps.len() {
-          println!("Adjusting timesteps for strength {}: starting at index {} (init_timestep={})", strength, start_idx, init_timestep);
-          timesteps = timesteps[start_idx..].to_vec();
-
-          // Set begin_index in the scheduler
-          scheduler.set_begin_index(start_idx);
-        }
-
-        // Return the actual number of inference steps we'll perform
-        let actual_steps = num_inference_steps.saturating_sub(t_start);
-        println!("Will perform {} inference steps due to strength {}", actual_steps, strength);
-      }
-    }
-
-    println!("Final timesteps for inference: {:?}", timesteps);
-    Ok(timesteps)
-  }
-
-  // Add this method to help initialize latents from an image (for img2img)
-  pub fn prepare_img2img_latents(
-    &mut self,
-    image_tensor: &Tensor, // Encoded image from VAE
-    timestep: usize,
-    batch_size: usize,
-    num_images_per_prompt: usize,
-    dtype: DType,
-    device: &Device,
-  ) -> Result<Tensor> {
-    // Store the dtype for consistent operations
-    self.dtype = dtype;
-
-    let latent_dim = image_tensor.dim(2)? / 8; // Assuming standard VAE downsampling
-
-    println!("Preparing img2img latents for timestep={}", timestep);
-    println!("  image_tensor: shape={:?}, dtype={:?}", image_tensor.shape(), image_tensor.dtype());
-
-    // Create noise tensor using the same approach as in main.rs
-    let noise_shape = Shape::from((batch_size * num_images_per_prompt, 4, latent_dim, latent_dim));
-    let noise = Tensor::randn(0f32, 1f32, noise_shape, device)?.to_dtype(dtype)?;
-    println!("  noise: shape={:?}, dtype={:?}", noise.shape(), noise.dtype());
-
-    // Encode image to latents
-    let encoded_latents = image_tensor.clone(); // Assuming already encoded
-
-    // Scale to match VAE expected scale - use affine for scalar multiplication
-    println!("  Using affine for VAE scaling by 0.18215");
-    let scaled_latents = encoded_latents.affine(0.18215, 0.0)?;
-    println!("  scaled_latents: shape={:?}, dtype={:?}", scaled_latents.shape(), scaled_latents.dtype());
-
-    // Add noise based on timestep (this handles the "strength" parameter)
-    let noised_latents = self.add_noise(&scaled_latents, noise, timestep)?;
-    println!("  noised_latents: shape={:?}, dtype={:?}", noised_latents.shape(), noised_latents.dtype());
-
-    Ok(noised_latents)
-  }
-
-  /// Apply classifier-free guidance to model output
-  pub fn apply_guidance(&self, model_output_uncond: &Tensor, model_output_text: &Tensor, guidance_scale: f64) -> Result<Tensor> {
-    // Skip guidance if scale is 1.0 (or very close to it)
-    if (guidance_scale - 1.0).abs() < 1e-5 {
-      return Ok(model_output_text.clone());
-    }
-
-    println!("Applying guidance with scale={:.4}", guidance_scale);
-    println!("  model_output_uncond: shape={:?}, dtype={:?}", model_output_uncond.shape(), model_output_uncond.dtype());
-    println!("  model_output_text: shape={:?}, dtype={:?}", model_output_text.shape(), model_output_text.dtype());
-
-    // Apply the classifier-free guidance formula:
-    // output = uncond + guidance_scale * (text - uncond)
-
-    // First calculate (text - uncond)
-    let guidance_diff = (model_output_text - model_output_uncond)?;
-    println!("  guidance_diff: shape={:?}, dtype={:?}", guidance_diff.shape(), guidance_diff.dtype());
-
-    // Then scale by guidance_scale using affine for scalar multiplication
-    let guidance_factor = guidance_scale - 1.0;
-    println!("  Using affine for scaling by {:.4}", guidance_factor);
-    let scaled_diff = guidance_diff.affine(guidance_factor, 0.0)?;
-    println!("  scaled_diff: shape={:?}, dtype={:?}", scaled_diff.shape(), scaled_diff.dtype());
-
-    // Finally add back to uncond: uncond + guidance_scale * (text - uncond)
-    let guided_output = (model_output_uncond + &scaled_diff)?;
-    println!("  guided_output: shape={:?}, dtype={:?}", guided_output.shape(), guided_output.dtype());
-
-    Ok(guided_output)
-  }
-
-  // Add this helper method to manage dtype consistency across the pipeline
-  pub fn ensure_dtype_consistency(&mut self, input_tensor: &Tensor) -> Result<()> {
-    // Store the current tensor's dtype for reference
-    self.dtype = input_tensor.dtype();
-    println!("Setting LCM scheduler dtype to match input: {:?}", self.dtype);
-    Ok(())
-  }
-
-  // Add a new method to prepare latents with explicit dtype control
-  pub fn prepare_latents(&mut self, batch_size: usize, height: usize, width: usize, num_channels_latents: usize, dtype: DType, device: &Device, generator: Option<u64>) -> Result<Tensor> {
-    // Store the requested dtype
-    self.dtype = dtype;
-
-    // Calculate latent dimensions
-    let latent_height = height / 8;
-    let latent_width = width / 8;
-
-    // Create random latents
-    let shape = Shape::from((batch_size, num_channels_latents, latent_height, latent_width));
-
-    // Note: Candle doesn't have a direct way to set random seeds for tensors
-    // So we'll just use regular random generation for now
-    let latents = if let Some(seed) = generator {
-      println!("Note: Using random latents (seed not supported in this implementation)");
-      // In a real implementation, you would want to set the random seed here
-      Tensor::randn(0f32, 1f32, shape, device)?
-    } else {
-      println!("Generating random latents");
-      Tensor::randn(0f32, 1f32, shape, device)?
-    };
-
-    // Convert to the requested dtype
-    let latents = latents.to_dtype(dtype)?;
-    println!("Generated latents: shape={:?}, dtype={:?}", latents.shape(), latents.dtype());
-
-    // Scale by the init_noise_sigma
-    let scaled_latents = latents.affine(self.init_noise_sigma, 0.0)?;
-
-    Ok(scaled_latents)
-  }
-
-  // Add this method to maintain dtype consistency across operations
-  pub fn match_dtype(&self, tensor: &Tensor, dtype: Option<DType>) -> Result<Tensor> {
-    match dtype {
-      Some(d) => tensor.to_dtype(d),
-      None => Ok(tensor.clone()),
-    }
-  }
-
-  // Add this public getter for the dtype field
-  pub fn dtype(&self) -> DType {
-    self.dtype
-  }
 }
diff --git a/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/mod.rs b/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/mod.rs
index d174da34d2..fe74fcf9d9 100644
--- a/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/mod.rs
+++ b/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/mod.rs
@@ -84,6 +84,8 @@ impl StableDiffusionConfig {
         };
         // https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/unet/config.json
         let unet = unet_2d::UNet2DConditionModelConfig {
+            time_embed_dim: None,
+            time_cond_proj_dim: None,
             blocks: vec![
                 bc(320, Some(1), 8),
                 bc(640, Some(1), 8),
@@ -153,6 +155,8 @@ impl StableDiffusionConfig {
         };
         // https://huggingface.co/stabilityai/stable-diffusion-2-1/blob/main/unet/config.json
         let unet = unet_2d::UNet2DConditionModelConfig {
+            time_cond_proj_dim: None,
+            time_embed_dim: None,
             blocks: vec![
                 bc(320, Some(1), 5),
                 bc(640, Some(1), 10),
@@ -237,6 +241,8 @@ impl StableDiffusionConfig {
         };
         // https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/unet/config.json
         let unet = unet_2d::UNet2DConditionModelConfig {
+            time_cond_proj_dim: None,
+            time_embed_dim: None,
             blocks: vec![
                 bc(320, None, 5),
                 bc(640, Some(2), 10),
@@ -306,6 +312,8 @@ impl StableDiffusionConfig {
         };
         // https://huggingface.co/stabilityai/sdxl-turbo/blob/main/unet/config.json
         let unet = unet_2d::UNet2DConditionModelConfig {
+            time_cond_proj_dim: None,
+            time_embed_dim: None,
             blocks: vec![
                 bc(320, None, 5),
                 bc(640, Some(2), 10),
@@ -405,6 +413,8 @@ impl StableDiffusionConfig {
         };
         // https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/unet/config.json
         let unet = unet_2d::UNet2DConditionModelConfig {
+            time_cond_proj_dim: None,
+            time_embed_dim: None,
             blocks: vec![
                 bc(320, None, 5),
                 bc(640, Some(2), 10),
diff --git a/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/unet_2d.rs b/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/unet_2d.rs
index 6215b5efda..421fbf269c 100644
--- a/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/unet_2d.rs
+++ b/crates/vendor/candle/candle-transformers/src/models/stable_diffusion/unet_2d.rs
@@ -10,6 +10,10 @@ use candle_nn as nn;
 use candle_nn::Module;
 use std::collections::HashMap;
 
+// fn save_tensor_as_numpy(tensor: &Tensor, filename: &str) -> Result<()> {
+//     // Use Candle's built-in NPY serialization
+//     tensor.write_npy(filename).map_err(|e| anyhow::anyhow!("{}", e))
+// }
 #[derive(Debug, Clone, Copy)]
 pub struct BlockConfig {
   pub out_channels: usize,
@@ -33,11 +37,13 @@ pub struct UNet2DConditionModelConfig {
   pub cross_attention_dim: usize,
   pub sliced_attention_size: Option<usize>,
   pub use_linear_projection: bool,
+  pub time_cond_proj_dim: Option<usize>,
+  pub time_embed_dim: Option<usize>,
 }
 
 impl Default for UNet2DConditionModelConfig {
   fn default() -> Self {
-    Self { center_input_sample: false, flip_sin_to_cos: true, freq_shift: 0., blocks: vec![BlockConfig { out_channels: 320, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 640, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 1280, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 1280, use_cross_attn: None, attention_head_dim: 8 }], layers_per_block: 2, downsample_padding: 1, mid_block_scale_factor: 1., norm_num_groups: 32, norm_eps: 1e-5, cross_attention_dim: 1280, sliced_attention_size: None, use_linear_projection: false }
+    Self { center_input_sample: false, flip_sin_to_cos: true, freq_shift: 0., blocks: vec![BlockConfig { out_channels: 320, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 640, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 1280, use_cross_attn: Some(1), attention_head_dim: 8 }, BlockConfig { out_channels: 1280, use_cross_attn: None, attention_head_dim: 8 }], layers_per_block: 2, downsample_padding: 1, mid_block_scale_factor: 1., norm_num_groups: 32, norm_eps: 1e-5, cross_attention_dim: 1280, sliced_attention_size: None, use_linear_projection: false, time_cond_proj_dim: Some(256), time_embed_dim: None }
   }
 }
 
@@ -73,8 +79,16 @@ impl UNet2DConditionModel {
     let b_channels = config.blocks[0].out_channels;
     let bl_channels = config.blocks.last().unwrap().out_channels;
     let bl_attention_head_dim = config.blocks.last().unwrap().attention_head_dim;
-    let time_embed_dim = b_channels * 4;
+    let time_embed_dim = config.time_embed_dim.unwrap_or(b_channels * 4);
     let conv_cfg = nn::Conv2dConfig { padding: 1, ..Default::default() };
+
+    // Log important configuration details
+    println!("Creating UNet2DConditionModel with {} blocks", n_blocks);
+    if let Some(proj_dim) = config.time_cond_proj_dim {
+      println!("UNet configured with time_cond_proj_dim: {}", proj_dim);
+      println!("This means the model expects guidance embeddings of dimension {}", proj_dim);
+    }
+
     let conv_in = conv2d(in_channels, b_channels, 3, conv_cfg, vs.pp("conv_in"))?;
 
     let time_proj = Timesteps::new(b_channels, config.flip_sin_to_cos, config.freq_shift);
@@ -241,18 +255,138 @@ impl UNet2DConditionModel {
 
     // Add guidance embedding if provided
     if let Some(guidance_emb) = guidance_embedding {
-      // Print shapes for debugging
-      println!("Guidance embedding shape: {:?}", guidance_emb.shape());
-      println!("Time embedding shape: {:?}", time_emb.shape());
+      // Get dimensions of both embeddings
+      let (batch_size, time_emb_dim) = time_emb.dims2()?;
+      let (guidance_batch, guidance_emb_dim) = guidance_emb.dims2()?;
 
-      // Get the batch size and embedding dimension
-      let (batch_size, emb_dim) = time_emb.dims2()?;
+      println!("==== Guidance Embedding Integration ====");
+      println!("Time embedding dims: [{}, {}]", batch_size, time_emb_dim);
+      println!("Guidance embedding dims: [{}, {}]", guidance_batch, guidance_emb_dim);
 
-      // Reshape guidance embedding to match time embedding dimensions
-      let guidance_emb = guidance_emb.reshape(&[1, emb_dim])?.expand(&[batch_size, emb_dim])?;
+      // Use time_cond_proj_dim from config to determine expected guidance embedding dimension
+      let expected_guidance_dim = self.config.time_cond_proj_dim.unwrap_or(time_emb_dim);
+      println!("Expected guidance dimension from config: {}", expected_guidance_dim);
+
+      // Check if dimensions match the expected projection dimension
+      let processed_guidance_emb = if guidance_emb_dim != time_emb_dim {
+        // Convert guidance_emb to f32 for calculations
+        let guidance_emb_f32 = guidance_emb.to_dtype(DType::F32)?;
+        println!("Converting guidance embedding to F32 for processing");
+
+        // If the guidance embedding dimension matches the expected projection dimension,
+        // we need to project it to match the time embedding dimension
+        if guidance_emb_dim == expected_guidance_dim {
+          println!("Projecting guidance embedding from {} to {}", guidance_emb_dim, time_emb_dim);
+
+          // Use linear interpolation
+          let src_positions: Vec<f32> = (0..guidance_emb_dim).map(|i| i as f32 / (guidance_emb_dim - 1) as f32).collect();
+          let dst_positions: Vec<f32> = (0..time_emb_dim).map(|i| i as f32 / (time_emb_dim - 1) as f32).collect();
+
+          // Get guidance embedding data (now f32)
+          let guidance_data = guidance_emb_f32.reshape((guidance_emb_dim,))?.to_vec1::<f32>()?;
+          println!("Reshaping guidance data to vector of length {}", guidance_data.len());
+
+          // Perform linear interpolation
+          let mut interpolated_data = Vec::with_capacity(time_emb_dim);
+          for &t in &dst_positions {
+            // Find the two source positions that surround the target position
+            let mut i = 0;
+            while i < src_positions.len() - 1 && src_positions[i + 1] < t {
+              i += 1;
+            }
+
+            if i == src_positions.len() - 1 {
+              // At or beyond the last position
+              interpolated_data.push(guidance_data[i]);
+            } else if t <= src_positions[0] {
+              // At or before the first position
+              interpolated_data.push(guidance_data[0]);
+            } else {
+              // Linear interpolation between two points
+              let t_norm = (t - src_positions[i]) / (src_positions[i + 1] - src_positions[i]);
+              let value = guidance_data[i] * (1.0 - t_norm) + guidance_data[i + 1] * t_norm;
+              interpolated_data.push(value);
+            }
+          }
+
+          println!("Created interpolated data of length {}", interpolated_data.len());
+
+          // Create tensor and convert to time_emb's dtype
+          let tensor = Tensor::from_vec(interpolated_data, (1, time_emb_dim), device)?;
+          let result = tensor.to_dtype(time_emb.dtype())?;
+          println!("Final projected guidance embedding shape: {:?}", result.shape());
+          result
+        } else {
+          // Unusual case - dimensions don't match expected pattern
+          println!("Warning: Guidance embedding dimension ({}) doesn't match expected projection dimension ({})", guidance_emb_dim, expected_guidance_dim);
+
+          // Fallback to generic interpolation
+          let guidance_data = guidance_emb_f32.reshape((guidance_emb_dim,))?.to_vec1::<f32>()?;
+
+          let src_positions: Vec<f32> = (0..guidance_emb_dim).map(|i| i as f32 / (guidance_emb_dim - 1) as f32).collect();
+          let dst_positions: Vec<f32> = (0..time_emb_dim).map(|i| i as f32 / (time_emb_dim - 1) as f32).collect();
+
+          let mut interpolated_data = Vec::with_capacity(time_emb_dim);
+          for &t in &dst_positions {
+            let mut i = 0;
+            while i < src_positions.len() - 1 && src_positions[i + 1] < t {
+              i += 1;
+            }
+
+            if i == src_positions.len() - 1 {
+              interpolated_data.push(guidance_data[i]);
+            } else if t <= src_positions[0] {
+              interpolated_data.push(guidance_data[0]);
+            } else {
+              let t_norm = (t - src_positions[i]) / (src_positions[i + 1] - src_positions[i]);
+              let value = guidance_data[i] * (1.0 - t_norm) + guidance_data[i + 1] * t_norm;
+              interpolated_data.push(value);
+            }
+          }
+
+          println!("Created interpolated data of length {}", interpolated_data.len());
+          let tensor = Tensor::from_vec(interpolated_data, (1, time_emb_dim), device)?;
+          let result = tensor.to_dtype(time_emb.dtype())?;
+          println!("Final fallback guidance embedding shape: {:?}", result.shape());
+          result
+        }
+      } else {
+        // Dimensions match, no processing needed
+        println!("Guidance dimension already matches time embedding dimension, no interpolation needed");
+        let result = guidance_emb.reshape((1, time_emb_dim))?.to_dtype(time_emb.dtype())?;
+        println!("Reshaped guidance embedding shape: {:?}", result.shape());
+        result
+      };
+
+      // Expand to match batch size if needed
+      let expanded_guidance = processed_guidance_emb.broadcast_as(time_emb.shape())?;
+      println!("Expanded guidance embedding shape: {:?}", expanded_guidance.shape());
+
+      // Save the processed and expanded guidance embeddings to disk if requested
+      if let Ok(save_dir) = std::env::var("CANDLE_SAVE_TENSORS_DIR") {
+        // Save pre-expansion (original projection)
+        let proc_path = format!("{}/unet_processed_guidance_embedding.npy", save_dir);
+        println!("Saving processed guidance embedding to {}", proc_path);
+        if let Err(e) = processed_guidance_emb.write_npy(&proc_path) {
+          println!("Error saving processed guidance embedding: {}", e);
+        } else {
+          println!("Successfully saved processed guidance embedding to {}", proc_path);
+        }
+
+        // Save post-expansion (final tensor used for addition to time_emb)
+        let exp_path = format!("{}/unet_expanded_guidance_embedding.npy", save_dir);
+        println!("Saving expanded guidance embedding to {}", exp_path);
+        if let Err(e) = expanded_guidance.write_npy(&exp_path) {
+          println!("Error saving expanded guidance embedding: {}", e);
+        } else {
+          println!("Successfully saved expanded guidance embedding to {}", exp_path);
+        }
+      }
 
       // Add to time embeddings
-      time_emb = (time_emb + guidance_emb)?;
+      time_emb = (time_emb + expanded_guidance)?;
+      println!("Final time embedding with guidance shape: {:?}", time_emb.shape());
+      println!("==== End Guidance Integration ====");
     }
 
     // 2. pre-process