S2.65: help and bench follow all four SAM 2 sizes; tests for the size configs and detection

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

apps/effectcraft-cli/src/main.rs

@@ -28,7 +28,7 @@
 //! effectcraft-cli bench --ops [--small] [--layers N] [--comps N] [--footage N]   everyday-operation timings
 //!     on a large generated project (open, save, auto-save, undo/redo, timeline, Project panel)
 //! effectcraft-cli bench --roto [--frames N] [--sam2 model.safetensors] [--mobilesam mobile_sam.pt] [--fixes] [--clip FRAMES --truth MASKS] [--json]   Roto Brush trackers on synthetic clips with known mattes
-//!     (classical engine, MobileSAM and SAM 2.1 Tiny when installed; IoU, flicker, re-appearance, seconds per frame)
+//!     (classical engine, MobileSAM and SAM 2.1 (any installed size) when installed; IoU, flicker, re-appearance, seconds per frame)
 //! effectcraft-cli script FILE.jsx [F.ecproj] | --eval CODE    run an After Effects-style script
 //! effectcraft-cli mcp [--autosave | --bridge PORT]            MCP server on stdio
 //!
@@ -1030,11 +1030,11 @@ fn bench_roto(args: &Args) -> Result<(), Failure> {
         note!("bench --roto: MobileSAM is not installed; comparing the classical engine only");
     }
     // `--sam2 FILE` uses the weights at FILE; otherwise the installed ones.
-    let sam2_file = args.opt("--sam2").map(std::path::PathBuf::from).or_else(|| session.model_path(&effectcraft_segment::SAM2_TINY));
+    let sam2_file = args.opt("--sam2").map(std::path::PathBuf::from).or_else(|| effectcraft_sam2::mask::SIZES.iter().find_map(|m| session.model_path(m)));
     let sam2: Option<std::sync::Arc<dyn MaskModel>> =
         sam2_file.and_then(|p| std::fs::read(p).ok()).and_then(|bytes| effectcraft_sam2::mask::load(&bytes).ok()).map(|m| m as std::sync::Arc<dyn MaskModel>);
     if sam2.is_none() {
-        note!("bench --roto: SAM 2.1 Tiny is not installed; no SAM 2 column");
+        note!("bench --roto: no SAM 2.1 size is installed; no SAM 2 column");
     }
     // `--fixes`: also count the correction strokes each needs before every frame reaches IoU 0.9 (up to 6).
     let fixes = args.flag("--fixes");
@@ -1053,7 +1053,8 @@ fn bench_roto(args: &Args) -> Result<(), Failure> {
             runs.extend(measure(&sc, &|s| Some(bench::run_flow(s, "mobilesam", Some(m.as_ref())))));
         }
         if let Some(m) = &sam2 {
-            runs.extend(measure(&sc, &|s| bench::run_tracker(s, "sam2-tiny", m.as_ref())));
+            let label = m.info().id.replacen("sam2.1-hiera-", "sam2-", 1);
+            runs.extend(measure(&sc, &|s| bench::run_tracker(s, &label, m.as_ref())));
         }
     }
     if args.flag("--json") {

crates/engine/src/analysis/models.rs

@@ -544,7 +544,7 @@ pub fn specs() -> Vec<CommandSpec> {
             "Use Roto Brush Model",
             [],
             None,
-            "{id: classical|mobilesam|sam2.1-hiera-tiny, wait?} (Roto Brush 2.0 / 3.0 use it; wait: loaded before returning)",
+            "{id: classical|mobilesam|sam2.1-hiera-tiny|sam2.1-hiera-small|sam2.1-hiera-base-plus|sam2.1-hiera-large, wait?} (Roto Brush 2.0 / 3.0 use it; wait: loaded before returning)",
             always,
             |s, p| select(s, p, Task::Mask)
         ),

crates/engine/src/tests/models.rs

@@ -37,7 +37,11 @@ fn models_are_listed_chosen_and_verified() {
         let list = s.execute_checked(&format!("{prefix}.models"), json!({})).unwrap();
         let ids: Vec<&str> = list["models"].as_array().unwrap().iter().map(|m| m["id"].as_str().unwrap()).collect();
         // Roto Brush also offers SAM 2.1 in four sizes (video trackers; see crates/sam2).
-        let expect: Vec<&str> = if task == Task::Mask { vec!["classical", model, "sam2.1-hiera-tiny", "sam2.1-hiera-small", "sam2.1-hiera-base-plus", "sam2.1-hiera-large"] } else { vec!["classical", model] };
+        let expect: Vec<&str> = if task == Task::Mask {
+            vec!["classical", model, "sam2.1-hiera-tiny", "sam2.1-hiera-small", "sam2.1-hiera-base-plus", "sam2.1-hiera-large"]
+        } else {
+            vec!["classical", model]
+        };
         assert_eq!(ids, expect);
         let m = &list["models"][1];
         assert_eq!((m["licence"].as_str(), m["installed"].as_bool()), (Some("Apache-2.0"), Some(false)));

crates/sam2/src/lib.rs

@@ -107,6 +107,36 @@ mod tests {
         assert!(Size::Large.repo().ends_with("large"));
     }
 
+    #[test]
+    fn configs_match_the_released_sizes() {
+        let blocks = |s| hiera::HieraConfig::for_size(s).blocks.iter().sum::<usize>();
+        assert_eq!([Size::Tiny, Size::Small, Size::BasePlus, Size::Large].map(blocks), [12, 16, 24, 48]);
+        for s in [Size::Tiny, Size::Small, Size::BasePlus, Size::Large] {
+            let c = hiera::HieraConfig::for_size(s);
+            assert_eq!((c.embed_dims.len(), c.heads.len(), c.windows.len()), (4, 4, 4));
+            assert!(c.global_blocks.iter().all(|g| *g < blocks(s)));
+        }
+    }
+
+    #[test]
+    fn size_comes_from_the_patch_embedding_and_depth() {
+        use candle_core::{DType, Device, Tensor};
+        let make = |width: usize, deep: bool| {
+            let mut m = std::collections::HashMap::new();
+            m.insert("vision_encoder.backbone.patch_embed.projection.weight".to_string(), Tensor::zeros((width, 3, 7, 7), DType::F32, &Device::Cpu).unwrap());
+            if deep {
+                m.insert("vision_encoder.backbone.blocks.15.layer_norm1.weight".to_string(), Tensor::zeros(4, DType::F32, &Device::Cpu).unwrap());
+            }
+            weights::Weights::from_map(m).unwrap()
+        };
+        assert_eq!(Size::detect(&make(96, false)).unwrap(), Size::Tiny);
+        assert_eq!(Size::detect(&make(96, true)).unwrap(), Size::Small);
+        assert_eq!(Size::detect(&make(112, false)).unwrap(), Size::BasePlus);
+        assert_eq!(Size::detect(&make(144, false)).unwrap(), Size::Large);
+        assert!(Size::detect(&make(64, false)).is_err());
+        assert!(Size::detect(&weights::Weights::from_map(Default::default()).unwrap()).is_err());
+    }
+
     #[test]
     fn errors_say_what_failed() {
         assert!(Error::Input("empty frame".into()).to_string().contains("empty frame"));

docs/roto-brush-sam2.md

@@ -106,7 +106,7 @@ encoder's 64 × 64 level matches PyTorch to 7e-6, the memory attention to 1.8e-5
 times the pieces on both devices (`EFFECTCRAFT_SAM2_PROFILE=1` prints a frame's breakdown).
 
 ## Benchmark
-`effectcraft-cli bench --roto [--frames N] [--sam2 model.safetensors] [--mobilesam mobile_sam.pt]` runs synthetic clips with
+`effectcraft-cli bench --roto [--frames N] [--sam2 model.safetensors (any size; default: the first installed)] [--mobilesam mobile_sam.pt]` runs synthetic clips with
 known mattes through all three. 24 frames, debug build, SAM 2 on the GPU; mean IoU (mean IoU after the object returns), seconds
 per frame: