Halves the weight memory (Large: about 900 MB to 450 MB). Off by default: the
video logit error grows from about 0.003 to 0.06-0.12 (masks unchanged) for a
speed gain of under 10 percent.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
egui_kittest test: a scripted tracker (sure, doubtful, unsure, absent, sure again)
tracks a layer and the Layer panel is rendered; the strip's pixels are green, amber,
red, slate and green at the right frames. ROTO_SPAN_SNAPSHOT saves the panel image.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Strided batched matmul with bias, GELU/ReLU and residual in the epilogue, layer
norm, softmax, window partition, pooling, patch embedding, RoPE and neck kernels
(WGSL, all tested against CPU definitions). GpuEncoder keeps activations on the
GPU and returns the 64x64 level and the decoder's skip features. On the RX 7900
GRE one frame encodes in 0.29 s (debug build) against 2.1 s on the CPU in release;
errors 6.8e-6 vs PyTorch, 1.2e-5 vs the CPU path.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Candle adds a bias on one thread; the Linear layers now add it (and apply GELU
for the MLP in) in one parallel pass. Encode 2.4 s -> 2.1 s, memory attention
2.4 -> 2.3 s; reference tests unchanged. gpu_matmul example: a tiled WGSL
matmul on the RX 7900 GRE reaches 1.9-3.6 TFLOPS against 0.16-0.47 for the CPU.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Candle's gelu_erf runs on one thread and was a third of the Hiera block time.
A rayon CustomOp1 calls the same erf per element (bit-identical, tested). All
reference tests unchanged. op_bench example times the block's operations.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
run_tracker follows the object with the model's video tracker on the same
synthetic clips (--sam2 FILE or the installed weights). 12 frames, dev build:
re-entry clip mean IoU 0.632 classical vs 0.975 SAM 2 (0.000 vs 0.967 after
the object returns); the other clips 0.98 vs 0.99.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Sam2Mask wraps the image path behind the Roto Brush's MaskModel interface
(cached frame encoding, bilinear mask resize). The engine loads the verified
weights natively (not on wasm). Checked with the real file: install, notice,
load, use by Roto Brush 3.0, deselect.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Weights loader with shape/finite validation (S2.10), patch embed, bicubic
positional embedding, window/global attention, query pooling. Reference test
(skips without weights): max abs error 2.6e-5 over all 12 blocks.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Synthetic Roto Brush clips with known mattes (movement, lighting change,
occlusion, leaving and re-entering) and a runner that compares the classical
engine with MobileSAM when installed. Needs no weights; SAM 2 plugs in later.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
UPDATE_APP gains our accent (#e0368f, our colour in ArtCraft Launcher). The prompt draws it
on the UPDATE chip, the buttons and the What's new underline, and says "by ArtCraft"
(craft-update's default developer); #2646 gave it our icon.
craft-update is pinned to storytold/craft-libs main at 1065653e7 (#434).
- `roxmltree`: 0.21.1 → 0.21.2
- `zip`: 9.1.0 → 9.1.3
Lockfile only. `symphonia` deliberately stays 0.5.5: effectcraft names it only to enable AIFF in filmcraft-codecs' symphonia, which is 0.5.5 at the pinned filmcraft rev. Moving to 0.6 waits for that pin to move.
The gains of an AAC source now come from the configuration FilmCraft decodes
it with, read from MP4/MOV (esds), Matroska (CodecPrivate) and LATM in MPEG
TS/PS (its first StreamMuxConfig). Channel configurations 3 to 6, 11 and 12
arrive reordered to WAVE order, 4 as L R C Cs, 7 in element order. A program
config element goes by its element positions, except the six structures in
which ffmpeg's encoder fills its own channel order across them (an LFE in a
side SCE, a back centre in a side pair), recognised by its "Lavc" comment.
ADTS can only signal configurations 1 to 7 that FilmCraft decodes, which the
channel count tells apart. A decode of fewer channels than the layout keeps
its first ones (7 of configuration 7's 8 in MPEG-TS LATM).
FilmCraft reorders AAC configurations 3 to 6, 11 and 12 to WAVE/SMPTE order
and leaves configuration 7 and program config elements as coded, so the
channel count alone can't say which order arrives: ffmpeg writes 7.1 as
configuration 7 (C L R first) and 6.1 as a PCE it fills in its own order
(L R C LFE Cs Ls Rs), while FilmCraft's encoder writes a centre-first PCE.
The order is now read from an MP4/MOV's AudioSpecificConfig when the source
opens; other containers keep a channel-count guess.
The ffmpeg oracle test gains configuration 12 (ffmpeg's 7.1 relabelled) and
skips PCE files ffmpeg 4.2 marks as stereo, which FilmCraft reads as two
channels.
CodeRabbit and FredApps both flagged that the preflight probed only the plain
PKG_CONFIG variable, while the pkg-config crate alsa-sys pins resolves the
target-scoped variants first and falls back pkg-config -> pkgconf; on a host
with only pkgconf the ci stage failed a build that would succeed. The probe
now goes through the same crate (0.3.34, already pinned via alsa-sys), so the
semantics match by construction, and a probe that ran but said no stays
distinct from one that could not run.
Also from review: audio_backend_hint is gated to Linux (bflatastic: dead code
on Windows/macOS under -D warnings), the fake-tool regression runs the real
xtask audio-backend task in a child process with a controlled environment
(pkgconf-only host, a probe that says no, one that cannot run, a HOST_PKG_CONFIG
override), the rerun-if-env lines stay out of the ci log, and CONTRIBUTING.md
asks for a usable alsa.pc and runtime library rather than development headers.
* fix(ui-egui): craft_ui::ComboBox for every combo box, so open menus grow back (egui#5225)
Repins craft-libs to aiongg/craft-libs 60d60d48 (storytold main with the shared
dropdown) and bans egui::ComboBox in clippy.toml.
* fix(ui-egui): font menu under the font field is a craft_ui combo menu that grows back
The filter row is the menu's header and only the font list scrolls, so a search
that matched nothing grows the menu back when it is cleared.
* chore(deps): craft-libs 4d7b82f7 (dropdown fixes)
* chore(deps): craft-libs 92d392cf (#387 merged)
* chore(xtask): pdf-deps expects craft-libs 92d392cf
Pin craft-libs crates to 3723e32a, which adds the group extra-tabs hook.
The comp tabs follow the panel tabs directly, with no drag bar between
them, and the group header holds only the lock and the panel menu.
craft-libs crates move to one rev, 37c907d2 (contains main's craft-update and craft-docpath pins).
Docs: native floating windows are Windows, macOS and X11 only; the UI estimate drops the
multi-monitor panel work that is done.
Picks up the startup prompt without the glyphs a font may lack (storytold/craft-libs#343, #344): the close mark is painted and the version pill is ASCII, so apps that panic on a missing glyph no longer take the prompt down.
Part of tracking issue storytold/craft-libs#319
Picks up the update prompt fix that draws its close mark and version pill
without glyphs a font may lack, so the prompt no longer takes the app down
when a bundled face has no ✕ or →.
Part of tracking issue storytold/craft-libs#319🤖 Generated with Codebuff
Co-authored-by: Codebuff <noreply@codebuff.com>
Picks up the fix that keeps the close request pending while the app holds unsaved work, the executable-name check in the Windows update helper, and the zip 9 / dirs-free build.
Part of tracking issue storytold/craft-libs#319
Startup check is opt-in for the desktop app, the Windows update helper is dispatched before argument parsing, and a successful update closes through the app's own close path.
aiongg
Brandon ThomasandClaude Opus 5.5
bflatastic
Kristian Myram
Dylan M. Taylor
tylerjenningsw
Kaedo17
dotCooCoo
aiongg
midasmatterhorn
skylar
