Activations stay on the device between layers (net/program.rs lowers the graph; every
operator is a WGSL kernel, one submission and read-back per run); faces of a photo go
in batches of up to 16; Settings > Faces shows each model's device and a CPU-only
switch (faces.gpu); the scan's pace limits GPU use; the GPU check is an activity row.
YuNet 58 -> 21 ms, SFace 40 -> 9.6 ms, AuraFace 295 -> 71 ms (heavily loaded machine).
Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An accelerator hook in the face interpreter takes Conv nodes; lightcraft-gpu implements it with a WGSL
implicit-GEMM convolution (any stride, padding, grouping; split-K for small maps; a direct kernel for depthwise
layers), each model's weights resident on a device of its own. After loading, detection and recognition check
the GPU in the background (agreement within 1e-3 and faster than the CPU) and otherwise stay on the CPU, which
remains the reference; the first GPU error drops it. Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An unedited raw's face tile is cut from its largest embedded preview, read
the planned way (header and JPEG, never the sensor data), when the preview
holds at least 3/4 of the tile's pixels across the face; otherwise, and for
edited photos and other files, it is developed as before. Tiles keep going
to the memory + on-disk thumbnail cache, now at 16 px size steps so the
thumbnail slider reuses them. The planned preview path carries an Adobe RGB
preview colour space and applies it instead of declining to the whole-file
read (Sony alpha 1 previews).
A sibling of spot.findDust for portraits: with the YuNet detector installed, small
round spots on each face's skin that are darker or redder than the surrounding skin
become heal spots (one undo step). The detector's box and landmarks define the skin
(ellipse, skin-colour gate, eye band / brows / nostrils / mouth cut out); a candidate
must stand out from a blurred copy beyond the sensitivity threshold and 2.5x the local
texture, then grow into a small roundish blob. Conservative: faces under 160 px get
nothing, faces with more than 12 candidates (stubble, freckles, pores) get nothing, at
most 6 per face (maxPerFace). Classical, no model beyond the detector.
Tests (synthetic, no model): cheek spots found at their positions, clean face empty,
spots on features ignored, small face ignored, speckled face treated as texture,
hostile boxes and an empty image do not panic. New LR-REM-BLEMISH parity row, Find
Blemishes section in docs/faces.md.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Settings ▸ Faces let the switch be turned on with no recognition model
chosen or installed: the box ticked, Settings closed without a word, and
recognition was on with nothing to run. Now the box is enabled only with a
chosen, installed recogniser (and a build that can run models), says
"Choose a model below (Use) to start." on hover, and that hint stays under
it on or off; a box left on by an older build can still be switched off.
faces.enable {enabled: true} refuses without such a model, so the control
channel and MCP cannot reach that state either, and removing the last
recogniser switches recognition off instead of leaving it on with nothing
chosen. Installing a model still chooses it and switches recognition on.
Tests: tests_face_models::recognition_cannot_be_turned_on_without_a_chosen_model
and tests_panels::recognise_faces_stays_off_until_a_model_is_chosen (fails
on main at the first click). Docs: faces.md, parity LR-PREF-PEOPLE.
CLAUDE.md gains the rule (a TaskGuard for every job that runs off the UI
thread and can take more than a moment; no progress window, panel or
toast of its own), and docs/background-tasks.md says how: kind, label,
progress, units, detail, the three ways to cancel, examples and tests.
Photo Merge's and the face scan's docs no longer describe the old
progress toast.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep current upstream UI behavior and the later detector contract checks. Validate candidate models before replacing working weights and records, with a synthetic regression that failed before the fix. Remove the unrelated GPU equivalence-test relaxation. All seven quality gates pass with the pinned hosted fonts.
Generated with Codex.
Photo > Detect Faces runs the YuNet detector (a small pure-Rust interpreter over the ONNX graph, no
runtime dependency) and adds the faces it finds as unnamed face regions in one undo step. The detector
and the recognition models are opt-in downloads: Settings > Faces shows each model's licence and
training-data note, Download stays disabled until the terms are accepted, and the file is fetched over
https with lightcraft-fetch from the model's own repository at a commit-pinned address, checked against
its recorded size and SHA-256 and then installed. Detect Faces offers the detector's download the first
time it is used. No model file is part of the repository (the real-model tests read LC_YUNET_MODEL; two
ignored tests download from the real hosts).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
