[YokiToki]

What

• New OpenSpec change 32-photo-trace-ai (planning only, no app code) and its row in openspec/ROADMAP.md. Target version 1.3.0, depends on 25a and 29.
• A standalone PoC in poc/ai/ (npm run poc:ai, not part of the app) that runs background-removal models on the straightened sheet in the browser (ONNX Runtime Web) and compares them with the classic segmenter. Results: poc/ai/FINDINGS.md. Adds onnxruntime-web as a dev dependency.

Why

The colour segmenter for photo cutouts is a set of heuristics with known weak spots. A neural engine, run in the browser, is added as an alternative. The classic path stays the default and is the baseline.

Plan highlights

• Engine switch Classic | AI; both produce the same Segmentation, so everything downstream is shared. segment.ts, colour.ts, outline.ts stay untouched (guarded by a frozen-output test).
• Hybrid edge: the network says what is the object, the classic paper-colour test places the exact edge (0.5 mm target).
• Model manager window: list, download, cancel, delete, choose the active model; weights stay in the browser (Cache API), several models side by side.
• Licence classes (permissive / non-commercial / copyleft) with per-class consent; non-permissive models (e.g. RMBG-1.4) are optional third-party integrations, never default, never bundled or hosted by the project.
• Backend is chosen per model (WebGPU / WASM), with capability detection.
• Task 1 is a gate: candidates are measured on the reference photos before the rest is built.

PoC findings so far (owner's photos, one thread)

• BiRefNet-lite at 1024 fails with std::bad_alloc on both backends: dropped.
• InSPyReNet fp16 (MIT): good, ~2 s on WebGPU, empty mask on WASM.
• ormbg (Apache-2.0): good, WASM only, 15–20 s.
• ISNet imgly fp16 (MIT): 0.25 s on WebGPU, misses part of the cutters.
• RMBG-1.4 (non-commercial): best quality, ~6 s on WASM.
• MVANet and BEN2 mark the whole sheet: dropped. Most networks lose part of the cutters that the classic path finds, so AI is an alternative for hard photos, not a better default.
• Still to measure: the no-flash photo on the shortlist, BiRefNet-lite 512, multi-threaded WASM.

For reviewers

Open questions are in design.md: huggingface.co vs self-hosted weights, minimum target device. Click-to-select (SAM) is deliberately out of scope.

🤖 Generated with Claude Code