dtourolleandClaude Opus 5
9a2b39b8e5
`models/LICENCE.md` recorded, on 2026-08-21, that no YOLO model trained on ADE20K existed in usable form — the stuff classes photography cares about, sky and vegetation and water, had no model to come from. Re-checked 2026-08-30: Ultralytics now ships a `semantic` task with ADE20K checkpoints, so `models/scene/` holds `yolo26s-sem-ade20k`. This is an addition, not a replacement. A semantic model labels every pixel but merges same-class pixels into one region, so it cannot tell three people apart — which is exactly what clicking a subject needs, and exactly what `segment/`'s COCO instance model already does. The scene tab grades per category and does not care that instances are merged. Keeping both is the point. ## The export is truncated, deliberately Ultralytics ends the graph with `Resize -> ArgMax -> Cast` and hands back a `[1, 640, 640]` u8 label map. The script cuts that tail and exposes the classifier's `[1, 150, 80, 80]` f32 logits instead, for two reasons. Cost: the Resize materialises 150 x 640 x 640 x f32, 246 MB, and ArgMax then reduces across the channel axis, striding 409,600 elements per comparison. On one loaded machine the full graph ran ~1160 ms against ~500 ms truncated — roughly four fifths of the time spent on work the application discards. Those numbers were measured under contention and are upper bounds, but the ratio is structural. Softness: ArgMax destroys the per-class scores, and the scene tab needs them. Softmax over the 150 channels, summed within each photographic category, yields per-category weights summing to 1 at every pixel. Feathering a partition of unity cannot double-grade a boundary, whereas feathering hard labels outward from two adjacent categories paints both grades into the overlap and haloes every horizon. The discarded upsample was never information: the graph's true spatial resolution is the 80x80 logit grid, and the application can resample from that itself. The tail is matched by op type and asserted before cutting, so an upstream graph change fails loudly in the exporter rather than quietly shipping a differently-shaped model. Nothing reads these weights yet — the decode path, the category descriptor grouping 150 classes into ~8 photographic ones, and the scene tab are still to come. At 24 MB this model also wants the runtime-asset treatment `models/face/` already gets on Android rather than `include_bytes!`; embedding it would put ~35 MB of weights in the binary. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
DarkRoom
A cross-platform, non-destructive RAW photo editor for Linux and Android.
Status: 0.9.0, and no longer a spike. A library opens, culls, develops and exports on both platforms, across eight tagged releases. What is not built is written down rather than merely absent — see docs/outstanding.md for the requirements that have no implementation and why, and docs/technical-debt.md for the compromises that were chosen.
Documentation
| Document | Contents |
|---|---|
| CONTRIBUTING.md | How to land a first change without reading the rest |
| requirements.md | What the software must do — 179 numbered requirements |
| architecture.md | How it is built — crates, GPU pipeline, data model, sync |
| technical-debt.md | Compromises taken deliberately, each with the condition that retires it |
| outstanding.md | What is not built, and whether that is a decision or a gap |
| code-health.md | What a contribution costs, per seam, measured |
| traceability.md | Generated: which requirement is claimed by which file |
| faces.md | Face detection and identity — the models, the licence problem, and what S14 measured |
Building
Desktop:
cargo run -p darkroom-desktop
Android (containerised toolchain, see docker/android):
./docker/android/build.sh cargo ndk -t arm64-v8a build --release
Git LFS is required for the model weights, and the toolchain pins itself. CONTRIBUTING.md has the details and the four commands CI will run against what you send.
Current state
Working. A catalog over a local folder, a Nextcloud account, or a folder a
sync client keeps in virtual-files mode — where a placeholder is treated as the
photograph rather than as a one-byte file. A virtualised library grid with a
capture-time timeline, ratings, labels, keywords, collections and a trash that
survives a crash mid-operation. Card ingest. Face detection and identity, with
the index syncing between devices. A develop pipeline of fifteen declared
operations fused into a single compute dispatch, plus the neighbourhood
operations that cannot be — clarity, texture, capture sharpening, noise
reduction, lens correction, spectral film simulation. Crop, straighten, spot
removal, gradient and subject-segmentation masks, named presets, and a
generated panel that no operation in ui/ is allowed to name. Export to JPEG,
PNG and 8- or 16-bit TIFF with resize and output sharpening.
The zero-copy display path works on desktop. The compute pass writes a texture that Slint composites directly, which is what ARCH §6.1 requires; the readback it forbids costs 96% of frame time at 4K, and
cargo run -p dr-gpu --example bench --features readback
still reproduces that measurement. The one exception is the Android develop view, which reads the frame back through the CPU because zero-copy there needs wgpu's Vulkan swapchain, and that tears a portrait window on a tablet whose panel is mounted landscape. It is debt, not a revision of the rule: the reasoning, the on-device measurements that forced it, and the three separate things any one of which would remove it are in technical-debt.md TD-1.
Not built. Plugins, compare and survey culling, focus peaking, burst grouping, AI denoise, tiled and progressive rendering, and most of the Android platform integration beyond running. The performance targets in §4.1 are unverified rather than unmet — the per-commit benchmark suite §8 requires does not exist, so nothing fails a build on a regression. docs/outstanding.md is the list, with the reasoning.
Licence
GPL-3.0-or-later.