AI-assisted automatic development of RAW photographs.

RAW decode and orientation

src/decode.rs uses rawler for RAW decode, 24 formats, and the database covers 725 bodies. Bayer files take rawler's normal demosaic path; non-2×2 RGB CFA data uses Autoshop's X-Trans geometric path, which fits color planes over a 5×5 neighborhood per CFA phase while retaining the measured photosite channel. That path closes zero-sample holes but remains an approximate X-Trans develop rather than a directional Markesteijn-class demosaic. src/render.rs applies EXIF orientation at the head of the displayed chain, before masks, straighten, and crop.

Develop engine and measured Lightroom parity

src/render.rs is an f32 pipeline with explicit linear-light operations where the algorithm requires them; the standard rawler output is gamma-encoded f32, so the implementation does not pretend every stage is uniformly linear. After orientation and optional denoise, anchored white balance precedes lens/manual vignetting and linear-light dehaze; exposure, contrast, whites, blacks, highlights, shadows, and the base/tone curve are combined in the tone LUT, then RGB curves, HSL, and color grading run in that order. Clarity and Texture, global color/detail, local masks, lens geometry, straighten, and crop follow; the Highlights control belongs to the tone LUT, with no separate highlight-reconstruction pass claimed.

The parity work in src/render.rs is measurement-driven. Period/step-response measurements refuted the earlier band-limited notch model for negative Texture; the current operator mixes fine Gaussian and coarse box low-passes against 45 anchors spanning nine periods and five slider levels. Radial feather uses a measured 290×11 (radius, feather) alpha LUT, with feather zero kept as an analytic hard edge. Brush dabs use k = (1 - ρm(h))n(h), cubic fits for ln(m) and ln(n) over hardness, screen accumulation, and a measured flow law; the held-out kernel RMS is 0.0109.

Masks and local segmentation

src/recipe.rs and src/render.rs implement radial, linear, brush, bitmap, luminance-range, and color-range masks with Add/Subtract/Intersect composition. src/segment.rs and python/segment.py add local BiRefNet subject selection, U²-Net fallback, OneFormer sky segmentation, and SAM 2.1 point-prompted object gestures. Cached alphas record the photo, mask subtype, orientation/click data, and backend generation, so provenance changes trigger re-derivation rather than silent reuse.

Lens correction and mask-coordinate transport

src/lensmeta.rs reads camera metadata corrections, including Sony tag 0x7037 distortion as a 16-sample piecewise-linear spline; the related camera knots are applied during render. src/lcp.rs reads Adobe .lcp perspective polynomials and solves their inverse for Lightroom mask-coordinate transport when camera knots are unavailable, while refusing unsupported fisheye profiles.

Measurements against real Lightroom exports showed that brush masks are rasterized before lens correction while radial masks are interpreted after it. src/render.rs and src/xmp.rs therefore transport mask coordinates through the engine's inverse geometry model instead of treating every mask as if it lived in one frame; the measured probe lands within 0.3 px.

Lightroom sidecar round-trip

src/xmp.rs is a hand-written sidecar reader/writer designed around conservative round trips: replace fields Autoshop owns, preserve unmodeled document content, and refuse unsupported semantics rather than silently flattening them. src/pipeline.rs connects that layer to the recipe/version store. The reader also imports Lightroom's sibling MaskBrushTable, validates its structure, and Brotli-decodes brush dab groups for the measured renderer; AI selection intent round-trips, but proprietary computed alpha is re-derived locally.

AI proposal, style fit, and generation

src/advisor/mod.rs turns the vision proposal into a bounded EditRecipe, and src/advisor/claude.rs supplies the Claude-based data-only verifier; pixels are not sent to that verifier. src/style.rs indexes prior RAW+XMP edits and retrieves similar examples, optionally with local SigLIP 2 embeddings.

src/fit.rs performs inverse rendering for match: luminance-CDF matching, exposure search, regularized engine-basis fitting and a residual tone curve are followed by closed-loop saturation and gated cast curves. src/generative.rs supports the configured gpt-image-2 used for the Part B reimagine; generation produces a lossy target, while fit/apply returns an editable deterministic full-resolution approximation. Style indexes, develop state, and segmentation alphas are cached locally.

Application and infrastructure

Rust (rustc/cargo 1.94, edition 2024) · rawler (RAW decode, 24 formats / 725 bodies) · image and qcms for raster/color I/O · rayon for row-parallel stages · clap, serde, and ureq · eframe/egui for the desktop GUI · tiny_http for serve. The embedded web UI is compiled with include_str!, so it has no runtime CDN or frontend build step.

The build workflow builds and tests the default and GUI feature sets on Ubuntu and macOS. The documented battery is 857 library (9 #[ignore]d forensic probes) / 14 CLI / 132 GUI / 2+2 contract tests, and both default and GUI Clippy runs are clean. The scripts/check_docs.py release gate re-derives pinned version, format, camera, dependency, toolchain, and battery claims from the tree.

Local ML sidecars use SCUNet for denoise, BiRefNet/U²-Net for subject masks, OneFormer for sky, SAM 2.1 for object prompts, and optional SigLIP 2 for style embeddings. Model weights are not stored in the repository.