LLR

LLR is a browser RAW editor for Sony shooters. It renders an ARW the way the camera did — Creative Look, DRO and all — and lets an LLM agent do the edit from one sentence.

Two things it does that Lightroom does not:

Three pairs, same crop of the same ARW, swapping in step:

Left: Sony's own RAW converter (Imaging Edge) ↔ LLR at the same settings. Middle: camera JPEG ↔ LLR at its defaults. Right: camera JPEG ↔ Lightroom (Camera FL profile, defaults) — watch the greens lose their colour, the petals turn, and the whole frame lift.

Measured (CIEDE2000 mean over the frame, per pixel and after a low-pass that drops texture and keeps what a flick between the two shows; method, per-region tables and a 124-frame sample):

Every ARW carries the calibration Imaging Edge renders from: the body's colour matrix and the tone curve and chroma terms of each Creative Look. LLR reads them back and runs that pipeline, so what you see first is Sony's own render — at RAW depth, with every slider still live. On top of that, camera match (a switch) applies what the body's JPEG does that Imaging Edge does not: the engine's own highlight roll-off without the LUT that costs saturation, and a fitted correction of a couple of L* and a few percent of chroma, measured against the camera's JPEGs. Lightroom's Camera Matching profiles approximate the same look from a fitted DCP, and saturated colour is where that drifts.

What that buys you, after the shot:

• All of the body's Creative Looks (ST, PT, VV, FL, IN, …) switchable per image, with the look's own tweaks on the camera's scale.
• DRO as the camera applied it, or any of Imaging Edge's levels.
• Sony's own RAW-domain noise reduction.

Anything that is not a Sony RAW renders through an Adobe DCP instead, and the engine is a per-image switch.

The Assistant is an agent with tools over the same state the sliders drive. It looks at the render, edits (Basic sliders, HSL, colour grading, curves, crop, masks, Creative Look, DRO, denoise), compares before/after, measures clipping, and iterates until the intent reads without over-correction. Every change is a history step: undo it, or take over by hand mid-way. The prompt above cost $0.40 on gpt-5.6-sol. Bring your own key; any provider pi-ai knows works.

Node 22+, pnpm 10, uv for the Python worker, and a WebGL2 browser.

pnpm install
pnpm dev        # api on :8790, web on the port Vite prints

Drop a RAW on the page. Then, optionally:

• Adobe DCP profiles for non-Sony RAWs: copy Camera/ from a DNG Converter or Lightroom install to vendor/adobe-camera-profiles/Camera/.
• The Assistant: export OPENAI_API_KEY / ANTHROPIC_API_KEY (or another provider's usual variable) before pnpm dev, then pick the model in the panel.
• Everything the server keeps — the catalog (folders, photos, edits), the imported files, decode cache — lives under ~/.cache/llr (LLR_CACHE_DIR), never in the checkout, and imports stay until you remove them from the library. Drop files or whole folders onto the folder tree to import them (a dropped folder keeps its structure); G and E switch between the library grid and the develop view. In the grid, the bar above it sorts by capture time, file modified time, camera, ISO and the rest, and groups by day, month, year, camera, lens or type. It behaves like a file manager: marquee-select from the empty space, Ctrl/Shift click and the arrow keys, F2 to rename in place, Delete to remove, and a right-click menu on a photo (or on the empty space) to open, rename, move to a folder, select, or import into the folder you are in.

• ARCHITECTURE.md — the three processes and why the browser's WebGL pass is the single source of truth for pixels.
• docs/sony-edit-internals.md and sony_repro/PIPELINE.md — how the Imaging Edge pipeline was recovered, and how faithful each stage is.
• docs/masking.md, docs/assistant-ux.md
• pnpm check runs typecheck, lint, and the web / api / worker test suites.