[bmachek]

AI Edit gets an AI retoucher: a vision model edits, Lightroom applies and renders the edit, the model looks at the render next to the original with measurements of its tones, and corrects — round by round, the way a person works at the Develop panel. The style engine stays as My style only (offline, no model), and can be the retoucher's starting point.

Why

Both existing paths stop short of good edits. The single-shot LLM edit (/v1/edit/recipe, unreachable from the plugin since #274) guessed absolute slider values for a file and profile it had never seen rendered — what vision models are worst at. The style engine never looks at the photo at all; only the guardrails do. Vision models are good at the step in between: judging a picture and saying what is wrong with it. The loop gives them that step.

What the branch contains

1. Merge of feature/ai-edit-xmp (a3726ea): lrg-develop (key registry, typed develop model, Lua/XMP readers and writers, mask builders), the style engine's typed white balance and the develop-settings experiments, onto current main. Conflicts resolved in Info.lua, Util.lua, style_edit.rs (the branch's warning list with main's guardrail result), Cargo.toml/lock, the golden-tests workflow. The retoucher is the first production caller of lrg-develop's Lua writer, applying through applyDevelopSettings + updateAISettings as experiment E2 settled.

2. The retoucher (1f46a97, and 930b7f2 from the live test):

Piece	What
lrg-api edit_loop.rs, routes/edit_session.rs	POST /v1/edit/sessions, POST /v1/edit/sessions/{id}/render, DELETE /v1/edit/sessions/{id}. The backend holds the conversation (in memory, 30 min idle expiry); the plugin drives it. Runs on lrg_providers::chat::ChatModel, so every provider the agent chat runs on works, local engines included. Every edit passes the existing guardrails
lrg-providers edit_loop.rs	System prompt, seven looks, the apply_edit/finish tools with a schema generated from the develop registry (Lightroom's own ranges, the photo's white-balance family), the opening and render texts
lrg-develop build/recipe.rs	Recipe → DevelopSettings: recipe aliases, colour-grading midtones/luminances/blending, black-and-white mix, white balance per family (a value on the other scale is dropped, never clamped), AI masks for subject, sky, background, ten people parts and seven landscape classes with sync ids derived from the target. And starting_values, the way back
lrg-imaging tone.rs	Luminance percentiles, histogram bands, clipping, the colour of the near-neutral tones, midtone saturation — what the model is told to trust over its eye
lrg-analysis style_engine	white_balance_into_global, so the photographer's style can be round 0
EditLoop.lua	Applies each round from the photo's starting point (keys an earlier round set go back; masks merge by CorrectionSyncID, the photo's own masks passed back raw), computes AI masks, renders at 1536 px, reports a mask that found nothing, reverts on discard
TaskAiEditPhotos.lua	Two editors. Retoucher: model (defaults to Analyze & Index's), look, own instructions, edits per photo (3), start from my style (≥ 5 examples), AI masks (LrC 15.3+). Review is Keep / Discard / Keep all on the finished edit. My style only: the previous flow, moved into runStyleEdits unchanged

Every edit is the complete edit measured from the starting point, so rounds are idempotent and the model can go back to an earlier, better round (best_round).

Errors surface (reviewed against Errors must surface): what did not reach the photo (clamped values, unknown mask targets, masks beyond six, keys the process version lacks, the style's white balance from the other file kind), a kept round's mask that found nothing, a failed record of the edit in the catalog or the index, edits the model sent that Lightroom could not use (named as the cause instead of "no tool calling") — all reach the review dialog and the run summary. persistEditRecipe now returns whether it recorded; both editors report a failure.

Test plan

• cargo fmt, cargo clippy --workspace --all-targets (zero warnings), cargo test --workspace (1534 passed). New: 6 end-to-end loop tests on a scripted model (edit → render → refine → finish with best_round; style as round 0 with both images; nudge then refuse a model without tool calls; edit budget; JPEG offset white balance; unusable edits named as the cause), recipe builder (incl. writing an apply table with a mask), tone measurement, tool schema
• busted (10 new specs for composing rounds, merging masks by sync id, revert, mask states; the one failure, aesthetic_refine_spec.lua, predates this branch — Lua 5.5 prints 1.0)
• luacheck, stylua, scripts/check-docs.py, pre-commit hooks
• Live against the dev binary over HTTP, with a scripted OpenAI-compatible server through the real openai_compatible adapter: start, two renders with Lightroom notes, a guardrail cap, back to round 2, the recipe stored on the photo's record, 404 after the end, DELETE
• Live in Lightroom Classic 16.0, test catalog, plugin from the checkout, driven through the agent bridge (EditLoop.run via eval_lua), seven CR3 photos:
  • Gemini 3.1 Pro: landscape with sky, backlit forest (twice), person in a forest, B&W portrait (look Black & white); 2–3 rounds, 25–50 s and 20–38k input tokens per photo. Dark backlit frames come out well exposed and natural; the model ends early when satisfied
  • Hybrid with 13 of the photographer's edits saved as training examples: style as round 0, then the model's corrections — which led to 930b7f2 (edits now merge onto the style instead of dropping its values)
  • Local Gemma 4 E4B (MLX): tool calls work through schema decoding, 3 rounds in ~50 s, but the edit stays timid and it dropped values in later rounds (renders now read the edit back to the model)
  • AI masks computed for subject, sky, background, face skin, iris and pupil and the landscape class vegetation; the photo's own masks kept and computed; same-sync-id corrections replaced in place; discard restores every written key and, with no own masks, an empty list removes the loop's — recorded in Dev-AI-Edit-XMP-Findings.md
  • Failure path: Gemini without credit (HTTP 402) reported in plain words, photo untouched
• A non-raw file (the test catalog had none), the dialog itself (Keep / Discard / Keep all, catalog_photo), virtual copies

Known behaviour

• The backend stores the kept recipe on the photo's index record when the session ends, before the user's Keep/Discard — as /v1/edit/style stores the recipe it generates.
• With a cloud model, renders of every edited photo go to that provider (stated in the help page).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HaMMvswyTLENR7Gdygjt41