Skip to content
skymanbpPublic

About

AI-assisted RAW photo developer: GPT vision advisor proposes an EditRecipe, a deterministic Rust engine renders it — Lightroom-compatible XMP sidecars, 24 RAW formats + baked images, measured (not guessed) Lightroom mask geometry, local GUI/web UI, AI denoise & segmentation sidecars

Topics

Resources

Contributing

Security policy

Stars

1.1k stars

Watchers

6 watching

Forks

Repository files navigation

AutoShade icon

AutoShade

Describe the picture. Keep the RAW.

AI decides what to change — a deterministic Rust engine does it.
An editable, Lightroom-ready recipe instead of AI pixels.


Release Platforms Rust License

Download v1.18.2  ·  Website  ·  Manual  ·  Showcase  ·  In depth


Desert canyon at dusk: straight conversion, generated target, the recovered recipe rendered on the RAW, and the same recipe on the AI-denoised RAW, with a 1:1 detail row

Straight conversion → generated target → recovered recipe on the 61 MP RAW → same recipe after AI denoise. Bottom row: 1:1 crops.


Tip

The generated picture is a target, never the delivery. An image model paints the look you describe; AutoShade fits an editable develop recipe to it and renders that recipe on the full-resolution RAW. The recipe carries the look, the sensor carries the detail — and in the recipe path, the AI never touches a pixel.

Two AI paths, kept apart in the app, the command line and this page:

  • AI generation & reverse-fit — an image model paints the picture you describe from your own frame (reimagine), and the reverse fit recovers that look as an editable recipe on the full RAW (match); one five-step strength control, Light to Full + AI review, sets how far the fit may go. The generated pixels are only ever a target.
  • AI analysis & auto develop — a vision model looks at the photo and proposes the recipe directly (analyze, auto), guided by your reference library or a typed direction; when its own visual review finds that develop clearly improvable, the model states how the finished picture should measure instead, the reverse fit finds the recipe, and the review picks between the two; AI adjust sends the develop you are looking at with a request ("darken the sky a little") and gets the complete adjusted recipe back, with any part it names selected by a content mask cut locally.

Highlights

🎯 Reverse-fit any look
reimagine generates the picture from your frame; match recovers a recipe from it — or from any finished rendition of the same shot. Look error 0.161 → 0.023.

🌌 A RAW denoiser trained here
Cleans the sensor mosaic before demosaic. Keeps 98.70 % of 17,817 faint stars on a real ISO 2500 frame.

📐 Every slider is a written law
Tone, Texture, Sharpening, masks and lens geometry are closed-form laws with named constants, each pinned by its own test.

🔁 Round-trips through Lightroom
An exported .xmp is merged byte for byte into Lightroom's own; the full develop survives a Lightroom rewrite and imports back exactly.

🖼️ Your library is the style guide
Point it at finished photos you love; each develop sees the 3 closest as a reference. Embeddings run locally.

🧩 Local AI masks, no API key
Subject, sky and point-prompted objects (BiRefNet, OneFormer, SAM 2.1), plus HDR, focus and noise stacking.

Showcase

Stone viaduct: straight conversion, generated target, and the recovered recipe rendered on the RAW, with a 1:1 detail row

Reverse-fit. Prompt: “a clearer afternoon, a little more contrast, a slightly deeper blue sky.” The recipe renders the look on the 9504 × 6336 RAW; the generated frame carries neither the detail nor the resolution.


Lakeside island town: straight conversion and three AI develops driven by three different direction texts

One RAW, four looks. Straight conversion, then moody, golden and vivid — only the direction text changes.


One 1:1 window of a 61 MP star field, two ways: AutoShade's neutral develop with no denoise and with its own denoiser at the 71 % default

ISO 2500 star field, 1:1. AutoShade with no denoise, then its own denoiser at 71 %. Clean colour, natural grain, faint stars intact.

Every number and prompt → Showcase

What is new here

Six techniques you will not find in another RAW developer. Each is written up in full — equations, figures, refused training runs — in AutoShade in depth.

Technique In one line
1 Generate, then reverse-fit An editable recipe from any finished look, solved at the distribution level
2 Evidence-gated local fit A divergence score D picks full solve or bounded mode; every local move must earn its place
3 Style by retrieval SigLIP 2 search over your own finished photos, not a preset
4 Every slider is a written law Tone, Texture, Sharpening, masks and lens geometry as closed-form laws, each pinned by its own test
5 A denoiser with pre-written pass marks Trained on this project's data; two later runs failed their marks and were refused
6 The camera's own look A per-photo base curve read from the embedded preview, like with like

How it works

AutoShade architecture: three front ends over one Rust library with the reference library, reverse-fit, local producers and the local-field analyzer; local Python sidecars for embeddings, descriptions, correspondence, segmentation and two denoisers, plus the depth and face sidecars (eight in all); opt-in external AI services
  • One engine, three front ends — desktop app, CLI and loopback web UI share one Rust library and one per-user develop store.
  • Local first — eight Python sidecars (masks, two denoisers, correspondence, descriptions, embeddings, depth for the lens blur, faces) run on your machine; pixels leave it only for an AI operation you ask for.
  • Non-destructive — your photo library stays read-only; generated pixels always land on their own labelled card.

Diagram generated from autoshade.architecture.json · zoom it online · full design in ARCHITECTURE.md

Measured numbers

What Measured
Automated test battery 2003 library / 28 CLI / 323 GUI / 2+2 contract tests
RAW coverage 24 extensions · 725 camera bodies · nine-camera zoo 9/9
Reverse-fit, stone viaduct look error 0.161 → 0.023 at confidence 0.63
Reverse-fit, desert canyon look error 0.129 → 0.048 · sky ΔE 3.8
Faint stars kept through the denoiser 98.70 % of 17,817 stars on an ISO 2500 frame

The full table — 20+ rows with conditions and sources — is in AutoShade in depth; scripts/check_docs.py re-derives the pinned claims.

Install and quickstart

Platform Get it
Windows AutoShade-Setup-1.18.2.exe — per-user, no admin. Or the portable autoshade-1.18.2-windows-x64.zip
macOS AutoShade-1.18.2-macos-universal.zip (Apple silicon + Intel), or -macos-cli.zip
Linux AutoShade-1.18.2-linux-x64.zip — command line only

All on the v1.18.2 release page. Python 3 is needed only for the local AI sidecars; model weights download on first use.

Note

macOS: the app is ad-hoc signed, not notarised — on first launch use System Settings → Privacy & Security → Open Anyway (once per machine).

Desktop app — it starts on a mode page (Landscape, Portraits, Normal; see Choosing a mode); then open a photo (Ctrl+O), move a slider, export with Ctrl+Shift+E. The original is never touched. Normal is a short one-click panel for phone and jpg photos (Normal mode).

CLI

# manual develop
autoshade apply photo.ARW recipe.json -o developed.tif

# AI develop, end to end
autoshade auto photo.ARW --guidance "natural color; protect highlights" -o developed.tif

# describe a look, then recover it as an editable recipe on the RAW
autoshade reimagine photo.ARW --prompt "a clearer afternoon, deeper blue sky"
autoshade match photo.ARW out/photo.reimagine.png --render

Upgrades, silent installs and every command: User manual.

Release assets and checksums

Built by GitHub Actions from the tag; checksums.txt carries every SHA-256. The RAW denoiser weights are uploaded by hand to the release the sidecar pins — releases/download/v1.6.0/autoshade-raw-denoise-v2.pth, mirrored byte-exact on Hugging Face — and refused unless size and SHA-256 match the pin in python/denoise_raw.py.

File Size SHA-256
autoshade.exe (CLI) 26,245,120 bytes 7470e35a7e82eac61b120eded894ed33703067bf9efb7e7ff9ebc79b29c3778d
autoshade-gui.exe (desktop app) 33,745,920 bytes 1ff4c3eb761bc16ddcdd2ab5dc5ecf368c7ff7c890db1598b8cc0d2e297e05ce
AutoShade-Setup-1.18.2.exe (installer) 16,928,266 bytes c204ca401a7050773f169fc715e7596f011f3f13fbbb7d53ae41f8a5c212d7dc
autoshade-1.18.2-windows-x64.zip (portable archive) 23,061,710 bytes 4c1b7b0d1e451531741129646c9dd49685ef54b679ff3ef2a0e3505392a32d97
AutoShade-1.18.2-macos-universal.zip (macOS app bundle) 46,095,029 bytes da51b67e241e0014505214211ee0e62717dbda2f3fc81562c8fc2942f054de40
AutoShade-1.18.2-linux-x64.zip (Linux command line only) 11,041,955 bytes 289bc538b33ded8cefd51ef59e1a45ba77ea08e6760be158f0bf29982379e96c
AutoShade-1.18.2-macos-cli.zip (macOS command line only) 20,055,706 bytes 9690f60b6b7a3a39c3f6a66540c776f8a0146722ef0764015a73085d8ef50c1b
autoshade-raw-denoise-v2.pth (RAW denoiser weights, fetched on demand from the v1.6.0 release) 130,590,559 bytes ffafa40a53f52092149db2fcf03636117ad6855e1068142d4f6b03b634e9f9c4
Build from source

Rust edition 2024, rustc/cargo 1.94.

cargo build --release                                   # CLI
cargo build --release --features gui --bin autoshade-gui  # desktop app

Local AI tools need Python packages (weights download on first use):

Tool Packages
Subject mask (BiRefNet) torchvision timm einops
Subject fallback (U²-Net) rembg
Sky / objects (OneFormer, SAM 2.1) transformers torch
AI denoise torch opencv-python numpy einops requests

Supported formats

Canon CR2 develop
.cr2 · Canon EOS 40D
Canon CR3 develop
.cr3 · Canon EOS R6
Nikon NEF develop
.nef · Nikon D700
Sony ARW develop
.arw · Sony α7 III
Olympus ORF develop
.orf · Olympus E-M5
Panasonic RW2 develop
.rw2 · Panasonic DMC-GX85
Pentax PEF develop
.pef · Pentax K-5
Ricoh DNG develop
.dng · Ricoh GR II
Fujifilm RAF X-Trans develop
.raf · Fujifilm X-S10 — X-Trans, approximate

Nine real CC0 files, fully decoded and neutral-rendered — not embedded previews.

Camera RAW — 24 extensions

arw, dng, raw, raf, nef, cr2, cr3, orf, rw2, pef, srw, 3fr,
fff, iiq, mef, mos, erf, kdc, dcr, dcs, crw, nrw, mrw, ari

Decoded by rawler 0.7.2, which carries 725 camera models. No embedded preview: 12 of the 24 formats store none. They are orf, srw, nrw, mef, mos, kdc, dcr, dcs, erf, iiq, crw, and ari; AutoShade renders its own neutral version instead.

Baked rasters — 8 extensions: jpg jpeg png tif tiff bmp webp gif, with ICC conversion.

Tech stack

Rust (rustc/cargo 1.94, edition 2024) · rawler · image · qcms · rayon · egui/eframe · tiny_http — Python sidecars on PyTorch.

RAW decode, 24 formats, 725 bodies · deterministic f32 develop engine · hand-rolled XMP layer (no XML crate) for byte-exact Lightroom merges. The current battery is 2003 library (1981 pass + 22 #[ignore]d forensic probes) / 28 CLI / 323 GUI / 2+2 contract tests. Per-subsystem notes: In depth → Implementation · equations and provenance: TECH_STACK.md.

Status and limitations

  • Release gates at v1.18.2 cover the CLI, desktop GUI, sidecar contracts, format fixtures and the deterministic renderer.
  • Not in scope: Adobe's rendering (AutoShade is its own develop engine; a sidecar carries the same settings to Lightroom, which renders them its own way), an exact X-Trans demosaic, a notarised macOS build, a Linux desktop app.
  • Honest markers: AI masks are local re-creations, not Adobe's pixels; the macOS app is CI-verified but has no reported interactive users yet.
  • Release ledger and rulings: ROADMAP.md.

License

MIT — see LICENSE. Showcase photographs © 2026 skymanbp, all rights reserved (not MIT).

Models, fonts and samples

Model weights download separately and remain their authors' property; every pinned model also has a byte-exact mirror on Hugging Face (Azng0/autoshade-mirror-*), checksum-verified either way — table in python/_mirror.py.

Model Purpose License
SCUNet AI denoise (baked sources) Apache-2.0
DRUNet-colour (DPIR), fine-tuned here as autoshade-raw-denoise-v2.pth AI denoise (RAW mosaic) MIT; fine-tuning pairs from RawNIND, CC BY-SA 4.0
BiRefNet Subject segmentation MIT
U²-Net Subject fallback Apache-2.0
OneFormer ADE20K Sky segmentation MIT
SAM 2.1 Point-prompted object masks Apache-2.0
MoGe-2 ViT-L (code in python/_moge/) AI lens blur depth map MIT (DINOv2 backbone code Apache-2.0)
MediaPipe Face Landmarker (BlazeFace + Face Mesh V2) Portraits mode face and eye outlines Apache-2.0
SigLIP 2 Style embeddings Apache-2.0
Qwen3-VL-2B-Instruct Local look descriptions Apache-2.0
Stable Diffusion 2.1 DIFT correspondence field CreativeML Open RAIL++-M

Fonts: subset Noto faces (SIL OFL) and a Lucide icon face (ISC) — texts in assets/fonts/. RAW samples: the nine format-grid files are from raw.pixls.us under CC0 1.0, SHA-256-verified.

Thanks to the rawler, image, qcms, rayon, clap, serde, ureq, egui/eframe, tiny_http and local-model communities.


Made for photographers who want to know what the AI changed — in numbers.

About

AI-assisted RAW photo developer: GPT vision advisor proposes an EditRecipe, a deterministic Rust engine renders it — Lightroom-compatible XMP sidecars, 24 RAW formats + baked images, measured (not guessed) Lightroom mask geometry, local GUI/web UI, AI denoise & segmentation sidecars

Topics

Resources

Contributing

Security policy

Stars

1.1k stars

Watchers

6 watching

Forks

Releases

Packages

Used by

Contributors

Languages