Restore Anything Pipeline:
Segment Anything Meets Image Restoration

Jiaxi Jiang · Christian Holz

Click anything in a photo. Give it its own restoration model. Dial it in with a slider.

_{Select by name, as a subject or with a click, then give each region its own model: restore just the face, blur only what is behind the person, remove a person with their shadow. (Model waits are sped up in all demos.)}

Most restoration tools treat a photo as one thing. Real photos are not: the face needs a face model, the sky is noisy, the sign is a JPEG mess, and the tourist in the back should go. The Restore Anything Pipeline (RAP) lets you select each part with SAM 3 and give it the right model, at the right strength, while you watch. It runs on your own GPU as a local web app.

This is the code for our RAP technical report (2023), rebuilt with current models and a full editing interface.

Highlights

• 🎯 Select anything. Click it, drag a box, paint it, or type its name: face, sky, text.
• 🧰 One model per region. Restore, denoise, deblur, fix JPEG, restore faces, colorize, brighten, erase, remove with shadows, or generate from a prompt.
• 🎚️ Instant sliders. Each model's main control is rendered in advance: drag and see it, or click a preview in the filmstrip.
• 🪄 One-click help. Auto enhance restores the background and every face; the app looks at your photo and suggests what to fix.
• 🖼️ Editor basics built in. Light and color adjustments per region, crop and straighten, undo/redo for everything, export to PNG/JPEG/WebP with 2×/4× AI upscaling and EXIF kept.
• 💾 Easy on memory. Models load the first time you use them and move to CPU memory when idle.

What's new since the 2023 report

The technical report paired Segment Anything with a controllable restoration model, so each object could get its own restoration, with several results to choose from. This release keeps that idea and builds it out:

• Selection: SAM 3, with text prompts (face, sky), boxes, a brush and soft-edged subject mattes on top of clicks.
• Models: many models instead of one, covering restoration, faces, colorization, low light, object removal and generative fill, each with continuous controls; adding another takes one small class (see Adding a restorer).
• Choosing a result: every control is pre-rendered, and a filmstrip previews it, so trying settings is instant.
• Editing: light and color adjustments per region, crop and straighten, undo/redo, Apply to chain steps, and export with 2×/4× upscaling.
• Interface: a local web app; models load when first needed.

Demos

Click, choose, slide

Click the bird, pick Enhance photo, and drag Detail. Hold Space to see the original.

Remove an object together with its shadow

Click the bench and choose Remove object + shadow (ObjectClear). Its shadow goes too.

Find by name, then replace it

Type sky, choose Replace with…, describe a sunset (FLUX.2 [klein]), then clean the noisy building with Remove noise (SCUNet).

Results

Colorize, apply, then Auto enhance the faces	Heavy JPEG cleaned up; FBCNN finds the quality by itself
A bench and its shadow removed	A new background from a prompt
Portrait look: background blurred by estimated depth	Night photo brightened

Quick start

1. Install (Linux with an NVIDIA GPU, conda):

git clone https://github.com/eth-siplab/RAP.git
cd RAP
bash scripts/setup.sh          # conda env "rap", third-party code at pinned commits, weights
hf auth login                  # SAM 3 checkpoints are gated on Hugging Face

Request access at https://huggingface.co/facebook/sam3 first. Some models are downloaded from Hugging Face the first time you use them:

• SD-2.1-base (PiSA-SR's base), from the sd2-community mirror; the Stability AI repository was taken down
• ObjectClear, FLUX.2 [klein], DDColor, HVI-CIDNet, BiRefNet (remote code pinned to a revision) and Depth Anything V2

2. Run:

conda activate rap
python -m rap --gpu 0 --port 7860

3. Open http://localhost:7860. On a remote server, forward the port first:

ssh -N -L 7860:localhost:7860 <user>@<server>

Pick one of the example photos on the start page, or drop in your own.

Hardware. An NVIDIA GPU with 8 GB or more runs every feature.

GPU memory	What changes
8–12 GB	Models take turns on the GPU (a few seconds when switching). Generative fill streams its weights from CPU memory: four variants in about 10–20 s.
16–24 GB	Most models stay loaded; switching is rarely needed.
28 GB or more	Everything stays on the GPU, including generative fill.

The app sizes itself to your GPU: by default about 60% of GPU memory holds model weights (--gpu-budget), and models that don't fit wait in CPU memory, capped at half of your RAM (32 GB of RAM recommended). If memory still runs out, it moves other models out, works in smaller batches and retries instead of failing. --max-gpu-memory limits the total on a shared GPU. The table was measured on a server GPU limited to 8, 12, 16 and 24 GB with that option, where the full test suite passes; a real consumer card has the same memory limits but may be slower. 12-megapixel photos work.

Your work survives a restart. Sessions are saved to ~/.cache/rap/sessions (--session-dir; empty keeps them in memory only), and reloading the page brings yours back. --cpu-threads (default 8) caps OpenCV/torch CPU threads, which matters on busy shared machines.

Tutorial

1. Tools. Select (V): click to add, right-click to remove. Box (B), brush (P), crop (C), pan (H). The line underneath says how the current tool works.
2. Find by name. Type a noun such as face, sky or text to select every match at once.
3. What to do with this region. Opens the menu of models and effects; the one that suits the region best is marked Recommended.
4. Slider and filmstrip. The filmstrip shows the main slider at several values; click one to jump there. Releasing the slider renders the exact value.
5. Compare (or hold Space) shows the original; Split gives a before/after divider you can drag.
6. Auto enhance restores the background with PiSA-SR and every face with CodeFormer.
7. Undo / redo work for everything, including Apply and crop.
8. Export as PNG, JPEG or WebP: set the quality, resize, or upscale 2×/4× with PiSA-SR.

Your first edit

1. Open a photo: click an example, drop a file anywhere, or paste from the clipboard.
2. Select something with a click. Add parts with more clicks and remove parts with right-clicks. Press Esc (or click beside the image) when the selection is right.
3. Choose what to do with it from its card on the right.
4. Tune the slider. Each region also has Adjust (exposure, contrast, color, clarity, …) and Strength, which blends with the original.
5. Everything else is the Background layer, which can have its own model too.
6. Export. Your work also survives a page reload: the address holds your session.

Tips

• Apply bakes the current result into the photo so you can chain steps: colorize → Apply → restore faces.
• Select subject mattes the main subject with soft edges (hair); Remove background makes the rest transparent.
• Removals and replacements change the photo itself: the background and the other regions are processed on the edited photo, so removing someone and then brightening everything works in one go. Hold Space to see the original.
• Click the logo to go back to the start page.

Keyboard and mouse

Key	Action
V	Select: click adds, right-click removes
B	Box select
P	Brush: paint adds, right-drag erases; [ ] change the size
C	Crop, rotate, straighten, flip (Enter applies)
H / middle-drag	Pan; the wheel zooms
F · 1	Fit · actual pixels
hold Space	Compare with the original
Ctrl+Z / Ctrl+Shift+Z	Undo / redo
Enter / Esc	Finish the current selection
Delete	Delete the selected region
double-click a slider	Reset it
?	All shortcuts, in the app

Models

Every model has continuous controls, and the main control is rendered in advance, so dragging it is instant.

Model	For	Controls	License
PiSA-SR (CVPR 2025)	real photos with mixed degradations; 2×/4× export	Detail, Fidelity	Apache-2.0
SCUNet (MIR 2023)	real noise: phones, high ISO, grainy scans	Texture (PSNR ↔ GAN model)	Apache-2.0
DRUNet (TPAMI 2022)	noise at a strength you choose	Noise level σ with a blind estimate	MIT
NAFNet (ECCV 2022)	camera shake, motion blur	Passes	MIT
FBCNN (ICCV 2021)	JPEG artifacts	Quality factor with a blind estimate	Apache-2.0
CodeFormer (NeurIPS 2022)	faces	Fidelity	S-Lab (non-commercial)
LaMa (WACV 2022)	erasing scratches, text, small objects	Mask expansion	Apache-2.0
ObjectClear (CVPR 2026)	removing objects with their shadows and reflections	Variant, Removal strength	S-Lab (non-commercial)
FLUX.2 [klein] 4B	generative fill from a prompt; on the background it replaces the background	Variant, Repaint strength	Apache-2.0
DDColor (ICCV 2023)	colorizing black-and-white photos	Color intensity	Apache-2.0
HVI-CIDNet (CVPR 2025)	low light	Brightness, Saturation	MIT

Also used: SAM 3 for selection, BiRefNet for Select subject, and Depth Anything V2 Small (Apache-2.0) for lens blur.

Effects per region: transparent background, solid color, lens blur (focused on the selected subject), and blur or pixelate for privacy, scaled to each selected part such as each face.

Adjust per region: exposure, contrast, highlights, shadows, whites, blacks, temperature, tint, vibrance, saturation, clarity, dehaze, sharpen, vignette, plus Auto tone and Auto color. These also work without a model, for example to brighten one backlit face.

How it stays fast

• Every restorer splits its work into a control-independent prepare and a cheap render. A sweep of the main control renders in one batch, and releasing the slider renders the exact value.
• Where the math allows, sweeps are almost free:
  • PiSA-SR's output is linear in its two strengths.
  • FBCNN's condition only enters its decoder.
  • SCUNet's Texture blends two outputs computed once.
  • DDColor and HVI-CIDNet controls act after the network.
• Tonal adjustments run in the browser. Export applies the same pipeline on the server (rap/adjust.py), matching the preview to within a few 8-bit levels.

For developers

Layout

rap/
  server.py            FastAPI app: sessions, segmentation, sweeps, history, apply, crop, export
  session.py           regions (binary or soft masks), brush strokes, undo/redo, compositing
  store.py             sessions saved to disk, so a restart keeps everyone's work
  models.py            lazy loading, GPU budget, idle offload
  segment.py           SAM 3 wrapper (points / box / text)
  subject.py           BiRefNet subject matting, Depth Anything V2 depth
  adjust.py            tonal adjustments (reference implementation for export)
  geometry.py          crop / rotate / straighten / flip
  restorers/
    base.py            Restorer interface
    pisasr.py          PiSA-SR on stock diffusers (LoRAs merged by hand) + upscaling
    kair.py            SCUNet and DRUNet denoisers (architectures vendored from KAIR in kair_arch.py)
    nafnet.py          NAFNet motion deblur (architecture vendored in nafnet_arch.py)
    fbcnn.py           FBCNN JPEG artifact removal (tiled for large crops)
    codeformer.py      CodeFormer with batched restore and paste-back
    lama.py            LaMa erase
    objectclear.py     ObjectClear object + effect removal
    genfill.py         FLUX.2 [klein] generative fill
    ddcolor.py         DDColor colorization
    cidnet.py          HVI-CIDNet low light
    effects.py         transparent, solid color, blur, pixelate, lens blur
  static/              single-page UI, no build step
scripts/               setup, weights, example images, README demos
examples/              demo images (see examples/examples.json for sources and licenses)
tests/                 end-to-end checks (API + headless browser)
docs/                  README images and demo GIFs

Adding a restorer

1. Subclass Restorer.
2. Declare its params. The first param is the one that gets swept.
3. Implement prepare(img, mask) and render (or render_many to batch).
4. Register it in rap/restorers/__init__.py.

Flags:

• needs_mask, context, blend: for restorers that edit inside a region themselves.
• uses_foreground, bg_only, allow_bg: for background layers.
• prompt_hint: for text-driven restorers.

Tests

tests/unit/ needs no GPU and runs on every push (GitHub Actions): adjustments, crop/rotate, undo/redo and saving/restoring sessions.

python -m pytest tests/unit

tests/run_all.sh drives a running server through its API and a headless browser (Playwright): segmentation, every restorer, adjustments, crop, undo/redo, export. It reports which scripts failed; each script also prints what it measured and saves screenshots to tests/out/.

pip install playwright && playwright install chromium
python -m rap --gpu 0                    # in one terminal
PYTHON=python tests/run_all.sh           # in another; RAP_URL=... to test another server

Regenerating the demos

With the server running: python scripts/record_demos.py records the clips (a drawn cursor clicks through the app), python scripts/make_gifs.py turns them into GIFs with the model waits sped up (needs ffmpeg), python scripts/make_ui_tour.py makes the annotated screenshot, and python scripts/make_showcase.py the JPEG and portrait comparisons.

Citation

@article{jiang2023restore,
  title={Restore anything pipeline: Segment anything meets image restoration},
  author={Jiang, Jiaxi and Holz, Christian},
  journal={arXiv preprint arXiv:2305.13093},
  year={2023}
}

Licenses

The code in this repository is released under the Apache License 2.0 (see LICENSE). Code taken from other projects keeps its original license, noted at the top of each file; the PiSA-SR wrapper follows the upstream inference code (Apache-2.0). The crop, undo and redo icons are adapted from Lucide (ISC). Each model keeps its own license; see the table above. CodeFormer and ObjectClear are non-commercial. The example photos are public domain or CC0 from Wikimedia Commons (credits in examples/examples.json), except the RAP demo images from 2023.