scumble

# Scumble

**A free, open-source desktop editor for AI inpainting.** Select part of a picture, describe what should be there,
and the result comes back as an editable layer, colour-matched to its surroundings. Scumble renders with FLUX 3 Image,
FLUX.2, GPT Image, Nano Banana, Seedream, Qwen Image Edit and more through your own API key, or locally on your own
ComfyUI, and keeps the work in layers you can export to PSD. Manual, videos and the dev blog:
[denrakeiw.com/scumble](https://www.denrakeiw.com/scumble).

> **Work in progress.** Scumble is in an early state (0.1.x). Not every feature has been
> tested end to end yet. Expect rough edges, keep backups of your images, and please report
> what breaks in the [issues](https://github.com/DenRakEiw/scumble/issues).

How it works: open an image, select an area (brush, shape, magic
wand, object hover or a text description), write a prompt, generate. The result lands as a
layer over the selection and can be blended in with colour match, erased in parts,
regenerated, stacked with filter and text layers, and exported with all layers to PSD or
OpenRaster. Or open the assistant and say what you want: it drives the same editor through
the same commands an external agent gets, one card per step, and asks before anything costs
money.

[![Scumble, explained by someone who did not ask: the video, 4:40](https://www.denrakeiw.com/scumble/videos/scumble-explained.jpg)](https://www.denrakeiw.com/scumble/videos)

**[Watch: Scumble, explained by someone who did not ask](https://www.denrakeiw.com/scumble/videos)** (4:40): two
voices, one sceptic, the whole app recorded in the app.

![The editor with a picture open: the tool column, the canvas, the layer stack and the selection panel](docs/images/editor.jpg)

Rendering happens on your own [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
(local or remote, for example on RunPod) or through API providers: Google (Nano Banana
2 / 2 Lite / Pro), OpenAI (GPT Image 2.5 Flare / Sunburst, 2), Black Forest Labs
(FLUX 3 Image, FLUX.2 max / pro / flex / klein, FLUX.1 Fill), ByteDance Seedream 5 and 4.5, Qwen Image Edit and Qwen Image 2.1,
Ideogram 4.5 (edits with the selection as its mask), and on Magnific its own Mystic, Ideogram mask inpainting and Image Expand outpainting (FLUX Pro, Ideogram, Seedream 4.5),
each through the model's own API where Scumble has one (for Seedream that is ByteDance's BytePlus ModelArk) or
through ToAPIs, fal.ai, Replicate, WaveSpeedAI, Comfy Cloud, Comfy Router, OpenRouter, Oxen.ai and Magnific. Object masks and background removal
run inside the app through ONNX Runtime (SAM2, BiRefNet, RMBG). The editor is the same
code as the ComfyUI node [Inpaint Canvas](https://github.com/DenRakEiw/ComfyUI-InpaintCanvas);
Scumble is the standalone window around it, plus recipes, plugins, an MCP server and the assistant.

Windows first (from the [Microsoft Store](https://apps.microsoft.com/detail/9NDBTNNMXF2R), the installer below or a portable zip), a Linux build (AppImage, .deb) that has not been tried on Linux yet, macOS is planned. Free software, GPL-3.0.
What has been verified so far: local rendering through ComfyUI, the in-app helper models,
the film pack, the command core, the MCP server, the tile engine on large documents and
auto-update, and among the API providers FLUX 3 Image on Black Forest Labs (with boxes in the prompt), OpenRouter and Comfy Router,
GPT Image 2.5 through OpenRouter, Ideogram 4 on fal.ai (with boxes), and the Realism Pass on an RTX 5090; the
other API providers and the assistant's model calls are untested against the live services.

## Features

### Select, prompt, generate

Paint a selection (or draw a rectangle, an ellipse, a lasso, use the magic wand, hover an
object, or type "the handbag" for a text selection), write the prompt in the Generate tab,
press Generate. The recipe decides where it runs: your ComfyUI, or a provider with a key.

![A rectangle selected around the handbag, the Generate tab with the prompt, the recipe settings and the seed](docs/images/selection.jpg)

- Selection by brush, rectangle, ellipse, lasso, magic wand, object hover (SAM2 in-app)
  or by text (SAM3 on the ComfyUI side); grow, shrink, feather, invert, from layer, saved
  selections.
- Recipes instead of node graphs: pick a model ("FLUX.2 [max]", "Nano Banana 2") and the
  provider it runs on (ToAPIs, its own API such as BytePlus ModelArk for Seedream, fal.ai, Replicate, WaveSpeedAI,
  Comfy Cloud, Comfy Router, OpenRouter, Oxen.ai, Magnific); import your own ComfyUI
  workflow as a recipe if it holds an Inpaint Canvas node, or a copy of a shipped model recipe with
  a variant of your own. Thirteen of Comfy's templates come as Comfy Cloud recipes: image edits (Boogu, Flux.2 Klein
  9B, Flux.2 dev, Mage Flow, Qwen Image 2.1 and 2509) and new images for Generate new (Anima, Flux.2 Klein 9B,
  Ideogram 4, Krea 2 Turbo, Mage Flow, Qwen Image 2.1, Z-Image Turbo); a workflow from Comfy Cloud becomes one as it is.
- API runs go out at the size the provider really takes (*Highres fix* picks the tier), with
  reference layers, and transparent results from the OpenAI image models land as cut-outs.
- Start from nothing: *Generate new* makes the base image from the prompt alone, locally
  or through a provider, and you edit it from there.
- Boxes in the prompt for FLUX 3 Image and Ideogram 4 (on fal): draw boxes on the picture with the Boxes tool (X) and say what each one
  does (add something new, keep, move or remove an element, place a reference layer, render words); the run tells
  the model where each change goes. With no boxes drawn, the selection goes as one box. One switch under the prompt
  turns them on and off, and the boxes are saved with the document. Ideogram 4 gets them as its structured caption
  (new things, words and kept elements; it has no move, remove or reference boxes).
- *Realism Pass (Windows only, RTX only)*, a refiner for generated pictures: from the Upscale dialog it sends the
  whole visible picture at 1× through DLSS 5 Neural Rendering on your own ComfyUI and adds the answer as a new layer
  above it, so generated skin, hair and fabric look less waxy. It needs a ComfyUI on Windows with an RTX 30, 40 or 50
  card and the community node pack ComfyUI-DLSS5-Enhancer with its runtime, which you install there; the
  [manual](docs/MANUAL.md)'s chapter of the same name has the steps. There is no Comfy Cloud version.
- Prompt upsampling through a stored API key, an OpenRouter, Oxen.ai or ToAPIs key, or a local Ollama /
  LM Studio, with your own prompt-writing rules as Markdown templates.

### Layers, filters, text

Every result is a layer. Match its colours to what is below it, mask it, erase parts, set a
blend mode, put filter layers and text on top. Nothing is baked in until you flatten.

![The layer stack: a vignette and a Kodak Portra 400 film look as filter layers, a text layer and the base](docs/images/layers.jpg)

- A full layer stack: paint, image, text and filter layers, masks, blend modes, opacity,
  retouch tools (clone, heal, smudge), transform, crop and extend, copy and paste of whole
  layers between tabs, SVG files as layers.
- Colour match per layer: a result that came back a shade off is matched to its
  surroundings (or to what lies below it) with one slider in the layer row, non-destructively;
  the slider can also stay at 40 % when the model's own tone is worth keeping.

![Colour match: the same result layer at Match 0 % with a visible cold rectangle, and at 100 % blended into the ground; below, the layer row with the Match slider](docs/images/colour-match.jpg)

- Filter layers on the GPU (WebGL2): grain with film presets, curves, levels, colour
  balance, HSL, LUT (.cube), vignette, normalise, sharpen, blur and more; a film pack plugin with
  film looks, halation, glow, bleach bypass, cross processing, split toning, light leaks,
  frames and control points.
- A shape tool (rectangle, ellipse, polygon, Bezier, freehand path, fill and outline),
  custom brushes from Photoshop `.abr` files, and 3D objects (`.glb`) placed into the picture
  with their own light.
- Export PNG, JPEG, WebP, PSD and ORA with layers, masks and selections, at a percentage or
  in a frame of a given size; an *AI label* panel writes the EU AI label into the file.

### The assistant

A chat column next to the canvas. It runs on your own API key, on Anthropic, OpenAI, Google
or any OpenAI-compatible endpoint (OpenRouter, DeepSeek, Moonshot / Kimi, Z.ai / GLM,
ToAPIs, WaveSpeed, Oxen.ai, or a local server), and drives the editor through the same 90+ commands
an external MCP client gets. Every call is a card you can open; everything that costs money,
queues on your ComfyUI, clears the undo stack or touches a layer that is not its own asks
first.

![The assistant panel: the request, the model's reply, two finished tool cards and a generate call waiting for Allow or Don't](docs/images/assistant.jpg)

- Ctrl+Z takes back each of its steps, and *Undo this turn* puts every document the turn
  touched back to what it was before, even when the turn was longer than the undo stack.
- Chats are saved as they go, with their screenshots, and can be reopened; *Settings >
  Assistant* deletes everything the assistant ever stored, your keys excepted.
- Your own model ids: *Settings > Language models* takes any model of any listed provider
  (an OpenRouter id, a model released after this version) for the assistant, for prompt
  upsampling, or both, on the key that provider already has.
- The whole feature, what leaves your machine per provider and what it costs:
  [docs/ASSISTANT.md](docs/ASSISTANT.md).

<img src="docs/images/settings-language-models.jpg" alt="Settings: the Language models section with provider, model id, name and the three checkboxes, then the prompt templates and the recipes" width="560">

### Large pictures, plugins, agents

- A tile engine keeps the picture, every layer and every mask in tiles; the pixel ker