Scumble
A free, open-source desktop editor built around AI models. Inpaint, edit, upscale and generate: select, say what should be there, and the result comes back as a layer of its own, colour-matched to its surroundings.
One picture from the first generation to the finished frame, 17 layers, without leaving the editor (sped up, the waits cut out).
Microsoft StoreWindows, signed by Microsoft, updates itself. Or
winget install 9NDBTNNMXF2R.
Download from GitHubWindows installer, portable zip, Linux AppImage and .deb, and the Claude Desktop extension.
Source on GitHubGPL-3.0. A star helps the project get found.
What it does
- Select by rectangle, brush, lasso, magic wand, object hover, a text description, or by depth, brightness and colour; type a prompt; the result lands as a new layer, colour-matched, with masks, 27 blend modes and an undo step of its own. Several separate areas can run one after another, each at its own size.
- Draw boxes on an empty canvas or on the picture, a caption each, and the picture is generated to that layout (FLUX 3 Image, its open Klein model and MOSS-Image on your own ComfyUI, Ideogram 4). Reference layers are named in the prompt as
@img1,@img2. - Layers, groups, masks, clipping; filter layers (curves, colour grading, HSL, grain, 46 film looks, lens blur, haze and dehaze on a depth map) with a live scope and a clipping warning; text layers; retouch tools (heal, remove, liquify, dodge and burn, frequency separation); in-app cutout, depth maps and automatic layer names.
- Relight with virtual lights from the picture's depth map. Deep paint: paint, a light brush and wrapped logos that take the picture's own light and keep its reflections. Contact shadows under cut-outs; 3D objects and a posable mannequin as depth and pose guides.
- Upscale the selection or the whole picture; a Realism Pass for generated skin, hair and fabric; green and blue screen keying with CorridorKey 2.0; one picture split into layers; outpainting in one step; a frame from a video as a picture or a layer.
- Pictures up to 15,000 pixels stay fluid; export to PNG, JPEG, WebP, TIFF, and layered PSD and ORA.
- The pixels come from your own ComfyUI (local or remote, with your LoRAs), Comfy Cloud, or image APIs on your own keys (FLUX 3 Image, FLUX.2, GPT Image, Nano Banana, Seedream, Qwen, Ideogram, Krea and more). No account, no subscription; keys stay in the OS credential store.
- Every command is an MCP tool, 124 of them: Claude Desktop installs it with one click from the release's
.mcpb, Claude Code and any MCP client drive the window you are looking at, and--runworks through a folder of files as a batch. A built-in assistant on your own key uses the same commands.