Scumble update: draw boxes, name your references with @img1, select by depth
Scumble update: draw boxes, name your references with @img1, select by depth
A few weeks ago I wrote here about Scumble, my free, open-source desktop editor for AI inpainting. It has had a lot of releases since (it's on 0.1.44 now), and three of the new features change how I work with my own generations every day. This is a quick tour of those three, with the tips I learned the hard way.
Quick recap: you select part of a picture, say what should be there, and the result comes back as a layer of its own, colour-matched to the picture around it. It runs on your own ComfyUI, on Comfy Cloud, or with just an API key (FLUX 3 Image, FLUX.2, Nano Banana, GPT Image, Seedream and more). Free, GPL-3.0, no account.
1. Box prompting: say where, not just what
The hardest part of inpainting has always been placement. You select an area, prompt "a cat", and the cat ends up squeezed into a corner, or twice as big as it should be. FLUX 3 Image from Black Forest Labs understands bounding boxes in the prompt, and Scumble now gives you a proper tool for them.
How it works
Pick the Boxes tool (X) and drag a box on the picture. Each box gets its own colour and a name.
Double-click a box to type what goes in it, right on the canvas ("a black cat sitting on the bench").
The Boxes switch under the prompt field turns on with your first box. Generate sends every box with your prompt, written the way the model's documentation describes it.
Six kinds of box
New: put what the description says into this box.
Keep: this element stays exactly where it is.
Move: take something from a source box to a new place (shown with a dashed source and an arrow).
Remove: take it out (hatched).
From reference: place one of your reference images (or a part of it) into the box.
Text: render these words here.
No boxes? Your selection becomes one. With the switch on and no boxes drawn, Generate sends your selection itself as a box, so the model knows where in the crop the change goes. In my tests the cat finally sat in the middle of the selection instead of at its edge.
Compose a picture from scratch. On an empty canvas, draw three boxes, describe each one, add a line about the background, and Generate new builds the picture with every element where you drew it. That's the GIF on the GitHub page: a café scene built from three boxes.
Tips
Draw the box at the size the thing should have. The model fills the box; a huge box gets a huge cat.
Boxes smaller than about 40 pixels often come back empty. Zoom in on the area and use a bigger crop.
On a new image, describe the background in the prompt too. Without it, the model may paint the empty space around the boxes as a transparency checkerboard (Scumble warns you when the prompt is empty).
Boxes work with FLUX 3 Image on every provider Scumble runs it on (Black Forest Labs, Comfy Router, OpenRouter, fal, WaveSpeed), and with Ideogram 4 on fal (New, Keep and Text boxes).
2. The prompt field: name your references with @img1
Edit models take extra pictures (a character, a product, a style), but every provider numbers them differently. "Use the jacket from image 2" breaks the moment the order changes. Scumble's prompt field now handles that for you.
Every reference layer is labelled img1, img2, ... in the layer panel, the reference list and on the canvas.
Type @ in the prompt and a list of your references opens, with thumbnails. Pick one, and it becomes a chip: a small round picture of the reference plus its label.
When you hit Generate, each
@img1goes out as the name the chosen model uses for that picture: "image 3" when the crop and the original go first, or whatever your local ComfyUI workflow expects. The status line tells you what each token became.Tokens follow their pictures. Move, hide, merge or delete references, and the prompt is rewritten so every chip still points at the same image. A chip whose reference is hidden is struck through, and Generate tells you why it won't run instead of sending the wrong picture.
+ Add reference at the end of the @ list loads new pictures as reference layers right from the prompt.
The prompt field has its own undo (Ctrl+Z / Ctrl+Y while you type), separate from the picture's.
Combined with boxes, this is my favourite trick: a From reference box plus @img1 in the prompt puts that object at that spot. In my live test a wall clock from a reference layer landed exactly inside its box.
3. Select by depth (new in 0.1.44)
This one is brand new and it's my favourite for finishing.
A depth map in one click. Selection › Depth computes a depth map of your picture inside the app, with Depth Anything V2 Small (an open model, 99 MB, Apache-2.0; download it once in Settings › Helpers). It runs on your GPU, or on the processor if you have none. A 15,000-pixel picture takes about a second. The map is saved with the document, follows crops and turns, and Show map draws it over the canvas.
Select by range. The Selection panel's Select by range block selects what lies between two depths, between two brightness levels, or near a colour:
a range bar over the picture's own histogram, so you see where the background starts;
soft ends as wide as you like (no hard cut-out edges);
Invert, a live tint on the canvas while you drag, and replace, add, subtract or intersect with what you had.
Limit a filter by depth. Any filter or fill layer can act only where the picture lies in a range. Haze that grows with distance is just a fill layer with a depth limit. Grade the background only, sharpen the foreground only, darken the far sky: no hand-painted masks. You can still paint on the layer's mask to refine it.
What I use it for
Replace a background without touching the subject: select far, then Generate a new background in that selection. The soft ends of the range keep the transition from looking cut out.
Atmosphere: a warm fill layer limited to the far range, blend mode Screen, low opacity. Instant depth.
Fix only the foreground: inpaint hands or a face with a near-range selection, so the background never changes.
Also new: Alt+click a layer's mask label to see the mask alone in black and white, and the Object tool now takes a box: drag a box around something and SAM2 selects the object in it.
Get it
Scumble is free and open source (GPL-3.0), with no account and no subscription. Windows (Microsoft Store or installer), Linux builds, and macOS is coming.
Microsoft Store: https://apps.microsoft.com/detail/9NDBTNNMXF2R (or
winget install 9NDBTNNMXF2R)Videos and manual: https://www.denrakeiw.com/scumble
It's built by one person, so if something breaks or you miss a feature, tell me in the comments or open an issue on GitHub.
And if you like it, a ⭐ on GitHub helps a lot.