Background Remover & Solid Colour
Remove image backgrounds on-device, export transparent PNG or a solid colour background; four-stage pipeline, editable mask, undo/redo, 15 MiB / 16 MP per image.Processed locally (model file must be downloaded)
Operations
Remove background (transparent PNG)
Remove the background of one image and export an alpha-preserving transparent PNG; supports feathered hair, mask brush edits, undo/redo and EXIF orientation alignment.
Change to a solid colour background
Replace the background with one opaque #RRGGBB colour, forcing alpha=255 per pixel; semi-transparent edges blend by effective alpha so no accidental transparent pixel remains.
Real examples and limits
Initial budget: one image, 15 MiB / 16 MP; input PNG/JPEG/WebP. Model weights must be downloaded on first use (not shipped in this round). Original-size export does not imply full-resolution inference. This delivery is a deterministic mask-composite slice: real model inference, WebGPU/WASM workers, browser Canvas decoding and WebP encoding are untested; a model manifest and licence audit are required before release.
FAQ
Are my images uploaded?
No. Pixels stay in your browser; there is no cloud processing in this round. Model weights must be downloaded once, but your image is never sent with that request.
Why is there a first-use model download?
Background removal needs a model weight file. The download stage shows real byte progress (e.g. 1024/4096 bytes); inference has no reliable percentage, so only its stage name is shown, never a fake progress bar.
Does it work without WebGPU?
Capability negotiation prefers WebGPU and falls back to WASM. If neither is available the tool says so explicitly instead of faking a result. A failed worker can be re-initialised up to 3 times before a clear error.
Will a transparent PNG lose its alpha?
No. Effective alpha = source alpha × mask / 255; a fixed sample preserves the 0/64/128/255 alpha sequence. A solid background forces alpha=255 per pixel, so no accidental transparency appears.
Which mask version gets exported?
Export always uses the current head of the undo/redo stack; tests compare a maskHash of the stack head with the exported result.
Do EXIF-oriented photos misalign the mask?
No. Source RGBA and mask are transformed together for orientations 1–8; a fixed Orientation=6 sample aligns 4×3 into 3×4.