Three steps to transparent or solid backgrounds
1. Pick a PNG, JPEG or WebP image within 15 MiB / 16 MP. On first use the model weight is downloaded and real byte progress is shown (e.g. 3 MiB / 6 MiB); a failed download retries up to 3 times and Cancel ends the task without leaving an endless background job.
2. The tool walks through four stages: model loading → inference → mask editing → background compositing. Inference has no percentage source, so only its stage name is displayed. WebGPU is preferred, WASM is the fallback, and a device supporting neither receives an explicit unsupported message. A crashed worker can be re-initialised.
3. Refine hair, white-product or transparent edges with the mask brush, using undo/redo; the export always uses the current stack head. “Remove background” yields a transparent PNG preserving 0–255 alpha; “Change background” takes an opaque #RRGGBB colour (e.g. #FFFFFF) and forces alpha=255 on every pixel. Filenames look like cutout-remove-background-20261115T000000Z.png.
Fixed samples: portrait hair (Orientation=6, 4×3→3×4, feathered mask values 32/96/160/224), white-background product (4×3, mask 0/128/255), transparent edges (4×4, source alpha 0/64/128/255). In this round these samples come from an injected mask matrix and are NOT real model inference results.