Remove a background through the shared local-first cutout pipeline, preview the clean edge, and hand an automatic result to interactive Select Subject refinement when the first whole-subject guess needs correction.
Where: Right-click an image → Remove Background; Image Editor → AI/Remove Background; Flowgen → Add → Image → Remove BG; or ask the AI assistant
Key ideas
- One segmentation path: All four product surfaces call the shared segmentation module instead of maintaining separate background-removal implementations.
- Color-decontaminated edges: The cutout estimates and removes the old background color from partially transparent edge pixels, then bleeds foreground color into transparent pixels so resizing does not pull the halo back.
- Local-first execution: The pipeline prefers a compatible local WebGPU backend and falls back to the bundled local imgly/IS-Net path. The imgly path requests GPU worker execution and retries on CPU/WASM when an adapter cannot compile it.
- Cold-start progress: The first run may need a large local model download. Progress is now surfaced instead of leaving the operation on an unexplained Processing state.
- Shipping BiRefNet path: Photographic cutouts prefer the 512px BiRefNet-lite browser model, verified on Apple/Metal. The incompatible 1024 export is not the shipping model.
- Screenshot/graphic fallback: When BiRefNet finds almost no salient subject, the resolver retries with imgly/IS-Net instead of returning a nearly blank transparent image.
- Interactive refinement: Canvas Remove Background exposes Refine with clicks; Image Editor exposes Refine with Object Select. Both preserve the automatic starting matte while letting you include/exclude specific objects.
Steps
- Choose the surface that matches the job: right-click for a quick canvas replacement, Image Editor for a layered edit, Remove BG for a reusable graph output, or the assistant for a broader task.
- Start the removal and keep the app open while the first-use model downloads; watch the progress status.
- Inspect hair, fur, glass, fabric, and other soft edges over both a light and dark temporary background.
- Refine the mask in Image Editor when the automatic subject choice or fine edge is not sufficient.
- Keep the transparent PNG or layer mask as the reusable source for later composition, vectorization, or generation references.
Choosing a surface
- Canvas context menu: quick remove-and-replace on a placed image.
- Image Editor: composite or selected-layer removal followed by mask/pixel refinement.
- Flowgen Remove BG: typed image_in → image_out step that preserves provenance and can feed later nodes.
- AI assistant: the same engine inside a larger requested edit.
Why the edge looks different
A soft matte alone is not enough: partially transparent pixels still contain some of the removed background. Color decontamination reconstructs foreground color before compositing, which is why a cutout can move onto a new background without a light or dark rim.
Current backend ladder
- BiRefNet-lite 512 on compatible WebGPU devices for photographic matting.
- Coverage guard below 0.5% prevents blank screenshot/graphic results.
- imgly/IS-Net local fallback for unsupported devices or non-salient graphics.
- Shared color decontamination and transparent-pixel color bleed for reusable edges.
Refine instead of restart
After the canvas dialog produces a result, Refine with clicks opens the same hover/click/box Select subject surface with the cutout alpha as its starting selection. In the Image Editor, Refine with Object Select traces the mask already applied to the layer and switches tools without a second removal run.
Limitations and important notes
- Automatic salient-subject detection can choose the wrong object in a page screenshot or multi-subject image.
- BiRefNet is registered as the preferred backend, but the tracked runtime still fetches the original graph. Apple adapters are gated before download; wider adapters can still fail and fall back. An experimental Split-rewritten graph clears the storage-buffer error but still exhausts browser memory and is not wired into the product.
- The first local model download is large and still comes from an external model CDN until the weights are self-hosted.
- The shipping BiRefNet path runs at 512 model resolution and is upsampled to the source size; a same-photo head-to-head against the 1024 imgly path remains an image-dependent quality judgment.
- The first local model files are still downloaded from external model hosts.
- Promptable refinement requires WebGPU; whole-subject removal retains a fallback.
Troubleshooting
The first run appears slow
Keep the tab active and watch the download/inference progress. Later runs reuse the cached model session.
The wrong objects were kept
Open the result in Image Editor and refine its selection/mask, or start from a tighter crop with one clear subject.
A rim is still visible after resizing
Inspect the original cutout at 100%, then refine or contract the mask slightly in Image Editor before scaling it down again.
Related guides
- Remove Background Node
- Use AI Fill, Erase, Restyle, outpainting, and Remove Background
- Image Editor layers, masks, adjustments, filters, and workspaces
- Image and Media Management
- Select a subject or object across flowmo
- Object Select node: click, add/remove objects, boxes, cutouts, and masks
- Image Editor Object Select: hover, click, and multi-object boxes