Choose the right subject-selection surface for a quick canvas cutout, a precise layered Image Editor selection, or a reusable Flowgen cutout-and-mask step.
Where: Canvas right-click → Select subject; Image Editor → Object Select; or Flowgen → Add → Edit → Image → Select Subject
Key ideas
- Selection and background removal are different jobs: Remove Background asks for one automatic foreground matte. Select Subject asks what you mean by the subject and accepts points, exclusions, a region, or a layer/composite scope.
- Three surfaces, three outcomes: Canvas Select subject applies a transparent replacement to selected elements; Image Editor Object Select creates an editable pixel selection; Flowgen Select Subject produces both a cutout and a mask as graph outputs.
- One shared local-first segmentation layer: Promptable surfaces reuse SAM 2.1 image embeddings for fast follow-up points. Whole-subject removal prefers BiRefNet-lite 512 for photographs and falls through to imgly/IS-Net when the first matte is nearly empty or the device cannot use it.
- Boxes mean objects, not pixels: A box is a region prompt that unions distinct detected objects. Use an ordinary rectangular marquee when you want every pixel in a rectangle regardless of object boundaries.
- Flowgen instructions stay source-connected: For assistant-authored flows, put the subject description in a connected Text source or the exact compatible linked-Brand field; do not write it as arbitrary inline node data.
Steps
- Choose the surface based on the result you need: quick applied cutout, editable selection in a layered Image Editor document, or reusable Flowgen cutout/mask outputs.
- For a placed website/video image, right-click → Select subject; hover, add or exclude points, box-select objects, compare tighter/looser candidates, inspect the preview backgrounds, and Apply.
- For layered editing, open Image Editor → Object Select (J), choose the selection mode, then hover/click or box-select and convert the result into a mask/edit.
- For Flowgen, add Select Subject, connect the image, then use a connected Text source or the linked project Brand’s exact compatible field for the subject description. Connect the source to the prompt field and the image to image_in.
- Run the node and wire Cutout or Mask to the next node that needs the result. Keep the source image connected so provenance and edits remain visible.
- If an automatic Remove Background result is close, use Refine with clicks/Object Select instead of restarting from nothing.
- Always review soft edges, holes, transparent materials, overlaps, and small disconnected parts at 100% before publishing or rendering.
Surface chooser
- Canvas: fastest path to replace the selected image/background with a transparent cutout.
- Image Editor: selection-first path for masks, compositing, local retouching, AI edits, and saved selections.
- Flowgen: graph-first path with explicit image_in, image_out Cutout, and mask_out Mask sockets.
- Remove Background: automatic whole-subject path when pointing is unnecessary.
Gesture vocabulary
- Hover: preview the object under the pointer.
- Click/positive point: include.
- Alt-click/negative point: exclude.
- Box: select/union objects inside the region.
- Tighter/Looser: move among part/whole candidate masks where exposed.
- Start over/Clear/Undo: correct the authored prompt without changing the source image.
Local model behavior
Interactive segmentation needs WebGPU and downloads the local model on first use. The expensive image encoding is reused while you refine the same source. Whole-subject matting also runs locally when supported and has a lower-capability fallback so a missing WebGPU path does not turn every cutout into an error.
Tips
- Prefer points for one object, a box for a small set of objects, and Remove Background for a clear single-subject portrait/product shot.
- Keep the mask as a separate asset when another edit may need to expand, contract, blur, invert, or reuse it.
- A good include point sits well inside the object; a good exclude point sits well inside the unwanted region.
Limitations and important notes
- Interactive point/box selection requires WebGPU.
- First use can be slow because model files are downloaded from external model hosts and compiled locally.
- Subject segmentation works on still images; video subject tracking/reframing is a separate Video Director capability.
- Automatic models can misread illustrations, screenshots, overlapping objects, glass, smoke, or very small details.
Troubleshooting
Select subject is unavailable but Remove Background works
The device likely lacks WebGPU. Use the whole-subject fallback or move the job to a WebGPU-capable browser for promptable selection.
The automatic cutout is blank on a screenshot or graphic
The shared engine should fall through from BiRefNet to imgly/IS-Net when the matte is nearly empty. Retry, then use Select Subject or Image Editor selection tools if the graphic has no clear salient subject.
Edges look correct on white but wrong on dark
Inspect both backgrounds, refine the selection/mask, and keep the color-decontaminated result rather than rebuilding transparency from a hard black/white matte.