Object Selection
From object to editable mask. Select an image region on the canvas, refine it with include and exclude prompts, then reuse the mask wherever the design needs it.
Select with the canvas
- Choose point prompts or an explicit box-hint mode; create a box by dragging (in any direction) or by tapping its two corners.
- Refine one result instead of restarting: each additional prompt continues from the candidate you kept, so the mask follows your edits rather than jumping to a new reading on every click.
- Set include or exclude polarity visibly. Shift remains a temporary exclude shortcut, separate from selection subtraction.
- Tap a prompt marker to remove just that prompt, drag it to move it, or press Backspace for the last one.
- Compare candidate masks before applying — with the panel arrows or the [ / ] keys while the tool is active; the candidate you reviewed is the one that is committed.
- Preview honestly: an overlay tint marks what the model highlighted, and a cutout preview fills what the mask will remove with checkerboard so you can see the applied result before committing to it.
- Three outputs, all consuming the exact reviewed candidate: an editable non-destructive raster mask, a temporary pixel selection (with replace, add, subtract, or intersect), or an extracted copy of the image on its own layer — the source stays untouched and the extracted layer is selected and ready to move.
Prompts are not object discovery
Candidate masks are alternative readings of the same prompt, not separate detected instances. A single click is genuinely ambiguous — the same point can mean a part, the whole object, or a group — so Varve keeps the competing readings and lets you choose, with the reviewed candidate pinned for Apply. A reviewed corpus of ambiguous real-photo prompts found the top-ranked candidate was acceptable in about two thirds of cases, with the rest one keyboard cycle away, so the order is a suggestion you can override. Multiple points may still describe one object, and Varve keeps disconnected pieces and holes instead of silently keeping only the largest component. Use a new prompt session when you mean a different object, then combine the accepted selections intentionally.
Describe it in words
Find by description runs a local text-conditioned detector (Grounding DINO Tiny) on the selected image. Type a concrete description such as red mug orperson. dog., inspect the image-backed highlight, confirm the region you mean, and send that box into the same candidate-review path. Detection is a discovery stage: it never applies a mask by itself, and the box is only a prompt for the segmentation model.
It is an explicit optional download and runs fully on-device. Be precise with wording: open-vocabulary detection is grounded in visual language, and a measured control showed it can return a confident box even when the described object is absent — so scores are model similarity, not proof, and the highlighted region must be verified. No match shows an honest empty result instead of a fabricated rectangle. Treat it as a background-scale task rather than a live filter: Varve runs it once per explicit search, shows measured stage timings and a running clock instead of a fake progress bar, lets you cancel, and releases the detector before segmentation starts. Point and box prompts stay available the whole time and need no download.
A subject estimate, fast by default
When you have an obvious subject but no prompts in mind, Selection Sources > Select subject proposes foreground regions on-device. The default Fast level runs the bundled U²-Net Light model — no download, no account, no GPU — and a model-free estimator still covers offline machines with no model at all. Balanced and High quality levels use optional installed models when they fit the device; the result always names the model that actually ran, and a higher-quality level never silently becomes a different one.
Proposals are ranked foreground estimates: preview a candidate first, then explicitly use it as a pixel selection or apply it as an editable mask. Every alternative stays one click away. Scores are shown as estimates, never as semantic confidence, and the copy never claims to know what an object is. Portrait (MODNet) is a reviewed portrait matte, not a guarantee that one person is isolated; framed or complex photographs can retain nearby background, so use Select specific object when one item is the target.
Choose the task, not a mystery model
Use a subject estimate for a quick foreground proposal, or choose Select specific object in Selection Sources when you can point to the object you mean. Object Selection then uses the same canvas prompts from every entry point, and Improve edges is available when the boundary needs care.
Auto picks the best-validated local segmentation model that your device can safely afford: capable providers are checked against the measured working-set budget before quality is compared, and a provider that cannot fit is refused with an explanation instead of risking a crash or a silent downgrade. Faster or higher-detail providers are explicit advanced choices, not surprises, and model names stay out of the normal workflow. An experimental EfficientSAM-Ti option exists for benchmark comparisons; it measured quality-equivalent to MobileSAM with a larger working set and cannot take mask prompts, so Auto never selects it. The candidate you review is the candidate that gets committed, and the result becomes an ordinary mask — reopening the document never requires the model again.
Selection and trimming are different tools
Object Selection creates a pixel mask from your prompts. Apply it as an editable mask, extract it to its own layer, or use it as a temporary pixel selection without changing the artwork. When you only need a tighter image frame, Trim to Subject can suggest object bounds separately; it shows the detected box for review before applying a non-destructive crop, and if the source it needs cannot be measured — an empty mask, unreadable pixels, a fully transparent image — the trim fails with an explanation and the document is left exactly as it was, instead of silently resetting your crop. The bounds workflow is not a hidden mask and does not claim semantic recognition.
For line art, use a colour-region selection
Object Selection is prompt-driven segmentation. For a closed ink drawing that needs flat colour, the Magic Wand can sample Visible artwork and write the fill to a separate Flats layer instead. Its edge expansion grows coverage by up to eight source pixels beneath antialiased ink; it does not repair open contours. This pixel workflow needs no segmentation model. See the linework-to-flats steps.
Local by design
Optional segmentation models are downloaded explicitly from Settings > Offline Models, verified before use, and run on the device through Varve's worker-backed inference path. Higher-quality subject estimates offer their optional model in place with its download size, and the bundled Fast level works without one. A model is not loaded at application startup, and image embeddings are kept in a bounded session cache rather than stored in your document.
Set the right expectation
Promptable segmentation follows your points and box; it is not a guarantee of semantic subject understanding or perfect alpha matting. Hair, glass, smoke, and similar boundaries may need the existing brush or trimap refinement tools. Select subject is a foreground estimate that works best when the subject stands out from its surroundings, and the model-free fallback is weakest on landscape or texture scenes. Review the mask against the final background before export.