Object Selection
Create an editable image mask from a canvas prompt, or use the result as a temporary pixel selection for downstream edits and export.
Workflow
- Select an image and activate Select Object, or choose Selection Sources > Select specific object when the target is one object rather than the foreground estimate.
- Choose Points for include/exclude prompts, or Box hint when a coarse region is easier to identify than a point.
- Set Prompt polarity explicitly. Shift remains a temporary exclude override for a point; it is not the same as subtracting an accepted selection.
- In Box hint mode, tap two opposite corners or drag a box. The box guides segmentation; it is not silently treated as a hard clipping boundary.
- Add more prompts to refine the temporary preview. If the candidate reaches an image edge without an extent prompt, Varve leaves it visible for inspection but blocks applying it; add an include point on that edge or draw a box around the full target. Tap a marker to remove it, drag a marker to move it, or press Backspace to remove the last one. Escape or Clear prompts discards the preview.
- Review the highlighted candidate and confirm it before choosing a Selection combination (replace, add, subtract, or intersect) if using Use as selection. Walk the candidates with the panel arrows or the [ / ] keys; the selection is validated against your prompts before it becomes selectable. Apply as mask always commits the reviewed candidate as an editable document mask. If you send the candidate to Generative Edit, Varve requires that first review and a second review in the compositing context; a moved image requires a fresh preview. Neither output runs inference again.
Magic Wand for linework and flats
- Activate Magic Wand and choose Visible artwork in Tool options when the sample should include the rendered active surface. Current layer keeps the existing single-raster workflow.
- Click inside a closed ink shape. Transparent artwork is sampled against white, so the selection can find the paper region while preserving antialiased line edges.
- Open Selection Sources, choose Create flats layer, set Fill color, then choose Fill pixel layer. Sampling and destination stay independent: the linework remains on its own editable node and the color lands on Flats.
- Use Edit > Undo or Redo to change the fill. For one combined PNG, select the linework and Flats layers together, group them, then export the group. Save the document to preserve both editable layers.
Visible-artwork samples are bounded and temporary. For contiguous visible-artwork selections, Gap closure radius closes only short contour openings within its 0–8 source-pixel limit; larger gaps can still reach the exterior. Edge expansion is a separate 0–8 source-pixel grow control for covering antialiased ink. Review the selection before filling.
Concept references stay visible in the document, but visible-artwork sampling excludes them until you explicitly opt in on the selected image.
What candidates mean
Candidate masks are alternative interpretations of the same prompts. They are not a list of separately discovered object instances. Multiple positive points can still describe one object, and a disconnected object or a hole is not automatically discarded or filled. Under-specified clicks are genuinely ambiguous: a point on a jacket can reasonably mean the jacket, the person, or the group, so Varve shows the competing candidates instead of pretending one is objectively correct. This is not a theoretical concern: on a reviewed corpus of real photographs with deliberately ambiguous single-click prompts, the provider's own top-ranked candidate was acceptable for the declared intent in two thirds of cases, and the remainder needed one keyboard cycle to reach an acceptable mask — so candidate order is a suggestion you can override, never a verdict. Adding prompts refines the active candidate; it does not silently switch to a different object. Use separate prompt sessions when you intend to select separate objects, then combine their accepted pixel selections deliberately.
Find a region by description
Find by description (Object Selection) runs a local text-conditioned detector, Grounding DINO Tiny, on the selected image. Type a concrete visual description such as red mug or person. dog., choose a match strictness, and review the returned boxes. Choosing one sends its box into the same Object Selection candidate path: the detection box is a discovery hint, never a finished mask, and nothing is applied without the normal candidate review.
The detector is an explicit 194 MB download and runs entirely on this device; the model's measured browser peak is 2.4 GB and Varve reserves 2.6 GB (2.6×109 bytes) for it, so low-memory sessions are refused before the download with a message that names the binding limit: either the device tier or a page without cross-origin isolation, both of which lower the safe budget. The editor worker runs the detector single-threaded, and that is a measured policy rather than a precaution: in a cross-origin isolated probe the threaded runtime never returned from session creation on the reference host, so threaded execution is unverified and not offered. The first browser run also pays a one-time session-compile cost that can take minutes on CPU. The panel shows measured stage timings — model load, image prep, tensor build, detection, detector release, and the total time to a reviewable result — and aborts with an explanation after eight minutes rather than searching forever; cancelling it also releases the detector instead of leaving it resident. After detections are materialized for review, the detector session is released before segmentation starts, and repeating the exact same search reuses the session's cached detections instead of re-running the model. Two honest limits matter. First, an open-vocabulary detector can highlight a plausible region even when the described object is absent — a measured control prompted dog on an elephant photograph returned a 0.71 box around the elephant — so scores are model similarity, not proof the object is present, and you must verify the highlighted region. Second, descriptions are grounded in concrete visual language: abstract or relational instructions are not parsed, and phrasing changes results. The panel reports an empty result rather than inventing a rectangle.
Select subject without a prompt
Selection Sources > Select subject proposes a foreground region for images with an obvious main subject. Choose an estimate quality first. Fast is the default and runs the bundled U²-Net Light model with no download, no account, and no GPU. Balanced prefers an installed IS-Net model, and High quality prefers an installed BiRefNet Lite model; when one is missing, the panel offers its download size explicitly and says which model actually ran. Varve shows the proposals for review first; only an explicit confirmation turns the reviewed candidate into a pixel selection or editable mask, and every alternative remains one click away.
The estimate is deliberately conservative about what it claims. Scores are shown as estimate scores rather than semantic confidence, and it never labels what an object is. A model-backed estimate works best when the subject stands out from its surroundings; a fully offline machine with no model falls back to the model-free heuristic, and the panel says so. Enclosed background-coloured regions stay inside a proposal rather than being punched out automatically, and the refine tools can remove them. For a specific object, use Object Selection with points or a box.
When a reviewed source-sized Object Selection is sent to Generative Edit without later brush or compound-mask changes, Varve keeps that source-resolution mask for inference and persistence; the scaled dialog preview is only for visual review. On reopen, the next edit decodes only the recorded source region, including the margin needed for grow, shrink, or feather settings. Editing or inverting the mask deliberately switches authority to the edited preview.
Accepted Generative Edits also retain compact selection evidence: the final mask fingerprint, source and placement identity, review checkpoints, and (for Object Selection) the chosen candidate, prompt geometry, rejected-candidate count, and target-topology diagnostics. This makes it possible to audit which reviewed pixels were used after reopening or moving the document between devices; it does not claim that a model understood an object label.
Model preference and honest fallbacks
Auto chooses only a validated, installed prompted provider that fits the current runtime and working-set budget. Faster local selects the optional MobileSAM split provider explicitly; it returns four candidates and is useful for testing a smaller download, but its real-photo browser gate found boundary-click ambiguity and its Node gate reached about 1.15 GB RSS on a 1920×2560 photograph. It is therefore experimental and is not silently selected by Auto. Higher detail selects the SAM2 Tiny workflow explicitly. Experimental — EfficientSAM-Ti installs and runs the upstream EfficientSAM split graphs for benchmarking: measured quality is equivalent to MobileSAM on the shared corpus, its predicted-IoU ranking is no stronger, it peak-used more memory than MobileSAM in the same run, and its decoder has no mask-prompt input, so it cannot be used for refinement round-trips. It is explicit-only and Auto never selects it.
If the requested provider cannot run, Varve explains why and keeps the prompts available. It does not silently replace prompted object selection with a foreground estimate. Model downloads are explicit, checksum-verified, and local.
Refine a mask or pixel selection
Open an image with a mask and choose Refine Mask in the Background Removal section or the Mask inspector. The brush has three explicit intents — Add / reveal, Subtract / hide, and Restore original — selectable in the inspector or with 1/2/3, with Alt as a hold-to-subtract shortcut. Brush size is measured in image pixels, and fast strokes are interpolated so no gaps appear between pointer samples.
Strokes are not clipped to the current pixel selection unless you enable Clip strokes to selection. That default exists because a refinement brush that can only paint inside the selection it is trying to grow cannot recover missing hair or foreground outside it.
Refine edges starts with Matting (binary-safe), because imported and segmented masks are often only 0/255. It solves a closed-form matting system over a spatial unknown band and can recover thin structures the current mask missed. Guided (already-soft edges) is an edge-aware smoother for masks that already carry partial coverage; it deliberately does not invent an alpha band from a hard boundary. If a solve is declined (for example, an extremely large unknown region on a constrained device), Varve says so and applies guided smoothing instead of committing a corrupt matte.
For difficult edges, Edit trimap paints categorical foreground, unknown, and background constraints. Unknown is a label, not 50% opacity; the matting solve estimates fractional coverage inside it. Manual constraints are preserved across re-solves, and your previous mask is used as a warm start and fallback.
Refine selection coverage
The Selection Sources inspector (and the Pixel Selection menu) refines area coverage without touching layer selection. Feather creates a Gaussian transition, Smooth reshapes the boundary, Grow and Shrink move it by a distance, Harden (contrast) steepens or flattens coverage, Threshold hardens it at a coverage cut, Antialias adds a one-pixel transition, Border selects a band inside, outside, or centered on the boundary, Shift edge translates a soft transition without flattening it, and Cleanup removes small islands or fills small enclosed holes.
Every operation is computed from the current selection and commits as a single undoable step. Marching ants trace the 50% coverage contour; feathered pixels beyond the outline are still partially selected, which is why paint and fill tools can appear to act slightly outside the visible boundary.
Models and privacy
The optional local model pair is downloaded explicitly from Settings > Offline Models, verified with its manifest checksums, and loaded lazily. Images are not uploaded by this workflow. Embeddings are temporary session data and are not stored in the document.
After applying, the document stores a normal raster mask resource, so reopening the document does not require the model to be installed.
For a tighter frame
Trim to Subject is a separate Crop & Bounds workflow. Its optional DETR model returns object rectangles only. Review the suggested bounds, choose a candidate when there is more than one, and apply them as a non-destructive crop. Use Object Selection when you need an editable mask or a temporary pixel selection.
Limitations
Object Selection is promptable segmentation, not semantic recognition. A click tells the model which region to segment; it does not guarantee that an arbitrary automatic result is the subject. A point near a textured object/background boundary can produce competing candidates, so review the candidate and add an exclude point, a deeper include point, or a box before accepting it. If a point lands on the proposed mask boundary, or a candidate touches the image frame without an extent prompt, Varve leaves the preview visible but blocks edits until you add stronger extent evidence. Hair, fur, glass, smoke, shadows, and translucent edges may need brush or trimap refinement.
Find by description is detection, not understanding. Grounding DINO Tiny is a general object-prompt detector: it does not read printed text, does not parse negation, counting, or relational instructions, and has a documented tendency to return confident boxes when the prompted object is absent. Treat every returned box as a proposal to verify, keep the default match strictness for unclear prompts, and use point or box Object Selection when you already know where the object is.
Hair and fur remain the hardest edges for any automatic matte: strands thinner than the analysis resolution, dark hair on dark backgrounds, backlighting, and motion blur all degrade the estimate. Varve re-estimates fractional coverage at source resolution with the closed-form matting refinement and keeps manual trimap constraints, but it does not claim laboratory matting accuracy. Glass, smoke, and reflections cannot be universally separated from a single photograph because their appearance depends on what is behind them; use the refine tools and check the composite over more than one background.
The Select subject estimate is a foreground proposal, not semantic recognition. The Fast level uses the bundled U²-Net Light model, optional higher-quality levels use installed IS-Net or BiRefNet Lite, and the model-free fallback (used when no model is available) is weakest on landscape or texture scenes. None of them labels what an object is, so Magic Wand and Object Selection remain the precise paths.