Visual Awareness
Varve is adding task-specific visual assistance for composition, selection, effects, and future anchors. The product surface is the design workflow, not a list of model names.
What is available
Object Selection creates an editable mask from a point or box prompt. Background removal provides local fast, auto, and high-quality mask workflows. Both persist normal document mask data, so reopening does not require a model.
Protect Faces is available in crop mode: it detects faces with the bundled YuNet model (233 KB, MIT, checksum-verified, runs locally) and repositions the crop window to keep them in frame without modifying source pixels.
What is being built
- Face anchors request semantic landmarks such as Left Eye instead of exposing backend landmark indices.
- Select Person will use a genuine person mask, not a face rectangle.
Privacy and performance
Images are analyzed only when a workflow asks for a capability. Results are cached by source asset and revision, not by canvas position. The face model is bundled and inference is local; larger model downloads are explicit and verified. Face detection is geometric visual awareness, not identity recognition.
Current limitation
Face detection is shipped and verified against OpenCV parity; hand/pose/person workflows are not yet available and are not presented as shipped features.