Visual awareness, on the designer's terms.
Varve is building task-specific visual tools that help with real design decisions without turning your document into a model playground.
Available today
Protect Faces is available in crop mode: a small bundled local model (233 KB, MIT) keeps detected faces in frame with a non-destructive source crop. It runs a whole-image pass plus bounded native-resolution windows, so faces down to roughly 10 px are still found on large photographs — the same three faces a whole-image downscale missed on our own test facade are detected again with the windows.
A detected box is an approximate face bound for framing, not a face, hair, or person mask, and there is no identity recognition. For a precise cutout use Object Selection or Background Removal, which are separate local paths. Detection is confidence-gated: “no faces detected” means none were retained above the floor, not that a photograph has no faces, and it never disables the manual matting tools.

In development
- Face-protected effects and static landmarks for meaningful anchors.
- Additional hand, pose, object, and person workflows only when a design feature needs them.
Local by default, honest by design
Images are not uploaded for these workflows. The face model is a small bundled model (233 KB, MIT) verified by checksum, and larger model downloads are explicit and checksum-verified when present. Importing an image does not silently trigger a general vision scan, and visual results do not become a hidden rendering dependency.