AI Eyes/Mouth mask components with one-click portrait setup

[SandeepSubba]

What

Adds Eyes and Mouth AI selections to both the Masks panel and Inpainting, plus a One-Click Portrait button that assembles the classic beauty-retouch stack automatically. The motivating workflow: skin softening needs subject minus eyes minus mouth, which previously meant manual brushing over every eye and lip.

How

Detection — YuNet (MIT, ~230 KB), auto-downloaded and SHA-verified exactly like the existing models. Unlike a plain box detector it returns 5 landmarks per face (eye centers, nose tip, mouth corners), so the masks are anchored on the real features:

• ellipses centered on the landmarks, sized by interocular / mouth-corner distance, rotated with head tilt
• input is letterboxed (aspect distortion measurably degrades landmark precision)
• a second detection pass on the face crop refines the keypoints (the face fills the 640px input instead of a fraction of it)
• multi-face: every detected face gets its ellipses; per-component Grow/Feather for fine-tuning

Threshold — 0.8, validated against a 64-image RAW sample library: every real face scored >= 0.81 (including tiny background faces, backlit, red stage light, glasses, headwear, tilted heads), while false positives (a bare back, foliage bokeh) topped out at 0.72.

One-Click Portrait (Masks panel) — builds three masks in one press:

• Skin Softening: foreground − eyes − mouth, negative clarity (if Frequency-separation skin smoothing (Smoothing / Texture / Scale, global + per-mask) #1326 lands, this can switch to frequency-separation smoothing)
• Eye Enhance: eyes mask with mild clarity/exposure/saturation
• Teeth Whitening: mouth mask with HSL yellows desaturate + lift, so red lips are untouched

Mask UX — visible "+ Add / − Subtract" chips on each mask (subtracting a component was previously three hover-submenu levels deep); Eyes/Mouth are top-level entries in the creation grid and component menus.

Undo correctness — new amendHistory store action folds async generation results (mask bitmaps, generative replace, quick erase) into the creating action's history entry. Previously each AI action left 2–3 history entries, so a single undo left an empty edit behind; now one action — including the whole portrait stack — is exactly one undo step.

Testing

• Placement verified across all people photos in a 64-image RAW library (frontal, 3/4, tilted, glasses, open-mouth expressions, multiple faces, dark/backlit/colored light) via a pipeline-identical harness, then confirmed in-app
• cargo check clean; tsc introduces no new errors
• Single undo after One-Click Portrait removes all three masks; existing edits untouched

🤖 Generated with Claude Code