Segmentation

Every rendered look is automatically segmented — each garment on each model gets its own mask — so that ReviewHub can attribute a grade to one garment on one model instead of only to a whole image. Garment-level grades are the highest-signal feedback the system receives.

Pipeline (v3)

1. SAM 3 person-detect → 3 model crops (fallback to fixed thirds)
2. Per manifest slot, SAM 3 concept prompts using the garment's NAME ("Quilted Technical Coat") — attribute-qualified names work; generic "jacket" does not
3. Sanity checks (area range + centroid position per slot); on failure, Gemini describes what is ACTUALLY worn (deviation handling) and SAM retries with that phrase — deviations are recorded (deviation_detected)
4. Layer subtraction so crops show only their own garment
5. Outputs per garment: mask SVG, crop PNG, and a grayscale hitmap where pixel value = garment index (1-based — 0 is background; ReviewHub reads the pixel under the cursor to know what you clicked)

Reliability

• Per-image SAM calls run in parallel; batches run 8 images concurrently
• A 10-minute reconciliation sweep segments anything that slipped through (fire-and-forget tasks die with their process; the sweep cannot miss)
• Failures land in pipeline_events with reasons