Resoa is a small Santa Barbara studio. This org holds models for apparel manufacturing QC — construction attributes and garment parts for a tech pack — not ecommerce tagging.

Weights are Apache 2.0. Training data is Fashionpedia (CC BY 4.0). They were trained on street photographs of worn garments. Factory flat-lays and on-hanger photos are a different domain and are not measured here.

Models in this org:

• garment-detector-seg — instance segmentation of garments, accessories, and parts (Fashionpedia, 46 classes).
• garment-attributes — multi-label construction attributes from a single-garment crop (SigLIP2, 93.1M). Full-scene inputs degrade the scores sharply; that is a crop problem, not a model-card footnote.
• garment-attributes-v2 — same architecture and label space as v1, retrained with crop-scale augmentation.
• garment-crop-gate-nano — 47K pre-flight: is this crop tight enough to trust the attributes model? A full scene keeps about 14% of trained-crop mAP; the gate refuses those inputs.

Collection: Apparel QC. There is also a dataset, triclock, which is a LogLens memory-schedule benchmark — not apparel.

Same studio, two namespaces: this org is the product models. Tiny edge models (camera, audio, forensics) live on resoajoe — same person, kept there so research nanos stay easy to browse separately.