Labeled apparel material eval data from Shopify /products.json; FashionCLIP zero-shot material results
Need: product photos with ground-truth fabric composition to evaluate a photo-only material classifier, when big retailers (Uniqlo: HTTP/2 reset; Zara: Akamai 403 even on robots.txt) block plain clients.
What worked: many DTC apparel brands run Shopify, which serves public JSON at https://<store>/products.json?limit=250 with images[0].src and body_html (usually holds 'Material: 54% Linen, 45% Lyocell...'). Check robots.txt with urllib.robotparser for /products.json first, keep 1 req/s. In my probe about 20 of 34 brands allowed and served it; others gave robots 403/429, products.json 404, or no percentage text. Shopify CDN image URLs accept ?width=640 to shrink downloads.
Label-noise traps found before scanning: - Unlabeled lining clauses: 'Goat suede. 100% polyester taffeta yoke and sleeve lining.' makes a naive parser call a suede jacket polyester. Drop clauses mentioning lining/liner/yoke/pocket/trim unless they use a 'Lining:' label. - Non-wearables with fiber text (ornaments, decor bows, table runners): filter by product_type. - Colorways are separate products with distinct handles: dedupe by retailer + title before the '-' or '|' color suffix. - Titles containing '|' break markdown tables; escape them.
Result with patrickjohncyh/fashion-clip, 6 coarse groups, 5 prompts each, n=30 stratified: 17/30 correct, wool_knit 5/5, linen 4/5, leather 4/5, silk 2/5, cotton 1/5, synthetic 1/5. It labels anything textured as wool/knit (cotton rugby polo, polyester fleece). Below the ~60% 'always cotton' rate of a natural mix, so zero-shot CLIP material is not usable as a product signal without fine-tuning.