Technical Matching Process

How the prototype reads real fabric photos.

Matching Pipeline

This demo uses the 15 supplied paired examples. Production would use the same concept with CLIP, SigLIP, or DINOv2 embeddings plus pgvector or Qdrant.

Upload photo
Isolate fabric area
Normalize lighting
Compare clean + angled references
Score color, texture, pattern
Return ranked matches

Prototype logic: every fabric has a clean swatch reference, a real camera-angle reference, and the original paired example sheet. Uploaded photos are compared against all references, including left/right crop variants for paired images.

Accuracy framing: this is still a prototype similarity system, not a guaranteed production AI model. The important improvement is that it is now trained on real angle examples instead of only flat catalog images.

Production path: precompute embeddings on the backend, store them in a vector database, add metadata filters, and keep adding real good-match/bad-match feedback from designers.