New AI Can Identify Any Fabric — And Slashes Costs by 90%
Cheaper, faster AI could speed up clothes shopping, recycling, and quality checks.
Researchers built a multi-agent AI system that recognizes a fabric's structure, hitting 93.94% top-1 accuracy on a newly curated 14-class benchmark. The trick: a CNN cascade handles the easy majority of images, and a vision-language model — which costs much more per image — is invoked only as a selective arbiter, constrained to a top-3 taxonomy-consistent choice. Tightening the VLM trigger cut API cost by roughly 90% with no measurable accuracy loss, and CPU inference runs in 93 ms or less without a VLM call. Each prediction carries a machine-readable reasoning record, offered as an entry point for future supply-chain documentation.
- A small, cheap AI handles most fabric photos; a big, costly AI is only called in when the small one is unsure.
- The hybrid reached about 94% accuracy and nearly 95% on the hardest fabric types, up from roughly 77%.
- Sending fewer than 10% of images to the expensive AI cut costs by around 90% with no accuracy loss.
Why It Matters
Cheaper, faster image AI means quicker textile recycling, better quality checks and lower costs passed to shoppers.