Introducing Pasma: local clothing recognition — demo and feedback
We're releasing Pasma at Vestes Archive: open-weight clothing recognition you can try locally in your browser.
Choose a photo of one garment → get the top three garment types, a category and visual material cues. The demo downloads ONNX weights and runs inference locally; your photo is not uploaded.
The release includes 20 garment types, 6 categories, 21 visual material scores, ONNX/CoreML exports and a Python example. Project code and fitted weights are Apache-2.0, with upstream notices retained.
Small-320 scored 89.12% Top-1 and 97.10% Top-3 on 1,103 held-out real photos in PyTorch FP32. The browser uses FP32 weights (~20–33 MiB depending on variant, plus runtime); 10.32 MiB refers to Small/Nano FP16 exports.
This is a v1 candidate. Material appearance is not fibre analysis, blouse/shirt confusion remains, and phone performance has not been validated. Evaluation data is private; the model card explains reproducibility and export limitations.
Try one garment and tell us where it struggles. Browser/device, model variant and whether Top-3 was useful are helpful details. You don't need to share a private photo. If the model helps your work, a like on the model page or a GitHub star is appreciated.