DRSY/MoTIS

Implements knowledge-distilled dual-encoders (6-12 layer Transformers) achieving CLIP-parity retrieval on MS COCO while reducing model size to 85-146MB and inference latency by 1.6-2.9x through layer pruning and supervised distillation. Provides multiple indexing strategies (linear scan, KMeans, Spotify Annoy) with lazy loading—encoding high-resolution images in the background while displaying thumbnails—and transpiles CLIP's tokenizer and preprocessing pipeline into native Swift/iOS via TorchScript.

126 stars. No commits in the last 6 months.