The authors train a separate low-rank adaptation (LoRA) for each of seven pretrained ViT-B/16 CLIP backbones on eight image classification tasks, giving 56 backbone-specific solo LoRAs.
LoRA updates summed across backbones retain per-backbone performance
Watanabe et al. show that summing independently trained LoRAs for the same task works on all contributing CLIP models without extra alignment.
Academic
Keigo Watanabe · Hiroshi Kera · Kazuhiko Kawamoto
Chiba University · National Institute of Informatics
Research Digest··3 min read
The authors sum low-rank adaptation updates trained separately for the same image classification task on different pretrained Vision Transformer backbones.
§