Skip to content

Latest commit

 

History

History
11 lines (7 loc) · 588 Bytes

denser-CLIP.md

File metadata and controls

11 lines (7 loc) · 588 Bytes

Denser CLIP representations



Max Woolf has some shallow resuls suggesting that we might have more powerful CLIP representations by incorporating the activations from additional penultimate layers: https://github.com/minimaxir/imgbeddings/blob/main/DESIGN.md

I think this merits looking into. really wouldn't be that hard, basically just need to add this representational strategy to a CLIP evaluation harness.