Can I build a style image library with pooled SigCLIP features for more robust and generalizable style transfer?

[BeiningWu]

Hi! Great work on USO. Quick question:

Is it feasible to create a style image library that stores pooled SigCLIP features (from your hierarchical projector) to achieve more generalizable style transfer for specific art styles?

The idea is to pre-compute and pool features from multiple images of the same style, then use these pooled features for better style generalization.

Any obvious technical barriers I should be aware of?

Thanks!

[fenfenfenfan]

Hey, that's a really cool idea!
Yep, that should totally work. SigLIP features are natively aligned to the textual space in USO, so those precomputed, saved features are basically like custom textual inversions but for style.
The only thing you'd need to figure out is the best way to "pool" them. We've found that just sticking them together (concatenating the tokens, e.g., two 64-token features become one 128-token feature) works well for blending styles.