Spaces:

---

AxleToe

/

Concept_Attn_SD3

like 0

Sleeping

I am building a visualization space for ConceptAttention paper (Diffusion Transformers Learn Highly Interpretable Features). The official ConceptAttention space seems to be currently down, I am creating this space to ensure that there's has a live, working playground to explore and study the paper's findings.

1. I have completely stripped out the heavy T5-XXL text encoder and its tokenizer, running only on half-precision CLIP text embeds.
2. I have optimized and successfully tested the entire patched attention-extraction pipeline on a Kaggle GPU instance.
   And the code is public. A zeroGPU grant would be incredibly helpful!
   thank you for your support.
   here's a demo: