Reproduce “Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Evelyn Turri, Davide Bucciarelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia
The authors transfer selected internal activations between text-conditioned image generations to combine prompts without blending pixels. They rank channels by absolute mean per sample, layer, and stream, then replace target image-stream values only within a source-derived spatial mask. On FLUX.1-schnell, masking raises joint DINO-I from 3.0 to 20.1 and joint CLIP-T from 32.4 to 44.1 on these pairs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.