Reproduce “SplitFlux: Learning to Decouple Content and Style from a Single Image” 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

Yitong Yang, Yinglin Wang, Changshuo Wang, Yongjun Zhang, Ziyang Chen, Shuting He

The SplitFlux framework enables robust content and style disentanglement from a single image within the Flux Diffusion Transformer model, facilitating flexible re-embedding of content into new contexts. It achieves superior content preservation and recontextualization compared to prior methods, while operating with significantly fewer training parameters.

Key insights

  • Within the Flux model, single stream blocks (20-57) are crucial for image generation, while double stream blocks (1-19) have negligible impact on semantics.
  • There is a functional specialization among single stream blocks: early blocks (20-29) primarily control image content, whereas later blocks (30-57) primarily determine image style.
  • Semantic boundary blocks (30-31) are critical for simultaneously processing content features and aggregating high-level style representations, requiring targeted adaptation to prevent content leakage.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.