Reproduce “Text-to-Image Models Need Less from Text Encoders Than You Think” 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
Nurit Spingarn, Noa Cohen, Tamar Rott Shaham, Tomer Michaeli
The authors test whether text-to-image models need contextual embeddings. They replace embeddings with representations preserving word identity, multi-token cohesion, and order, but removing prompt context, while leaving the image model unchanged. Across three diffusion-transformer systems, BoPTW guides generations about as well as full embeddings; on GenEval Single Object, its SD 3 non-inferiority rate is 88%, with long-prompt and text-rendering exceptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.