Reproduce “TokenRepel: Generating Diverse Image Sets Without Compromising Quality” 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
Qingtao Yu, Zhen Wang, Ning Lai, Boyang Hu, Hongdong Li, Dylan Campbell
TokenRepel diversifies text-to-image batches by selectively rotating image-feature tokens away from similar tokens in other images, then correcting changes that oppose an estimated high-density direction. The training-free method works with pretrained architectures without external guidance models or additional training. On FLUX.2-Klein-9B, DreamSim rises from 0.1964 to 0.2643, while GenEval falls from 0.8813 to 0.8525 and latency is 1.32× baseline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.