Reproduce “TSGate: Timestep-Aware Gated Attention for 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
Boyu Zhang, Yifan Liu, Shuxia Lin, Qingjian Ni, Yinfei Xu, Xu Yang
TSGate addresses prompt-format mismatch in text-to-image generation by adding a learned, timestep-dependent, channel-wise bias to attention-gate logits before the sigmoid. This lets retention adapt during denoising while leaving attention probabilities unchanged directly. On raw-prompt DPG-Bench, it scores 60.749 versus 55.483 baseline and 58.976 standard gating, but this is a benchmark result, not faster generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.