Reproduce “UltraDiff: Transferring High-Fidelity Priors to Compressed Latent Spaces for High-Resolution Image Generation” 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
Jingjing Ren, Haitian Zheng, Haoyu Chen, Connelly Barnes, Zhifei Zhang, Yuqian Zhou, Zhe Lin, Lei Zhu
UltraDiff trains an f32 latent diffusion student using f8 guidance through MixTeacher, which shares attention across paired representations and transfers the teacher’s f32 denoising direction. At 512 × 512, GenEval is 0.72, versus 0.61 after f8 initialization and 0.73 for the f8 anchor. With 20 steps on one A100, 2048 × 2048 generation takes 6.2 seconds versus 249.8; 2K–4K adaptation uses flow matching alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.