Reproduce “Lipschitz Singularities in Diffusion Models” 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
Zhantao Yang, Ruili Feng, Han Zhang, Yujun Shen, Kai Zhu, Lianghua Huang, Yifei Zhang, Yu Liu and 3 more
Researchers identified and addressed "Lipschitz singularities" in diffusion models, a phenomenon where the noise prediction network's Lipschitz constant explodes near the zero-time point. Their proposed Early Timestep-Shared Diffusion Model (E-TSDM) improves image generation quality, lowering FID on FFHQ 256x256 from 9.50 to 6.62, and significantly enhances the efficiency of fast sampling algorithms.
Key insights
- Lipschitz singularities arise because the derivative of common noise schedules (`dσt/dt`) tends to infinity as `t → 0`, causing the partial derivative of the noise prediction network with respect to time to also explode.
- By sharing timestep conditions in a specific early interval, E-TSDM effectively forces the Lipschitz constant within those sub-intervals to zero, leading to an overall smoother and more stable network.
- A theoretical error analysis confirms that the approximation introduced by timestep sharing is bounded and converges to zero, ensuring its validity without significant fidelity loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.