Reproduce “Forward-Value Discretization: Faster Stochastic Diffusion Sampling Without High-Order Approximation” 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
Yuchen Jiao, Na Li, Changxiao Cai, Yuting Wei, Gen Li
F-DDPM is a training-free stochastic diffusion sampler that changes where the pretrained model is evaluated. It uses a lookahead estimate of a cleaner state, then performs the stochastic update with noise, requiring one model evaluation per iteration. On unconditional CIFAR-10 with EDM, it achieves FID 8.63 at 9 evaluations, outperforming tested stochastic baselines at that budget in the reported comparison, but not establishing universal speed or quality advantages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.