Forward-Value Discretization: Faster Stochastic Diffusion Sampling Without High-Order Approximation
F-DDPM achieved lower FID than the tested stochastic baselines at matched model-evaluation budgets across four image datasets.
The authors prove that higher-order approximations do not generally improve stochastic samplers’ convergence order, even with exact score functions. Their training-free method instead evaluates the model at a cleaner lookahead point, reusing the previous prediction and injected noise. It retains first-order convergence with one model evaluation per iteration. Across CIFAR-10, ImageNet, FFHQ and LSUN, it delivered competitive sample quality against deterministic baselines using 1.5 times as many evaluations.
ArXiv · AI/CL/LG's note
The authors prove that higher-order approximations do not generally improve stochastic samplers’ convergence order, even with exact score functions. Their training-free method instead evaluates the model at a cleaner lookahead point, reusing the previous prediction and injected noise. It retains first-order convergence with one model evaluation per iteration. Across CIFAR-10, ImageNet, FFHQ and LSUN, it delivered competitive sample quality against deterministic baselines using 1.5 times as many evaluations.
ArXiv · AI/CL/LG's note
score 5