Reproduce “Learning Fast Samplers for Diffusion Models by Differentiating Through Sample Quality” 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
Daniel Watson, William Chan, Jonathan Ho, Mohammad Norouzi
A new method, Differentiable Diffusion Sampler Search (DDSS), optimizes fast samplers for pre-trained diffusion models by directly optimizing perceptual quality metrics. This approach substantially improved few-step sample quality, achieving an FID of 13.77 for 5-step generation on CIFAR10, outperforming DDIM's 32.66.
Key insights
- Direct optimization of perceptual metrics like KID, rather than log-likelihood, is crucial for achieving high visual sample quality with few diffusion steps.
- Relaxing the "matching marginals" constraint, a design principle in DDIM, allows for discovering more effective sampling paths and leads to superior perceptual quality.
- Optimizing sampler parameters in a perceptually meaningful feature space (e.g., InceptionV3 features for KID) is essential; direct pixel-space optimization yields poor results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.