Reproduce “Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance” 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
Aidan Gleich, Scott C. Schmidler
Researchers from Duke University developed SMC-MLMC, a training-free diffusion guidance method that uses Sequential Monte Carlo with Multilevel Monte Carlo to achieve unbiased and computationally efficient conditional generation. This approach attained 95.6% Top-1 accuracy on CIFAR-10 and demonstrated a 3x lower cost-per-success than prior methods, establishing new benchmarks for accuracy and efficiency.
Key insights
- Replacing heuristic point-estimates with unbiased Monte Carlo estimators for marginal likelihood within an SMC framework enables asymptotically exact conditional sampling.
- Multilevel Monte Carlo is crucial for making theoretically unbiased likelihood estimation computationally feasible in complex diffusion models, overcoming the 'rare-event estimation' challenge.
- The SMC-MLMC reweighting mechanism can rigorously correct the bias introduced by existing training-free guidance heuristics when they are used as initial proposal distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.