Reproduce “Sublinear iterations can suffice even for DDPMs” 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
Matthew S. Zhang, Stephen Huan, Jerry Huang, Nicholas M. Boffi, Sitan Chen, Sinho Chewi
The authors introduce DDRaM, a randomized-midpoint sampler for stochastic DDPM reverse processes. It queries a score at a randomly chosen intermediate point in each step, then completes the update. Under score-accuracy, bounded-second-moment, and time-varying smoothness assumptions, its step bound scales with the square root of dimension at fixed parameters. On AFHQv2, it improves image-quality metrics over two baselines at score-evaluation budgets.
Key insights
- Randomized midpoint methods, previously effective in log-concave sampling, can be successfully adapted to SDE-based diffusion models to achieve debiasing and superior accuracy.
- The 'shifted composition rule' framework provides a robust analytical tool for translating weak and strong error bounds into KL divergence guarantees for complex, randomized SDE discretizations.
- Sublinear complexity for DDPMs is achievable without altering the core reverse process, making theoretical improvements directly relevant to practical implementations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.