Reproduce “Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation” 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
Bowen Zheng, Tianming Yang
Researchers from the Chinese Academy of Sciences propose a "generative pre-training" paradigm for diffusion models, enabling efficient one-step image generation by fine-tuning a pre-trained diffusion U-Net with only a Generative Adversarial Network objective. This approach drastically reduces training data requirements by 1-2 orders of magnitude compared to traditional distillation, achieving an FID of 1.54 on CIFAR-10 and showing strong generalization across various datasets.
Key insights
- Diffusion model training implicitly acts as "generative pre-training," embedding universal generative capabilities into the model's parameters.
- This latent generative knowledge can be efficiently "unlocked" for one-step generation by fine-tuning with only a Generative Adversarial Network (GAN) objective, eliminating the need for instance-level distillation losses.
- The authors identify an "inconsistent local minima problem" in traditional distillation, where multi-step and single-step models converge to different optimal states, hindering effective knowledge transfer via direct output mimicry.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.