Reproduce “Correction Training for Diffusion Models” 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
Donghoon Ahn, Runqian Wang, Alexei A. Efros, Taesung Park, Yash Belhe
The authors introduce a corrector for diffusion image generation that refines the denoiser’s clean-image prediction before each sampling update. Trained on the denoiser’s own predictions, the corrector can be applied repeatedly. On class-conditional ImageNet-256 with DiT-XL/2 at 250 network evaluations, it lowers FID-50K from 9.58 unguided to 2.07, versus 2.51 for CFG, with lower precision but higher recall than CFG.
Key insights
- Repeated correction often improves generation even though it can increase reconstruction error on inputs resembling the training distribution, suggesting that training-distribution reconstruction error alone does not predict sampling quality.
- Corrected intermediate states become harder to distinguish from training-distribution states in classifier two-sample tests; this is consistent with reducing the training–inference gap but does not prove that mechanism.
- Higher fidelity can come with lower recall, and too many correction iterations may reduce quality or impose overly regular structure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.