Reproduce “Test-Time Iterative Error Correction for Efficient 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
Yunshan Zhong, Weiqi Yan, Yuxin Zhang
IEC refines image generation from an efficient diffusion model without retraining, by evaluating the model again on provisional outputs at selected denoising steps. This aims to limit errors that can compound across steps, at the cost of extra computation. On CIFAR-10, W8A8 sampling with 100 DDIM steps improves FID from 4.32 to 3.88 with 20% time overhead; this result covers one tested setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.