Read the original at HF Daily Papers
Researchers introduce a correction step for diffusion models that reduces the training-inference gap by refining intermediate denoising predictions to prevent early artifacts from persisting in final samples.
Carried by: HF Daily Papers. First seen: .