Correction Training to Mitigate Exposure Bias in Diffusion Models
October 14, 2026
Correction training reduces the training-inference gap in diffusion models by training a corrector to refine imperfect predictions from the base denoiser. This iterative refinement at intermediate denoising steps prevents early artifacts from accumulating during the sampling process.
HOW THIS AFFECTS YOU
●
builderYou can improve the visual fidelity of diffusion-based pipelines by implementing iterative correction steps.
●
researcherYou can address exposure bias in diffusion models by explicitly training for error correction during sampling.