New arXiv Paper Proposes 'Correction Training' to Fix Exposure Bias in Diffusion Models

It is the mismatch between training and inference: the denoiser is trained on real images interpolated with noise, but at generation time it receives inputs constructed from its own imperfect earlier predictions, a state it never saw during training, which can cause errors to accumulate.

CFG steers generation toward a condition by blending two predictions, while correction explicitly learns to fix the base model's prediction against the ground-truth target. The paper reports correction achieves performance competitive with CFG, and combining both improves diversity and reduces oversaturation.

Yes, according to the paper, the refined predictions can be distilled back into the base diffusion model, retaining quality gains without additional inference-time overhead.

This is research, not a production release, but it is relevant to any team working on image, video, or visual asset generation. The practical step is reproducing the results on your own models before making adoption decisions, especially for unguided generation.