Correction training reduces diffusion models' training-inference gap
Original title: Correction Training for Diffusion Models
A corrector refines intermediate predictions and can be distilled back into the base model.
The paper adds a correction step for diffusion models to reduce exposure bias, where inference inputs differ from training inputs. The corrector takes the base denoiser's prediction and is trained against the same ground-truth target, then iteratively refines clean-sample predictions during sampling. Evaluated on DiT, SiT, and JiT, the method improves unguided and unconditional FID and is competitive with classifier-free guidance. Combining correction with CFG improved diversity and reduced CFG oversaturation. Distilling refined predictions back into the base model keeps quality gains without extra inference cost.
Unconfirmed: The authors say exposure bias might be one reason diffusion models struggle without extra help such as CFG.
- The method was evaluated on DiT, SiT, and JiT.
- It improves previously reported unguided and unconditional FID scores.
- Performance is competitive with classifier-free guidance.
- Correction plus CFG reduced typical CFG oversaturation.
- Distillation removes the extra inference-time cost.
From the original article
1 min read at huggingface.co
Iterative sampling in diffusion models can lead to error accumulation due to exposure bias. The denoiser is trained on real images interpolated with noise, but at inference its inputs are constructed from its own imperfect predictions from the earlier steps, something it has never seen at training time. This might be one reason diffusion models often struggle without additional help, such as CFG.
By Donghoon Ahn, Runqian Wang, Alexei A. Efros · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.