Input Perturbation Reduces Exposure Bias in Diffusion Models

Method and contribution

Studies accumulated sampling errors in diffusion models and trains with perturbed inputs to simulate inference-time prediction errors.

Addresses the same training–sampling mismatch through input perturbation, providing a comparison with scheduled sampling on model-generated states.

Read in context

Learning to render LaTeX, SMILES, HTML and music notation →

Citation

Mang Ning, Enver Sangineto, Angelo Porrello, Simone Calderara, Rita Cucchiara. Input Perturbation Reduces Exposure Bias in Diffusion Models. ICML 2023.

@inproceedings{pmlr-v202-ning23a,
  title = {{Input Perturbation Reduces Exposure Bias in Diffusion Models}},
  author = {Ning, Mang and Sangineto, Enver and Porrello, Angelo and Calderara, Simone and Cucchiara, Rita},
  year = {2023},
  booktitle = {ICML 2023},
  url = {https://proceedings.mlr.press/v202/ning23a.html}
}