DMD

One-step Diffusion with Distribution Matching Distillation

Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, Taesung Park · CVPR 2024

Read paper · arXiv:2311.18828

Use the score difference between real and generated distributions to construct a distribution-matching generator update.

What you will implement

Implement the score-difference target and paired regression term; inspect mode coverage using an analytic bimodal distribution.

• 19. DMD: Score-Difference Distillation

  Implement the distribution-matching generator gradient, paired regression, and separate fake-score update.

  Implement score_difference, dmd_loss, fake_score_loss. toy.py provides the training and visualization scaffold.