DMD2

Improved Distribution Matching Distillation for Fast Image Synthesis

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

Read paper · arXiv:2405.14867

Remove DMD's paired-regression requirement using faster fake-score updates and adversarial auxiliary training.

What you will implement

Implement two-timescale updates and auxiliary objectives; verify parameter-update boundaries in a small experiment.

• 20. DMD2: Two-Time-Scale and Adversarial Updates

  Implement the critic-to-generator update ratio and non-saturating GAN contribution on top of distribution matching.

  Implement generator_due, adversarial_losses, generator_loss. toy.py provides the training and visualization scaffold.