Parallel Decoding Distillation for Fast Image and Video Generation
PDD speeds diffusion generation by predicting several denoising steps in one network pass.
The paper frames Parallel Decoding Distillation as a simpler alternative to VSD and adversarial-loss distillation, which it says can be hard to optimize and prone to mode collapse. The method is designed to work with pre-trained diffusion and flow matching models and allows different inference budgets. The authors report state-of-the-art results at 4-8 function evaluations on LTX-2.3, Wan 14B, and Qwen-Image, with stronger video diversity. HF Daily Papers' note
The paper frames Parallel Decoding Distillation as a simpler alternative to VSD and adversarial-loss distillation, which it says can be hard to optimize and prone to mode collapse. The method is designed to work with pre-trained diffusion and flow matching models and allows different inference budgets. The authors report state-of-the-art results at 4-8 function evaluations on LTX-2.3, Wan 14B, and Qwen-Image, with stronger video diversity. HF Daily Papers' note
score 5