NVIDIA’s PDD speeds image and video generation with 4–8-step distillation

nvidia · hf · 2026-07-29

NVIDIA’s paper proposes Parallel Decoding Distillation (PDD), a trajectory-based distillation method to speed up diffusion and flow-matching models for image and video generation.
The method is designed to be simpler and more scalable than VSD/adversarial-loss-based approaches, which are hard to optimize and can collapse diversity.
PDD predicts multiple denoising steps per network evaluation, supports a variable number of function evaluations (NFE), and works with any pre-trained model.
The authors report SOTA results at 4–8 NFE on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image, with better video diversity as well.

More from Multimodal