Parallel Decoding Distillation for Fast Image and Video Generation

Researchers introduced Parallel Decoding Distillation, a simpler training method that speeds up diffusion and flow-based image and video generation.

The method predicts multiple denoising steps per evaluation, enabling faster sampling with pretrained models.

It reports state-of-the-art results at 4 to 8 network evaluations on several text-to-image and text-to-video systems.

The approach also reduces mode collapse and improves motion diversity, making fast generation more practical for deployment.