Parallel Decoding Distillation for Fast Image and Video Generation
PDD speeds diffusion generation by having each network pass predict several denoising steps.
The paper presents Parallel Decoding Distillation as a trajectory-based alternative to VSD and adversarial distillation losses. The authors say it works with pre-trained diffusion and flow-matching models and can run at variable NFE counts. In their reported tests, PDD reaches state-of-the-art results at 4-8 NFE on LTX-2.3, Wan 14B, and Qwen-Image. They also report a notable gain in video diversity.
ArXiv · AI/CL/LG's note
The paper presents Parallel Decoding Distillation as a trajectory-based alternative to VSD and adversarial distillation losses. The authors say it works with pre-trained diffusion and flow-matching models and can run at variable NFE counts. In their reported tests, PDD reaches state-of-the-art results at 4-8 NFE on LTX-2.3, Wan 14B, and Qwen-Image. They also report a notable gain in video diversity.
ArXiv · AI/CL/LG's note
score 5