Parallel Decoding Distillation for Fast Image and Video Generation
2026-07-28 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionMachine Learning
AI summaryⓘ
The authors explain that creating videos with diffusion or flow models takes a long time because the process repeats many steps. They propose a new method called Parallel Decoding Distillation (PDD) that speeds up video generation by predicting several steps at once, making it faster and easier to use with existing models. Unlike previous methods, PDD avoids complicated calculations and improves the variety and motion in the generated videos. Their approach also works well across different video and image generation tasks with fewer steps needed.
Diffusion modelsFlow modelsVideo generationSampling stepsVariational score distillationAdversarial lossesMode collapseParallel Decoding DistillationFunction evaluationsText-to-video
Authors
Neta Shaul, Chao Liu, Arash Vahdat, Julius Berner
Abstract
Generation in video diffusion or flow models is computationally expensive due to the slow and iterative sampling process. Current state-of-the-art (SOTA) acceleration methods heavily rely on variational score distillation (VSD) and adversarial losses to distill diffusion models into few-step generators. Albeit achieving high-quality video generation, these training losses are notoriously hard to optimize and suffer from mode collapse, leading to loss of video diversity and lack of motion. In this paper, we introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method for fast inference of diffusion and flow matching models. Our architecture and training procedure are compatible with any pre-trained model and support sampling with a varying number of function evaluations (NFE). PDD accelerates generation by predicting multiple denoising steps per network evaluation. Conceptually, it learns a representation of the mean velocity without regressing its derivative using JVPs or finite-difference approximations. Our method achieves SOTA performance with 4-8 NFE on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image. Moreover, PDD presents a significant improvement in generated video diversity.