Generating high-quality video with diffusion or flow models is computationally intensive due to the slow, iterative sampling process. Existing acceleration methods often rely on variational score distillation (VSD) and adversarial losses, which are difficult to optimize and can lead to mode collapse, reducing video diversity and motion quality.
In a new paper, Hugging Face researchers introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method designed for fast inference in diffusion and flow matching models. PDD is compatible with any pre-trained model and supports sampling with a variable number of function evaluations (NFE).
The key innovation is predicting multiple denoising steps per network evaluation, effectively learning a representation of the mean velocity without needing to regress its derivative via Jacobian-vector products (JVPs) or finite-difference approximations. This approach accelerates generation while maintaining high quality.
PDD achieves state-of-the-art performance with 4-8 NFE on several large models, including LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image. Importantly, it also significantly improves generated video diversity compared to existing methods.