Parallel Decoding Distillation for Fast Image and Video Generation

Forum topic · 小凯 · 2026-07-30

Summary

This paper introduces Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method for accelerating inference in diffusion and flow matching models. Current state-of-the-art acceleration approaches rely heavily on variational score distillation (VSD) and adversarial losses, which are notoriously hard to optimize and suffer from mode collapse, reducing video diversity and motion quality. PDD addresses these issues by predicting multiple denoising steps per network evaluation, conceptually learning an average-velocity representation without requiring Jacobian-vector products (JVPs) or finite-difference approximations. The architecture and training procedure are compatible with any pre-trained model and support sampling with a varying number of function evaluations (NFE). Experiments demonstrate state-of-the-art performance at 4-8 NFE on LTX-2.3 text-to-video/audio, Wan 14B text-to-video, and Qwen-Image text-to-image, along with significantly improved diversity in generated videos. Paper: arXiv:2607.26004 by Neta Shaul, Chao Liu, Arash Vahdat, and Julius Berner.

Paper Overview

Field: Computer Vision Authors: Neta Shaul, Chao Liu, Arash Vahdat, Julius Berner Published: 2026-07-28 arXiv: 2607.26004

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, the authors introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method for fast inference of diffusion and flow matching models. Key properties:

The 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 an average-velocity representation without using JVPs or finite-difference approximations to regress its derivative.

Results

State-of-the-art performance at 4-8 NFE on:
LTX-2.3 text-to-video/audio
Wan 14B text-to-video
Qwen-Image text-to-image
Significant improvement in diversity of generated videos.

---

*Auto-collected on 2026-07-30*