---
title: "Parallel Decoding Distillation for Fast Image and Video Generation"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.26004v1"
permalink: "https://feed.astralor.com/items/itm_1ae47a9cc16625ee55333ee5"
published: "2026-07-28T17:20:00.000Z"
score: 78
channel: "ai"
tags: ["LTX-2.3","Qwen-Image","Wan 14B","Parallel Decoding Distillation","Diffusion Models","Video Generation"]
---

# Parallel Decoding Distillation for Fast Image and Video Generation

## AI commentary

> This paper's PDD method is clever: bypassing complex distillation losses via multi-step prediction, it boosts efficiency and diversity while maintaining quality. For researchers and engineers working on video generation acceleration, this is one of the most noteworthy advances—it could replace the existing VSD paradigm and make few-step generation truly practical.

## At a glance

This paper introduces Parallel Decoding Distillation (PDD), a trajectory-based distillation method for fast inference of diffusion and flow matching models. PDD predicts multiple denoising steps per network evaluation, achieving SOTA performance with 4-8 NFE while avoiding mode collapse from traditional VSD and adversarial losses, improving video diversity. Validated on LTX-2.3, Wan 14B, and Qwen-Image models.

## Key points

- PDD predicts multiple denoising steps per network forward pass, reducing inference to 4-8 steps.
- No VSD or adversarial losses needed, training is simpler and avoids mode collapse.
- Achieves SOTA on LTX-2.3 text-to-video/audio, Wan 14B text-to-video, and Qwen-Image text-to-image.
- Significant improvement in generated video diversity and motion quality.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_1ae47a9cc16625ee55333ee5