Data-Forcing Distillation: Restoring Diversity and Fidelity in Few-Step Video Generation

Siyi Chen1 Shaowei Liu2,3 Yixuan Jia1 Zian Wang2 Huan Ling2 Qing Qu1 Jun Gao1,2
1University of Michigan  ·  2NVIDIA  ·  3University of Illinois Urbana-Champaign
One line of code
DFD is a single-line change to the DMD2 training loop — no GAN needed.
100–300 steps of finetuning
A short post-training stage restores diversity and fidelity.
Improve on 3 video generation tasks
Text-to-video (Wan2.1), image-to-video (Cosmos), and autoregressive generation (Self Forcing).

TL;DR We introduce real-data supervision into the DMD loss, removing the over-saturation and limited-diversity problems of DMD. With only a single line of code change, we improve on three video generation tasks.

Overview

A 60-second silent walkthrough of the method and qualitative results.

DFD Restores Diversity

Image-to-Video Results

Autoregressive Video Generation Results

Text-to-Video Results

Abstract

Method

Ablation Studies

Quantitative Results

Limitation

BibTeX

@misc{dfd2026,
  title  = {Data-Forcing Distillation: Restoring Diversity and Fidelity
            in Few-Step Video Generation},
  author = {Chen, Siyi and Liu, Shaowei and Jia, Yixuan and Wang, Zian
            and Ling, Huan and Qu, Qing and Gao, Jun},
  year   = {2026}
}