SIGGRAPH Asia 2026 Conference Papers

Uncertainty DMD

Restoring Diversity in Few-Step Autoregressive Video Distillation

Zixuan Duan*, Xunzhi Xiang*, Yabo Chen*, Xin Zhang, Changhan Liu, Haibin Huang, Chi Zhang, Qi Fan, Xuelong Li

* Equal contribution    Corresponding authors

Nanjing University  ·  TeleAI  ·  Fudan University

Recovering per-prompt diversity through structured uncertainty, while preserving comparable visual quality.

arXiv Explore the results ↓

Video results

Diversity that persists over time

Each panel shows nine generations using the same prompt seeds across all methods.

Prompt 01 / 09

“Time lapse of sunrise on Mars”

Uncertainty DMD Ours
DMD Baseline
DMD + GAN Baseline

Method

Uncertainty generation within DMD training

Uncertainty DMD adds two lightweight perturbations to the autoregressive distillation pipeline. The model architecture and the original DMD objective remain unchanged.

Uncertainty DMD method overview showing time-step uncertainty, stochastic cache writing, and DMD training

01

Time-step uncertainty

Perturbed denoising timesteps produce different first chunks from different noise seeds, restoring variation at the start of generation.

02

Stochastic cache writing

The cache is perturbed before each append operation so distinct generation trajectories continue across subsequent chunks.

Paper

Abstract

Few-step Distribution Matching Distillation enables efficient autoregressive video generation, but often maps different noise seeds to highly similar first chunks. The deterministic autoregressive cache then propagates this collapsed state, reducing sample diversity and motion dynamics.

We introduce Uncertainty DMD, a lightweight framework that injects structured uncertainty through first-chunk timestep perturbation and stochastic cache writing. The same perturbations apply during training and inference, require no architectural modifications, and recover diverse generation trajectories while preserving comparable visual quality.

Reference

Citation

@misc{duan2026uncertaintydmdrestoringdiversity,
  title     = {Uncertainty DMD: Restoring Diversity in Few-Step
               Autoregressive Video Distillation},
  author={Zixuan Duan and Xunzhi Xiang and Yabo Chen and Xin Zhang and Changhan Liu and Haibin Huang and Chi Zhang and Qi Fan and Xuelong Li},
      year={2026},
      eprint={2609.11265},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.11265}, 
}
Nanjing UniversityNanjing, China
TeleAIInstitute of Artificial Intelligence, China Telecom
Fudan UniversityShanghai, China