Uniform Discrete Diffusion with Metric Path for Video Generation

Arxiv 2025
Haoge Deng 1,3,5*,   Ting Pan 2,3,5*,   Fan Zhang 5*,   Yang Liu 4,5*,   Zhuoyan Luo 5,   Yufeng Cui 5,   Wenxuan Wang 5,   Chunhua Shen 4,   Shiguang Shan 2,3,   Zhaoxiang Zhang 1,3†,   Xinlong Wang 5†‡

1 National Laboratory of Pattern Recognition, CASIA

,  

2 Key Laboratory of Intelligent Information Processing, ICT, CAS

,  

3 University of Chinese Academy of Sciences

,  

4 Zhejiang University

,

5 Beijing Academy of Artificial Intelligence

,
* Equal contribution, † Corresponding author, ‡ Project leader,
Arxiv Paper Code

Abstract

Continuous-space video generation has advanced rapidly, while discrete approaches lag behind due to error accumulation and long-context inconsistency. In this work, we revisit discrete generative modeling and present Uniform discRete diffuSion with metric pAth (URSA), a simple yet powerful framework that bridges the gap with continuous approaches for the scalable video generation. At its core, URSA formulates the video generation task as an iterative global refinement of discrete spatiotemporal tokens. It integrates two key designs: a Linearized Metric Path and a Resolution-dependent Timestep Shifting mechanism. These designs enable URSA to scale efficiently to high-resolution image synthesis and long-duration video generation, while requiring significantly fewer inference steps. Additionally, we introduce an asynchronous temporal fine-tuning strategy that unifies versatile tasks within a single model, including interpolation and image-to-video generation. Extensive experiments on challenging video and image generation benchmarks demonstrate that URSA consistently outperforms existing discrete methods and achieves performance comparable to state-of-the-art continuous diffusion methods.

Method

Illustration of different image/video generation paradigms. Discrete-space approaches such as AR and MDM adopt non-refinable local generation, where produced tokens are fixed once generated. In contrast, URSA introduces iterative global refinement, conceptually aligning discrete methods with continuous-space approaches, and substantially narrowing their performance gap.

Global refinement via token distance in embedding space. Starting from categorical noise x0 (left), our framework refines data based on token distance to get target data x1 (right), enabling hierarchical structure generation from global semantics to fine details.

Quantitative Results

URSA rivals Sora-like text-to-video generation models despite using a discrete video tokenizer.

URSA emerges frame-conditioned video generation by accurately modeling the future motion.

URSA performs on par with the state-of-the-art models in generating high-resolution images.

Visual Results

Text-to-video Visualization

Image-to-video Visualization

Autoregressive video generation Visualization

Autoregressive video generation Visualization with First-End Frame Conditioning

BibTeX


        @article{deng2025ursa,
            title={Uniform Discrete Diffusion with Metric Path for Video Generation},
            author={Deng, Haoge and Pan, Ting and Zhang, Fan and Liu, Yang and Luo, Zhuoyan and Cui, Yufeng and Shen, Chunhua and Shan, Shiguang and Zhang, Zhaoxiang and Wang, Xinlong},
            journal={arXiv preprint arXiv:2510.24717},
            year={2025}
        }