Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation
Authors: Shanchuan Lin, Ceyuan Yang, Hao He, Jianwen Jiang, Yuxi Ren, Xin Xia, Yang Zhao, Xuefeng Xiao, Lu Jiang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments demonstrate that our 8B model achieves real-time, 24fps, streaming video generation at 736 416 resolution on a single H100, or 1280 720 on 8 H100 up to a minute long (1440 frames). |
| Researcher Affiliation | Collaboration | Shanchuan Lin: Corresponding author: EMAIL Hao He: The Chinese University of Hong Kong. Internship at Byte Dance Seed. Jianwen Jiang: Byte Dance Intelligent Creation Lab. |
| Pseudocode | No | The paper describes methods and architectures but does not include any explicitly labeled pseudocode or algorithm blocks with structured formatting. |
| Open Source Code | No | The model, data, and code used in this paper is proprietary and not currently scheduled for release. However, we describe our methodology in details for reproducibility. |
| Open Datasets | Yes | We evaluate our method on the standard VBench I2V benchmark [32] on both 120-frame short-video generation and 1440-frame long-video generation. |
| Dataset Splits | Yes | We evaluate our method on the standard VBench I2V benchmark [32] on both 120-frame short-video generation and 1440-frame long-video generation. |
| Hardware Specification | Yes | Our 8B model achieves real-time, 24fps, streaming video generation at 736 416 resolution on a single H100, or 1280 720 on 8 H100 up to a minute long (1440 frames). |
| Software Dependencies | Yes | We implement block causal attention using Flash Attention 3 [79] in a for-loop. |
| Experiment Setup | Yes | We use Adam W optimizer [57] with a learning rate of 1e-5 and a weight decay scale of 0.01 throughout the process. We first train on 736 416 (equivalent to 640 480 by area) 5-second videos for 20k iterations with a batch size of 256. |