BiWM advances open-source interactive video world models with bidirectional autoregression

Arxiv CS.CV (computer vision)

BiWM transitions bidirectional video diffusion models into an autoregressive paradigm, improving interactivity of video world models. It eliminates multiple stages needed by existing causal pipelines, such as control fine-tuning and causal initialization.

See 5 more sources

What Makes Video World Model Latents Action-Relevant: Prediction over Reconstruction5 days agoJewon Yeom, Hanseul Kim, Jeongjae Park, Sungmok Jung, Jaejin Lee, Taesup Kim

BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression3 days agoShaohao Rui, Xiaofeng Mao, Zhanyu Zhang, Peijia Lin, Yansong Zhu, Yibo Zhang, Haibin Wan, Weijie Ma

TBD-VLA: Temporal Block Diffusion Vision Language Action Model5 days agoSung-Wook Lee, Xuhui Kang, Yen-Ling Kuo

DisCo: World Models with Discrete Camera Motion Control5 days agoHongrui Huang, Junke Wang, Quanhao Li, Yu-Gang Jiang, Zuxuan Wu

Emergent Semantic Representations in World Models through Physical Interaction without Linguistic Supervision16 days agoJiayi Fang

Arxiv CS.CV (computer vision)·Yifan XuChao ZhangRuifei MaFei GaoZhifei YangJiaxing QiZhipeng Chen

4 days ago