BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression

A new open-source framework called BiWM aims to improve the interactivity of video world models by adopting a bidirectional autoregressive approach, a departure from existing causal pipelines that require multiple training stages and often suffer from quality degradation due to error accumulation [1][2]. The framework, detailed in a paper submitted to arXiv on June 8, 2026, is described as the first full-stack system for interactive video world models under the bidirectional autoregressive paradigm, jointly optimizing generation quality and inference speed [1][2]. While recent models such as Yume-1.5 and Matrix-Game-3.0 have demonstrated the fidelity and stable long-horizon rollout benefits of a bidirectional autoregressive approach, open-source frameworks like minWM have only supported causal models [2]. BiWM streamlines the training process into two stages, down from the four required by minWM. From a pretrained video backbone, the framework injects camera control through fine-tuning, followed by a few-step Distribution Matching Distillation stage that converts the backbone into an action- and camera-controllable world model, converging in a few hundred steps on 8xH200 GPUs [2]. The framework supports a range of model backbones, including Wan2.1-1.3B, Wan2.2-5B, HunyuanVideo-1.5-8B, and LTX-2.3-22B, and also allows for secondary fine-tuning of existing bidirectional models [2]. BiWM enables real-world camera control in scenarios where minWM loses controllability and integrates pluggable history compression for long rollouts [2]. To address the mode-seeking degradation common in Distribution Matching Distillation, the framework incorporates GAN and mass-covering forward-KL objectives to preserve scene dynamics [2]. An optional NVFP4 4-bit training and inference pipeline is also offered [2]. The authors have open-sourced BiWM to support resource-constrained research and high-fidelity environment simulation [1][2].