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

BiWM introduces the first open-source framework for bidirectional autoregressive video world models, reducing training complexity from four stages to two while maintaining generation quality. The framework supports multiple model architectures and enables real-world camera control with improved long-horizon rollouts through self-correcting error propagation.

BiWM addresses a critical gap in open-source video world model infrastructure by implementing bidirectional autoregression, a paradigm that recent proprietary models like Yume-1.5 and Matrix-Game-3.0 have proven superior to traditional causal approaches. The framework's significance lies in democratizing advanced world model development—previously, researchers faced a stark choice between using limited open-source tools built for causal models or relying on proprietary systems. The bidirectional approach achieves better error correction through self-referential prediction, enabling more stable long-horizon rollouts essential for interactive environments.

The technical innovation reduces implementation overhead substantially. Traditional causal pipelines require control fine-tuning, autoregressive training, causal initialization, and few-step distillation—four distinct stages prone to error accumulation. BiWM collapses this to two stages: camera control injection and Distribution Matching Distillation, converging in hundreds of steps on accessible hardware. The framework's universality across model sizes (1.3B to 22B parameters) and architectures demonstrates robust generalization.

For the AI research community, this release lowers barriers to advanced world model research by 50% in training complexity while improving output fidelity. Institutions with limited compute resources can now develop interactive video models previously accessible only to well-funded labs. The integration of pluggable history compression and optional 4-bit quantization further extends accessibility. The addition of GAN and forward-KL objectives specifically addresses mode-seeking degradation inherent in distribution matching, a sophisticated solution to a known problem in the field. Open-sourcing this work establishes a new foundation layer for embodied AI and robotics simulation research.

• →BiWM reduces video world model training from four stages to two, accelerating development cycles by approximately 50%
• →First open-source bidirectional autoregressive framework enables real-world camera control that causal approaches cannot achieve
• →Supports model scaling from 1.3B to 22B parameters with unified training recipe across architectures
• →Integrated history compression and 4-bit quantization make advanced world models accessible to resource-constrained researchers
• →GAN and forward-KL objectives preserve scene dynamics while countering mode-seeking degradation in distribution matching