BiWM: Open-source bidirectional video world models cut training from 4 to 2 stages
Two-stage training on 8 H200s beats causal models in fidelity and control.
A team of researchers from China has introduced BiWM (Bidirectional World Models), the first full-stack open-source framework for interactive video world models under the bidirectional autoregressive paradigm. Unlike previous causal pipelines that require four stages—control fine-tuning, autoregressive training, causal initialization, and few-step distillation—BiWM collapses the process into just two stages: a fine-tuning step for camera control injection, followed by a few-step Distribution Matching Distillation (DMD) stage. The entire pipeline converges in a few hundred training steps on eight H200 GPUs, dramatically lowering the barrier to entry for high-fidelity interactive video generation.
BiWM supports a wide range of pretrained backbones—Wan2.1-1.3B, Wan2.2-5B, HunyuanVideo-1.5-8B, and LTX-2.3-22B—and works with existing bidirectional video diffusion models. It introduces real-world camera control (where prior work like minWM loses controllability), pluggable history compression (FramePack and PackForcing style) for long rollouts, and an optional NVFP4 4-bit training/inference pipeline. To counter the mode-seeking degradation typical of DMD, the framework adds GAN and mass-covering forward-KL losses that preserve scene dynamics. The authors have open-sourced BiWM to push forward resource-constrained research and high-fidelity environment simulation.
- BiWM reduces training stages from 4 (minWM) to 2, converging in a few hundred steps on 8x H200 GPUs.
- Supports 4 backbone sizes (1.3B to 22B) including Wan, HunyuanVideo, and LTX, plus fine-tuning of existing bidirectional models.
- Adds real-world camera control, pluggable history compression for long rollouts, and a 4-bit NVFP4 pipeline.
Why It Matters
Makes high-quality, interactive video world models accessible to resource-constrained labs and accelerates simulation research.