BiWM is an open-source minimalist training framework for video world models, whose core positioning is to transform a pretrained bidirectional video diffusion model into an action-controllable autoregressive world model. According to the materials, it is the first such open-source training framework.
In terms of purpose, BiWM is aimed at the training and inference of bidirectional autoregressive video world models, supports three types of conditional generation, text-to-video (t2v), image-to-video (i2v), and video-to-video (v2v), and achieves them within the same model.
Main capabilities and workflow: the framework completes the transformation through two training stages. The first stage is camera-control fine-tuning, using 81 discrete camera actions; the second stage is few-step DMD distillation, used for chunk-by-chunk autoregressive generation. Unlike causal attention pipelines, BiWM preserves full bidirectional attention within the current chunk and its history. Camera actions are computed as action_label by translation * 9 + rotation, thereby producing 81 combined actions. In terms of action injection, the action path injects the per-latent-frame discrete actions into AdaLN, while keeping camera descriptions out of the captions.
Supported backbone models: Wan2.1 / Wan2.2, as well as HunyuanVideo-1.5 / LTX-2.3. The materials note that Wan and HY15/LTX23 should use separate environments. Among them, HY15 uses Qwen2.5-VL, and LTX23 uses Gemma-3; for HY15's dedicated environment configuration and troubleshooting, see hunyuan/ENV_SETUP.md. Wan2.1 uses the corresponding scripts under scripts/wan21/. For HunyuanVideo-1.5 discrete-action training, the base weights need to be placed under ckpts/HunyuanVideo-1.5 before running.
Usage: the materials provide instructions for dataset download and the expected directory structure; each JSON record contains a scene-only caption and action_frames. All scripts can be run from any working directory, automatically resolve the repository root, and allow paths to be overridden via environment variables. Common overridable environment variables include BIWM_VIDEO_DIR, BIWM_CAPTION_JSON, OUTPUT_DIR, PYTHON_BIN, MAX_TRAIN_STEPS, and CHECKPOINTING_STEPS. The materials also mention Wan distillationmodel inference. The specific installation and run commands are not fully given in the materials and are not yet provided.
The project is built on top of FastVideo, minWM, Wan, HunyuanVideo-1.5, and LTX-Video, and is released under the Apache License 2.0.
Limitations: the materials explicitly require separate environments for different backbones; details such as hardware requirements, training duration, VRAM requirements, and complete installation steps are not yet provided.
