# Preview Model Weights This directory contains the local weights required by the Preview Model. ## Weight Mapping | Component | Local path | Description | | --- | --- | --- | | Video generation model (VGen) | `checkpoints/diffusion_pytorch_model.bin` | Main video diffusion model. | | Action Expert | `checkpoints/action_model/diffusion_pytorch_model.bin` | Action prediction model. | | VAE | `vae/diffusion_pytorch_model.safetensors` | Video latent encoder and decoder. | | VAE latent statistics | `vae/latent_stats.pth` | Latent normalization statistics used by the VAE. | | Qwen3.5 VLM | [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) | Vision-language model used for instruction and image conditioning. | The `config.json` files next to each model define the corresponding model architecture. ## Expected Layout ```text preview_model/ ├── checkpoints/ │ ├── config.json │ ├── diffusion_pytorch_model.bin │ └── action_model/ │ ├── config.json │ └── diffusion_pytorch_model.bin ├── vae/ │ ├── config.json │ ├── diffusion_pytorch_model.safetensors │ └── latent_stats.pth └── stats.json ``` `stats.json` contains the state and action normalization statistics used during policy inference. ## Runtime Paths Use `preview_model/checkpoints` as the `CKPT` directory, use `preview_model/vae` as the configured VAE path, and set the VAE statistics path to `preview_model/vae/latent_stats.pth`. Download Qwen3.5-2B from the link above and set `vlm_model_id` to its local directory or directly to `Qwen/Qwen3.5-2B` when Hugging Face downloads are available.