PanDWM: Panoramic 4D Driving World Models

Checkpoints for PanDWM, a panoramic driving world model that generates 360° driving video in four denoising steps per window and lifts the generated frames into 3D and 4D Gaussians.

Paper (coming soon) | Website (coming soon) | Code

Files

File	Component	Size	SHA-256
video/pandwm_video_v84i4250_ema.pth	Four-step autoregressive video student (EMA weights, training step 4250)	13 GB	4def33264008d987fc3a9053a7a3f88d917cd0b467158f6dd109517e359d46cd
gaussian/pandwm_gs_decoder_G3b_2000.pth	Feed-forward Gaussian decoder (3D and 4D Gaussians)	220 MB	6450e980772adc46e2303960be06a26ab98f24805df2547af5a9ef63ec403da1
refiner/pandwm_render_refiner_v1_4000.pth	Render refiner for laterally shifted views	13.6 GB	824fde5807c5060c1443cabd23a5823d65723c63541be3d9c4ddb606efb3b25d

All files are PyTorch state dicts saved with torch.save. Verify them after download with sha256sum -c SHA256SUMS.

Required third-party weights (not included)

Weights	Used for	Source
Stable Diffusion 3.5 Medium	VAE and text encoders of the video model and the refiner	stabilityai/stable-diffusion-3.5-medium
Depth Anything 3 (DA3-LARGE-1.1, DA3METRIC-LARGE)	Range relabelling of the Gaussians	depth-anything/DA3-LARGE-1.1, depth-anything/DA3METRIC-LARGE, code: Depth-Anything-3
I3D (i3d_pretrained_400.pt)	FVD	TATS

Inception (FID), VGG (LPIPS) and the other evaluation backbones are downloaded by torch/torchmetrics on first use.

Usage

See the code repository for environment setup, data preparation and the scripts that reproduce each table.

License

The video model and the render refiner are derived from Stable Diffusion 3.5 Medium and are distributed under the Stability AI Community License. Powered by Stability AI.

Third-party weights listed above are not redistributed here and remain under their own licenses. Training and evaluation use the nuScenes dataset under its terms of use.

Citation

@article{sun2026pandwm,
  title={PanDWM: Panoramic 4D Driving World Models},
  author={Sun, Jiajun and Zhang, Zeyu and Ren, Zeyu and Jiang, Guangfeng and Zhao, Yang and Tang, Hao},
  year={2026}
}