VideoNeuMat — weights

Model weights for VideoNeuMat: Neural Material Extraction from Generative Video Models (SIGGRAPH 2026).

• Project page: https://bowenxueai.github.io/VideoNeuMat/
• Code: https://github.com/bowenxueai/videoneumatcode

The pipeline turns a text prompt (or image) into an 81-frame 1024×1024 material video with a fine-tuned Wan-2.1 video model, then a feed-forward LRM extracts a re-renderable neural material (BRDF + displacement) from that video.

Files

path	what	base model
wan14b/step-10000.safetensors	T2V material generator — full fine-tuned DiT (27 GB)	Wan2.1-T2V-14B
wan14b_t2v_lora/step-9000.safetensors	T2V material generator — LoRA (293 MB)	Wan2.1-T2V-14B
wan14b_i2v/step-9500.safetensors	I2V material generator — full fine-tuned DiT (31 GB)	Wan2.1-I2V-14B
lrm/latent_module.pth	LRM encoder — material → latent (feed-forward)	—
lrm/mlp.pth	shared neural-material MLP decoder	—

All three generators produce the same 81-pose sparse-rig material video that the LRM consumes. The base Wan-2.1 weights (Wan-AI/Wan2.1-T2V-14B, Wan-AI/Wan2.1-I2V-14B) are downloaded separately from the official Wan-AI repos — see the code repo's README and release/download_weights.sh.

The generator weights are fine-tunes of Wan-2.1 (Apache-2.0). The VideoNeuMat code is MIT-licensed.

Citation

@inproceedings{xue2026videoneumat,
  author    = {Xue, Bowen and Hadadan, Saeed and Zeng, Zheng and Rousselle, Fabrice and Montazeri, Zahra and Hasan, Milos},
  title     = {VideoNeuMat: Neural Material Extraction from Generative Video Models},
  booktitle = {ACM SIGGRAPH 2026 Conference Papers},
  year      = {2026},
}