VideoNeuMat: Neural Material Extraction from Generative Video Models
VideoNeuMat tries to extract standalone material assets from generative video models instead of just using the videos themselves. The authors fine-tune Wan 2.1 14B to produce controlled material sample videos, essentially a virtual gonioreflectometer that keeps the model’s realism while imposing a measurement pattern.
They then train a Large Reconstruction Model, finetuned from a smaller Wan 1.3B backbone, to reconstruct compact neural materials from those videos. The paper says 17 generated frames are enough for single-pass inference, and the resulting materials generalize to novel lighting and viewing conditions. For game devs, the appeal is obvious: if this holds up beyond the paper, it could reduce the need for expensive capture rigs or heavily manual material authoring, while giving technical artists and graphics programmers another route to high-diversity PBR-like assets.
“effectively creating a "virtual gonioreflectometer"”
- what
- VideoNeuMat extracts reusable neural material assets from generative video models in a 2-stage pipeline.
- who
- Authors: Bowen Xue, Saeed Hadadan, Zheng Zeng, Fabrice Rousselle, Zahra Montazeri, and Milos Hasan.
- when
- Submitted Feb 6, 2026; revised May 15, 2026 (v2).
- impact
- Could reduce dependence on manual material capture and help generate diverse, reusable materials for rendering pipelines.
Promising workflow for material creation and asset reuse
Follow graphics updates
See relevant stories in your personalized news feed.
Discussion