MatSynth: A Modern PBR Materials Dataset


SurfaceNet basecolor/metallic




Quantitative evaluation

We evaluate performances of the SurfaceNet model on the basecolor/metallic workflow, to set a new baseline for future work shifting to this workflow.
Similarly to the diffuse/specular counterpart, the model is trained on the MatSynth dataset for 250,000 steps (approx. 56 hours), with batch size of 8 and learning rate of 2e-5. We enable the adversarial discriminator after 100,000 steps.
Evaluation is conducted in terms of RMSE for all maps (except the Normal map for wich we compute the cosine distance). Additionally we compute the LPIPS/SSIM metrics for the rendering and basecolor which can be treated as natural images.
Rendering metrics are computed on 4 different lighting conditions (see the paper for details), while using the remaining as input to the model.

Metric Rendering Basecolor Normal Height Roughness Metallic
RMSE ↓ 0.091 0.090 0.250 0.135 0.106
Cos dist ↓ 0.543
LPIPS ↓ 0.275 0.273
SSIM ↑ 0.649 0.653


Qualitative evaluation

Qualitative results of the SurfaceNet model on the basecolor/metallic workflow, show that the model is able to correctly estimate the metallic component of a material.
The model is additionally able to estimate the height (normalized between 0 and 1) which can be used to displace the mesh when rendering.

Input Rendering Basecolor Normal Height Roughness Metallic
GT
SurfaceNet
GT
SurfaceNet
GT
SurfaceNet
GT
SurfaceNet