Texture Translation of PBR Materials Based on Pix2pix-Turbo

Jiamu Liu

2024

Abstract

Physically Based Rendering (PBR), a high-quality method in 3D model rendering, is widely used in modern games and 3D short films. However, generating corresponding PBR textures is relatively complex and challenging. This paper proposes a new task called PBR texture translation. The task involves generating corresponding texture maps such as height, normal, and roughness maps based on the base color image of a given PBR texture using an image-to-image translation model. Additionally, this paper improves the latest image translation model, pix2pix-turbo, by incorporating a classifier and expert models, and specifically adjusting the text-image alignment via a Text Prompt through experiments. After training on the MatSynth dataset, the model achieved a Minimum Mean Squared Error (MSE) of 1181.41 and a maximum Structural Similarity Index (SSIM) of 0.614 on the height texture of the test set, reducing MSE by 1,443.53 and improving SSIM by 0.13 compared to the original model. The contributions of this research include proposing the PBR texture translation task and improving the pix2pix-turbo model to make it more suitable for texture translation tasks.

Paper Citation

in Harvard Style

Liu J. (2024). Texture Translation of PBR Materials Based on Pix2pix-Turbo. In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-754-2, SciTePress, pages 322-328. DOI: 10.5220/0013516500004619

in Bibtex Style

@conference{daml24,
author={Jiamu Liu},
title={Texture Translation of PBR Materials Based on Pix2pix-Turbo},
booktitle={Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2024},
pages={322-328},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013516500004619},
isbn={978-989-758-754-2},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML
TI - Texture Translation of PBR Materials Based on Pix2pix-Turbo
SN - 978-989-758-754-2
AU - Liu J.
PY - 2024
SP - 322
EP - 328
DO - 10.5220/0013516500004619
PB - SciTePress