Does Native 3D Texture Generation Necessarily Require 3D Assets for Training?
Tex-Zero trains native 3D texture models from constructed 2D image planes, without real 3D assets.
Tex-Zero shows that native 3D texture generation can be trained without real 3D asset datasets. High-quality 2D images are turned into training samples by placing each image on a 3D plane and applying patch-wise random rotations and aggregation. A Tex-Zero VAE trained only on this constructed data reconstructs real 3D assets with high quality, and a Tex-Zero DiT built on it generates fine-grained textures conditioned on multi-view images. The work argues that detailed color, not scanned geometry, is the essential training signal.
- Tex-Zero trains native 3D texture generation without real 3D assets.
- 2D images become planes with patch-wise rotations and aggregation.
- Tex-Zero VAE reconstructs real 3D assets never seen in training.
- Tex-Zero DiT is also trained only on constructed image data.
- AI models
- Tex-Zero VAETex-Zero DiT
From the original article
1 min read at huggingface.co
Native 3D texture generation synthesizes colors directly in 3D space for a given geometry, conditioned on multi-view reference images. It is generally believed that training such models requires large-scale, high-quality real 3D asset data, whose acquisition remains a long-standing and challenging problem.
By Jiangshan Wang, Zeqiang Lai, Jiayi Guo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.