SUPRA-TEX: 3D Generation with Super-Aligned Texture (arXiv 2610.22352)

Forum topic · 小凯 · 2026-10-17

Summary

SUPRA-Tex is a new method for single-image 3D generation that dresses a given 3D mesh with relightable PBR materials while faithfully preserving the visual details of a reference image. The paper, authored by Chen Yang, Jiemin Fang, Fanpeng Meng, Zihao Huang, Jichen Pang, Guanjun Wu, Zhikuan Bao, and Qi Tian (arXiv 2610.22352), addresses two common failure modes in prior work: artifacts from fusing inconsistent multi-view predictions and lost fine detail from implicit image-feature conditioning. SUPRA-Tex decouples the problem into visible-material estimation, surface alignment, and observation-guided 3D completion. An MLLM-based image editing model predicts lighting-separated PBR maps in the reference view, which are projected onto visible surfaces and encoded as observation latents sharing the generator's latent space. An interleaved flow Transformer treats these latents as fixed read-only anchors, alternating between unobserved-region and full-surface updates to produce complete material latents. Trained only on public datasets, SUPRA-Tex achieves state-of-the-art results in both geometry-conditioned texturing and end-to-end asset generation, matching or exceeding leading commercial systems. Project page: https://supra-3d.github.io/supratex/

Paper Overview

Field: Computer Vision (CV)
Authors: Chen Yang, Jiemin Fang, Fanpeng Meng, Zihao Huang, Jichen Pang, Guanjun Wu, Zhikuan Bao, Qi Tian
arXiv: 2610.22352
Project page: https://supra-3d.github.io/supratex/

Introduction

Single image 3D generation aims to dress a given 3D mesh with relightable materials while faithfully preserving the visual details of a reference image. Existing methods either introduce artifacts when fusing inconsistent multi-view predictions or struggle to preserve fine details through implicit image-feature conditioning.

Method

To resolve these challenges, SUPRA-Tex is built on the principle of decoupling visible-material estimation, surface alignment, and observation-guided 3D completion. This division of labor allows the 3D generator — trained on limited 3D data — to focus its capacity on preserving visible evidence and completing unobserved surfaces.

1. Train an MLLM-based image editing model to predict lighting-separated PBR maps in the reference view. 2. Project these maps onto visible surfaces and encode them as observation latents that share the latent space of the generation target. 3. An interleaved flow Transformer uses these latents as fixed read-only anchors, alternating between unobserved-region updates and full-surface updates to generate complete material latents.

Results

Trained only on public datasets, SUPRA-Tex achieves state-of-the-art performance in both:

Geometry-conditioned texture generation
End-to-end 3D asset generation

It matches or exceeds leading commercial systems.