SUPRA-Tex ('3D Generation with Super-Aligned Texture') is an open-source model from Huawei, Shanghai Jiao Tong University, and Huazhong University of Science and Technology that turns a single image into a fully textured 3D asset. Unlike typical image-to-3D models, it preserves fine surface details like lettering, patterns, and material boundaries, and outputs a relightable GLB with base color, metallic, and roughness (PBR) textures.

How is SUPRA-Tex different from other image-to-3D tools?

Most image-to-3D generators produce a single baked-in color texture that smears or hallucinates fine details and can't be relit. SUPRA-Tex instead predicts true PBR materials in the reference view, encodes them as surface-aligned 3D observations, and completes the unseen regions with an 'Interleaved Flow Transformer.' The result retains the input image's details (text stays readable) and renders correctly under new lighting.

What can you actually do with SUPRA-Tex?

Two modes: image-to-3D (generate a textured asset from one photo) and mesh texturing (texture an existing mesh with matching camera and material observations). There's also an optional dual-view mode that generates a back view so the unseen side also follows an image, and a --refine pass for very small text and intricate textures.

What hardware does SUPRA-Tex need?

Linux with an NVIDIA GPU and CUDA 12.4. The full image-to-3D pipeline has been tested on an RTX 3090 (24 GB), where CPU offload is enabled automatically; 48 GB+ GPUs are faster. The models are on Hugging Face (2.1B completion model + a 33.6M-param material adapter), and the code is MIT-licensed, though the material adapter uses Qwen's non-commercial research license.

SUPRA-Tex: Huawei Turns One Photo Into a Relightable 3D Asset That Keeps the Lettering

By Prahlad Menon Published 2026-10-21 3 min read

Most “photo to 3D” tools fall apart on the thing that matters most: the surface. Point one at a photo of a labeled switch panel or a branded water bottle and you get a plausible shape wrapped in a blurry, smeared texture — the lettering turns to mush, the metal doesn’t look like metal, and you can’t relight it without it falling apart.

SUPRA-Tex, from Huawei with SJTU and HUST, is built specifically to fix the texture problem. Its name says it: Super-Aligned Texture.

What It Does

SUPRA-Tex generates a textured 3D asset from a single image while preserving the details that usually get destroyed — lettering, patterns, and material boundaries. The pipeline:

Predicts PBR materials in the reference view — not just a flat color, but base color, metallic, and roughness maps.
Encodes them as surface-aligned 3D observations — the visible pixels become ground-truth texture on the geometry, locked to where they actually belong.
Completes the unseen regions with an Interleaved Flow Transformer — the back and sides you never saw get materials that match.

The output is a relightable GLB with proper PBR channels, so it behaves correctly under new lighting in any engine — Blender, Unreal, three.js, whatever.

The comparison that’s making the rounds (vs. Meshy 7) tells the story: on an image of a model jet with “Supra-Tex” printed on the fuselage, the competitor’s text is unreadable while SUPRA-Tex keeps it crisp and surface-locked.

Two Modes, Plus a Clever Back-View Trick

Image-to-3D — one photo in, textured GLB out.
Mesh texturing — already have geometry? Give it a mesh, camera, a foreground image, and reference-view material maps, and it completes the PBR textures directly.

The genuinely interesting part is dual-view texturing. Single-view inference only knows what the front shows; everything else is completion (i.e., educated guessing). Dual-view adds a generated back observation so the unseen side also follows a real image. The back view is painted by a foundation image model (their implementation uses GPT-6 Astra via the Codex CLI), reviewed, then fused with the front — front taking priority where they overlap. Because only that one stage is external, the whole thing gets better automatically as image models improve.

Why It Matters

This is the same lesson we keep running into: in generative 3D, the geometry is often the easy part and the texture is where quality lives. We’ve covered turning phone video into 3D digital twins and Kiln’s self-correcting text-to-3D loop; SUPRA-Tex attacks the specific failure mode that makes most of these unusable for real products — unreadable text and fake-looking materials.

For anyone doing e-commerce assets, game props, product viz, or AR, “the label stays legible and the asset relights properly” is the difference between a demo and a deliverable. And it runs on a single 24 GB card.

It’s open-source (MIT, with the material adapter under Qwen’s non-commercial research license), the models are on Hugging Face, and there’s a live demo to try before you install anything.

Repo: github.com/Supra-3D/SUPRA-Tex · Project page: supra-3d.github.io/supratex · Demo: HF Space · Paper: arXiv:2610.22352