SUPRA-Tex

SUPRA-Tex3D Generation with Super-Aligned Texture

Built entirely on
open-source data.

Commercial comparisons Real2Sim Gallery

AUTHORS & RESOURCES

  1. 1Huawei Inc.
  2. 2Shanghai Jiao Tong University
  3. 3Huazhong University of Science and Technology

* Equal contribution† Project lead Corresponding authors

arXiv GitHub Models HF Demo
Take flightCompare with MeshyReturn to image
Start with a reference image.
01 / COMMERCIAL COMPARISONS

Competitive with
commercial models.

Trained entirely on open-source data.Compare alignment and fine detail in 3D.

COMPARE WITH
SUPRA-TexMeshy 7.1 Drag to rotate · Scroll to zoom

Commercial comparisons

SUPRA-Tex OursMeshy 7.1

Original reference

Fighter

Original fighter reference image
02 / HUMAN PREFERENCE

Preferred,
head to head.

Preferred for visual quality in our five comparisons.

45real & synthetic inputs
512paired evaluations
View full study
Win rate · ties excluded
03 3D TEXTURING

From image
to textured 3D.

Explore reference-guided textures, novel views and relighting across diverse objects.

Reference images · Textured results · Relighting
04 REAL2SIM

Real2Sim.

Explore a 3D scene built from one image, with simulated robot interaction and relighting.

05 REFERENCE ALIGNMENT

Even the small print.

Compare lettering, symbols and fine markings.

Reference image
Reference image of an industrial switch with a wiring table and caution label
Rows and symbols in the wiring table.
ReferenceSUPRA-Tex
Reference detail: wiring table, circuit symbols, and caution text
SUPRA-Tex rendered detail of the same wiring table, circuit symbols, and caution text
Drag to compare

One reference image · matched camera · different lighting.

06 RELIGHTABLE MATERIALS

Hold the shape.
Change the light.

Explore seven models under changing light.
Inspect their albedo, roughness and metallic maps.

LIVE 3D
Loading 3D…
REFERENCEAmber bird brooch reference image

Amber bird brooch

Drag to rotate · Shift + scroll to zoom

One reference image · provided geometry · PBR materials

Loading 3D model…
07 IMAGE-BASED CONTROL

Make it yours.

Choose a back-view design, then inspect the corresponding result from three angles.

FRONT REFERENCEFIXED
Fixed front reference: a blue-and-white porcelain plate with a floral pattern

The front reference and geometry stay fixed.

CHOOSE A BACK DESIGN
GENERATED RESULTBack view
Crane and peonies on the generated plate, back view
Crane & peonies

More views. More control. Use additional views or edited images without retraining the 3D generator.

LIVE 3D / BACK DESIGN STUDY

Crane & peonies

The supplied meshes use reprojected albedo with interactive PBR lighting.
Reference image
SUPRA-Tex result
09 THE RESEARCH

Observe first.
Generate in 3D.

Estimate materials from the reference, align them to the mesh, then complete the texture in 3D.

01ESTIMATE
Estimated albedo of the reference mug

Separate material from light.

Use 2D generative priors to estimate albedo, metallic and roughness maps from the reference image.

02ALIGN
Observed materials projected onto the visible mug surface, with unobserved regions shown in green

Anchor to the surface.

Project reliable observations onto the mesh, then encode them in the generator's native material latent space.

03COMPLETE
The paper's illustration of completed material latents for the mug

Complete in 3D.

An interleaved flow transformer alternates unobserved-region and full-surface updates, guided by fixed observation memory.

10 MEASURED FIDELITY

Same geometry.
Higher fidelity.

SUPRA-Tex achieves lower perceptual error than Pixal3D* on both benchmarks, with geometry held fixed within each comparison.

Compared on the same mesh Pixal3D* SUPRA-Tex

LPIPS measures perceptual error · lower is better
01 / GIVEN GEOMETRY

Isolate the texture.

One image. The same ground-truth mesh.
16 matched views · SUPRA-Tex‡ (Qwen-Image-2.1).

DTC-Tex100 assets
0.0500.042
16.0%lower error
Complex-Tex110 assets
0.0460.038
17.4%lower error

Closer agreement on object pixels.

Foreground PSNR ↑
DTC-Tex
+3.20 dB
Complex-Tex
+1.85 dB
02 / PREDICTED GEOMETRY

Lower error on predicted meshes.

One image. A shared predicted mesh.
End-to-end LPIPS evaluated across 4 views.

DTC-Tex100 assets
0.0710.025
64.8%lower error
Complex-Tex110 assets
0.0700.029
58.6%lower error

Controlled comparison.

The shared mesh isolates differences in texture generation.

* Pixal3D is the open-source TRELLIS.2-based variant used in this study. Given-geometry highlights use SUPRA-Tex‡ with Qwen-Image-2.1 for 2D material estimation.

Explore the evidence.

Given geometry12 configurations · 12 metrics

Each method textures the ground-truth mesh from one reference image. Scores average the reference view and 15 novel views under matched lighting and cameras. The reference camera is supplied where the method interface supports it.

Each cell: DTC-Tex / Complex-Tex. PSNR is in dB. ↑ Higher is better; ↓ lower is better. — No PBR output. ‡ Qwen-Image-2.1 material estimation.

Given-geometry texture generation — full benchmark results
MethodPSNR ↑Foreground PSNR ↑SSIM ↑LPIPS ↓CLIP-I ↑FDDINOv3 ↓Albedo PSNR ↑Albedo foreground PSNR ↑Albedo SSIM ↑Albedo LPIPS ↓Metallic MAE ↓Roughness MAE ↓
Step1X-3D23.12 / 23.6114.21 / 13.970.914 / 0.9240.083 / 0.0660.851 / 0.84833.30 / 30.93— / —— / —— / —— / —— / —— / —
MV-Adapter24.64 / 24.5915.73 / 14.950.917 / 0.9260.077 / 0.0640.858 / 0.84933.17 / 31.19— / —— / —— / —— / —— / —— / —
Ink3D25.82 / 26.1517.08 / 16.690.921 / 0.9350.076 / 0.0600.859 / 0.86435.01 / 30.86— / —— / —— / —— / —— / —— / —
UniTEX25.42 / 25.7816.54 / 16.190.925 / 0.9400.071 / 0.0560.877 / 0.87229.24 / 27.42— / —— / —— / —— / —— / —— / —
MVPainter26.44 / 25.7217.59 / 16.190.924 / 0.9330.077 / 0.0670.856 / 0.84633.40 / 33.8724.17 / 23.0515.30 / 13.510.911 / 0.9220.089 / 0.0790.121 / 0.2280.119 / 0.188
TEXGen27.24 / 27.9018.33 / 18.250.932 / 0.9450.072 / 0.0530.848 / 0.87151.11 / 32.89— / —— / —— / —— / —— / —— / —
LumiTex28.29 / 27.7719.38 / 18.210.939 / 0.9470.062 / 0.0510.892 / 0.88724.42 / 23.2926.30 / 25.2617.39 / 15.680.928 / 0.9380.069 / 0.0600.352 / 0.2560.098 / 0.181
Hunyuan3D 2.126.80 / 25.8317.89 / 16.190.926 / 0.9310.070 / 0.0630.877 / 0.85728.18 / 31.0524.00 / 22.8915.09 / 13.250.910 / 0.9180.079 / 0.0740.142 / 0.2470.079 / 0.156
TRELLIS.228.60 / 27.9719.73 / 18.410.935 / 0.9450.062 / 0.0510.880 / 0.87632.73 / 27.4225.78 / 25.7716.90 / 16.170.921 / 0.9350.070 / 0.0580.063 / 0.1310.059 / 0.111
Pixal3D*29.19 / 28.8420.47 / 19.440.950 / 0.9500.050 / 0.0460.908 / 0.89920.35 / 19.5026.39 / 26.6217.60 / 17.250.937 / 0.9350.057 / 0.0560.051 / 0.1450.076 / 0.122
SUPRA-Tex30.89 / 30.0522.34 / 20.640.961 / 0.9600.042 / 0.0390.927 / 0.91516.51 / 17.3628.20 / 27.6119.55 / 18.120.951 / 0.9520.048 / 0.0460.043 / 0.1090.041 / 0.084
SUPRA-Tex‡32.08 / 30.6623.67 / 21.290.962 / 0.9610.042 / 0.0380.928 / 0.92116.31 / 16.0829.68 / 27.7421.14 / 18.270.955 / 0.9530.048 / 0.0470.033 / 0.1150.038 / 0.076

PSNR, SSIM and LPIPS use the full rendered image unless marked foreground; foreground PSNR uses the GT object mask. Metallic and roughness MAE use the intersection of predicted and GT masks. Scores average over views and valid objects. Pixal3D* denotes the open-source TRELLIS.2-based variant used in this study. SUPRA-Tex uses FireRed-Image-Edit-1.1 for 2D material estimation; SUPRA-Tex‡ uses Qwen-Image-2.1.

End-to-end generation7 methods · 9 metrics

One reference image, with no ground-truth geometry or camera during generation. SUPRA-Tex and Pixal3D* share the same predicted geometry. Reconstruction scores use four views, CLIP-I uses the reference view, and distribution scores compare 16-view rendered sets.

Each cell: DTC-Tex / Complex-Tex. PSNR is in dB. † Similarity-aligned to ground truth for scoring only.

End-to-end single-image generation — full benchmark results
MethodPSNR ↑Foreground PSNR ↑SSIM ↑LPIPS ↓CLIP-I ↑FDDINOv3 ↓FDCLIP ↓FID ↓CMMD ↓
Step1X-3D19.70 / 20.6612.00 / 12.360.894 / 0.9000.115 / 0.0990.844 / 0.88233.4 / 26.319.6 / 17.933.5 / 31.414.40 / 8.73
Hunyuan3D 2.121.97 / 21.8714.52 / 13.830.904 / 0.9080.095 / 0.0870.835 / 0.84334.8 / 31.222.7 / 23.233.6 / 35.512.78 / 11.39
UniLat3D†19.78 / 20.4111.89 / 11.890.900 / 0.9020.105 / 0.0960.822 / 0.83034.2 / 29.827.2 / 27.536.1 / 35.227.17 / 21.57
CUPID†20.26 / 20.6812.38 / 12.120.906 / 0.9050.093 / 0.0900.850 / 0.85232.1 / 28.125.1 / 24.433.0 / 34.124.18 / 18.44
TRELLIS.2†22.59 / 22.2215.71 / 14.320.907 / 0.9090.098 / 0.0870.816 / 0.83041.7 / 32.626.2 / 23.432.2 / 32.016.63 / 10.49
Pixal3D*26.88 / 26.2419.90 / 18.640.945 / 0.9400.071 / 0.0700.910 / 0.90927.2 / 22.916.5 / 15.926.0 / 25.88.29 / 5.96
SUPRA-Tex28.65 / 28.7320.79 / 20.240.967 / 0.9630.025 / 0.0290.971 / 0.96821.1 / 19.49.9 / 9.323.0 / 23.34.97 / 4.00
11 OPEN SOURCE

Open by design.

Built on data and algorithms from the open-source community, SUPRA-Tex achieves results comparable to commercial systems in input alignment and visual quality. We are open-sourcing SUPRA-Tex to give back to the community that made this work possible.

If this work helps or inspires you,

Star on GitHub

Citation

@article{yang2026supratex,
  title   = {{SUPRA-Tex}: 3D Generation with Super-Aligned Texture},
  author  = {Yang, Chen and Fang, Jiemin and Meng, Fanpeng and Huang, Zihao
             and Pang, Jichen and Wu, Guanjun and Bao, Zhikuan and Tian, Qi},
  journal = {arXiv preprint arXiv:2610.22352},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.22352}
}

SUPRA-Tex — Aircraft demonstration