Alaya Lab · Technical report · 2026
Sprint3D: Accelerating End-to-End 3D Asset Generation through Pipeline Optimization
From one image to a textured 3D asset in 4.05 s,
33× faster than the released Pixal3D.
Image to textured GLB · one H200 GPU · median of 77 inputs · split by mean stage times
Base (Pixal3D)
133 s
Ours
4.05 s33× FASTER
NetworksMesh post-processingInput and other
OutputsAssets generated by our pipeline, each from one image.
INTERACTIVE · WEBGL
The 3D viewer starts when this section is on screen.
Drag to rotate · scroll or pinch to zoom · right-drag to pan
Web copies: 100k faces, 1024² textures. Exported assets: about 1M faces, 4096² textures.
Single-image generationInput-view normal maps against the ground truth, following the Pixal3D paper.
TOYS4K TEST · 2,653 OBJECTS
At the same export settings our pipeline is 33× faster and on par with or better than the released Pixal3D on the normal metrics, with a slightly lower silhouette IoU.
| Method | Silhouette IoU(%) ↑ | Normal PSNR(dB, overlap region) ↑ | Mean angular error(°) ↓ | Chamfer distance(×10⁻³) ↓ | F-score@0.01 ↑ | ULIP-2↑ | Uni3D↑ | Time per image(s) ↓ |
|---|---|---|---|---|---|---|---|---|
| TRELLIS.2 | 79.55 | 17.15 | 25.06 | 28.01 | 0.523 | 0.424 | 0.394 | – |
| TripoSG | 59.87 | 14.22 | 38.87 | 58.62 | 0.239 | 0.397 | 0.373 | – |
| Hunyuan3D-2.1 | 59.82 | 14.31 | 39.33 | 59.94 | 0.264 | 0.442 | 0.403 | – |
| Base (Pixal3D) | 96.97 | 20.87 | 14.25 | 26.61 | 0.392 | 0.423 | 0.391 | 133.0 |
| Ours | 96.78 | 21.56 | 13.00 | 25.55 | 0.408 | 0.415 | 0.385 | 4.05 |
Chamfer distance and F-score use each method's own frame, without alignment. Ours is slightly lower on ULIP-2 and Uni3D, which the report traces to thicker inner walls of the remeshed shells.
Secondary comparison · 400 validation objects
Validation set
These objects also served model selection. The same picture holds, with a slightly lower mean silhouette IoU from six failure objects.
| Method | Silhouette IoU(%) ↑ | Normal PSNR(dB, overlap region) ↑ | Mean angular error(°) ↓ | Chamfer distance(×10⁻³) ↓ | F-score@0.01 ↑ | ULIP-2↑ | Uni3D↑ | Time per image(s) ↓ |
|---|---|---|---|---|---|---|---|---|
| TRELLIS.2 | 71.88 | 15.80 | 28.92 | 34.44 | 0.414 | 0.404 | 0.367 | – |
| TripoSG | 60.35 | 13.39 | 40.45 | 57.33 | 0.207 | 0.361 | 0.333 | – |
| Hunyuan3D-2.1 | 64.90 | 14.47 | 34.56 | 42.28 | 0.286 | 0.418 | 0.375 | – |
| Base (Pixal3D) | 96.60 | 19.20 | 17.22 | 38.50 | 0.314 | 0.406 | 0.367 | 133.0 |
| Ours | 96.33 | 19.82 | 15.92 | 37.24 | 0.332 | 0.395 | 0.359 | 4.05 |
BibTeXTechnical report, 2026. arXiv identifier to follow.
@misc{kuno2026sprint3d,
title = {Sprint3D: Accelerating End-to-End 3D Asset Generation through Pipeline Optimization},
author = {Kuno, Sho and Yu, Ruihan and Yu, Yonghao and
Lin, Guixu and Fu, Lian and Niu, Muyao and
Tsai, Yu-Ju and Huang, Zheng-Hui and Huang, Yi-Chuan and
Yang, Siqi and Lan, Fengbo and Wu, Erwin and
Zhang, Kaipeng and Wang, Zhixiang},
year = {2026},
note = {Alaya Lab technical report}
}