File size: 4,166 Bytes
ff5c1e6 69cb29c ff5c1e6 69cb29c ff5c1e6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 | ---
license: cc-by-nc-4.0
base_model: TencentARC/Pixal3D
pipeline_tag: image-to-3d
tags:
- image-to-3d
- 3d-generation
- mesh
- lora
- distillation
- pixal3d
---
# Sprint3D student weights
Sprint3D is a fast image-to-3D asset pipeline built on the publicly released
[Pixal3D](https://github.com/TencentARC/Pixal3D) pipeline. It rebuilds the mesh post-processing on
the GPU and replaces the four rectified flows of Pixal3D by distilled *students*, cutting the median
time per asset from 133 s to 4.05 s on one H200 GPU at the same export settings (33×).
This repository holds only the students: four LoRA adapters (EMA weights) on top of the Pixal3D flow
models, plus `students.json`, which records for each student its base flow model, the LoRA layout and
the sampling schedule. The Pixal3D base model, the DINOv3 encoder and the other upstream models are
not redistributed here; the code fetches them from their original repositories at pinned revisions.
- Code: https://github.com/AlayaLab/sprint3d
- Project page: https://alaya-lab.github.io/sprint3d/
- Technical report: [arXiv:2611.16701](https://arxiv.org/abs/2611.16701), *Sprint3D: Accelerating End-to-End 3D Asset Generation through Pipeline Optimization*
## Files
| File | Student | Steps | Size | sha256 |
|---|---|---|---|---|
| `structure_student.safetensors` | sparse structure flow (`ss_flow_img_dit_1_3B_64_bf16`) | 4 | 1,274,085,008 B | `a938130271c96eaaf67643bb428aa2aa8e53033b09608ea3eedc5531d9b97396` |
| `coarse_shape_student.safetensors` | coarse shape flow (`slat_flow_img2shape_dit_1_3B_512_bf16`) | 1 | 1,305,542,328 B | `59be306522816b4d1defa525cfecf3f9666a161902b245fee387cd7a34fb9a66` |
| `fine_shape_student.safetensors` | fine shape flow (`slat_flow_img2shape_dit_1_3B_1024_bf16`) | 1 | 1,305,542,328 B | `6beb37c82f5f618f84d9bc988564d41e3971ff85c42ef06217e580186e7dd583` |
| `texture_student.safetensors` | texture flow (`slat_flow_imgshape2tex_dit_1_3B_1024_bf16`) | 2 | 1,305,542,328 B | `ecc4ae096358120cde9ef2a6bbdff8bfee5c04e5f3bc49db84183d0f7d37887c` |
| `students.json` | metadata: base model revision, LoRA layout, schedules, hashes | | 6,362 B | `310db80728d8f0ff5c95ae73ce43bb3c544a0560d7812ba339e099683cd311e5` |
Each adapter has rank 256 (alpha 256) on the attention and MLP linears of all 30 transformer blocks
(240 adapters, 480 tensors), stored in float32. The students run 4 + 1 + 1 + 2 = 8 network
evaluations in place of the 76 of the released schedules. Base model: `TencentARC/Pixal3D` at
revision `0b31f9160aa400719af409098bff7936a932f726` (MIT).
## Use
The intended way to use these files is through the code repository, which downloads every weight at
its pinned revision and verifies sizes and sha256 values:
```bash
git clone --recursive https://github.com/AlayaLab/sprint3d.git
cd sprint3d
python scripts/download_weights.py # students + upstream models, about 32.6 GB
python scripts/verify_weights.py
```
`students.json` documents the key convention (`backbone.model.<path>.lora_A.weight` /
`lora_B.weight`, update `(alpha / rank) * lora_B @ lora_A`) and the fold rule, so the adapters can
also be applied to the Pixal3D flow models without this code.
## License
The student weights are released under **CC BY-NC 4.0**. They were trained on subsets of the
TRELLIS-500K object lists; parts of that data (HSSD and ABO under CC BY-NC, 3D-FUTURE for research
use only) do not allow commercial use, and the weights inherit that restriction. The code is
Apache-2.0; the Pixal3D base model is MIT and is fetched separately. Sprint3D is not affiliated
with, endorsed by or reviewed by Tencent or the authors of Pixal3D.
## Citation
```bibtex
@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},
eprint = {2611.16701},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}
```
|