Sprint3D student weights

Sprint3D is a fast image-to-3D asset pipeline built on the publicly released 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: Sprint3D: Accelerating End-to-End 3D Asset Generation through Pipeline Optimization (arXiv link to follow)

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:

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

@article{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},
  journal = {arXiv preprint},
  year    = {2026}
}