Alchemy3D-1M

Alchemy3D-1M

A million-scale dataset for versatile 3D asset editing.

Alchemy3D-1M is the training corpus for Alchemy3D: source–target 3D editing pairs on PBR-ready assets, with text instructions and image conditions. It contains 1.25M unique assets and 1.38M editing pairs across 7 edit types.

Due to Hugging Face storage limits, this repository includes only training metadata, structured latents at resolution 512, and rendered conditions. For GLBs and latents at resolution 1024, please use the ModelScope counterpart Alchemy3D-1M.

Dataset Layout

The dataset contain encoded latents for sparse structure (ss_latents), geometry (shape_latents/512), PBR material (pbr_latents/512). csv files are metadata.

Latents are packed as part_XXX.zip shards.

Usage

Download the dataset:

hf download libadi/Alchemy3D-1M --repo-type dataset --local-dir Alchemy3D-1M

Then unzip all zip files.

License

Apache License 2.0. Downstream 3D generators and renderers you apply to these assets may have their own terms.

Citation

@misc{li2026scalingversatile3dassets,
      title={Scaling Versatile 3D Assets Editing with a Million-Scale Dataset}, 
      author={Badi Li and Tianxin Huang and Yu Zhou and Wei-Shi Zheng and Yi Ma and Shenghua Gao},
      year={2026},
      eprint={2609.34271},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.34271}, 
}