---
pipeline_tag: text-to-3d
---

# AssetFormer: Modular 3D Assets Generation with Autoregressive Transformer

AssetFormer is an autoregressive Transformer-based model designed to generate modular 3D assets from textual descriptions. By adapting module sequencing and decoding techniques inspired by language models, the framework enhances the quality of 3D asset generation composed of primitives that adhere to constrained design parameters.

- **Paper:** [AssetFormer: Modular 3D Assets Generation with Autoregressive Transformer](https://huggingface.co/papers/2602.12100)
- **Repository:** [https://github.com/Advocate99/AssetFormer](https://github.com/Advocate99/AssetFormer)

## Installation

To get started, clone the official repository and install the dependencies:

```bash
git clone https://github.com/Advocate99/AssetFormer.git
cd AssetFormer
conda create -n assetformer python=3.12
conda activate assetformer
pip install -r requirements.txt
```

## Preparation

1. Download the `flan-t5-xl` models and place them in the `./pretrained_models/t5-ckpt/` folder:
    ```bash
    huggingface-cli download google/flan-t5-xl --local-dir ./pretrained_models/t5-ckpt/flan-t5-xl
    ```

2. Download the `inference_model.pt` from this Hugging Face repository and place it in the `./pretrained_models/` directory.

## Inference

Run the following command to sample 3D assets as JSON files:

```bash
python sample.py --gpt-ckpt ./pretrained_models/inference_model.pt
```

After sampling, you can use the Blender scripts provided in the official repository (`./blender_script/`) to render the 3D assets with the modular fbx files.

## Citation

If you find this work useful, please kindly cite:

```bibtex
@article{zhu2026assetformer,
  title={AssetFormer: Modular 3D Assets Generation with Autoregressive Transformer},
  author={Zhu, Lingting and Qian, Shengju and Fan, Haidi and Dong, Jiayu and Jin, Zhenchao and Zhou, Siwei and Dong, Gen and Wang, Xin and Yu, Lequan},
  journal={arXiv preprint arXiv:2602.12100},
  year={2026}
}
```

## Acknowledgement
The codebase is developed based on [LlamaGen](https://github.com/FoundationVision/LlamaGen).