---
license: mit
library_name: transformers
pipeline_tag: image-to-3d
tags:
- lego
- 3d-generation
- autoregressive
- transformer
- llama
- dinov2
- clip
- siggraph-asia-2025
---

# LegoACE: Autoregressive Construction Engine for Expressive LEGO® Assemblies

Official model weights for **LegoACE**, presented at **SIGGRAPH Asia 2025**.

LegoACE is an autoregressive transformer that generates LEGO® assemblies as
sequences of placed bricks. This repository hosts two pretrained variants:

| Subfolder | Conditioning | Encoder | Training steps |
|-----------|--------------|---------|----------------|
| `mv/`     | Multi-view images (4 views) | [DINOv2-base](https://huggingface.co/facebook/dinov2-base) | 520K |
| `text/`   | Text descriptions           | [CLIP ViT-B/32](https://huggingface.co/openai/clip-vit-base-patch32) | 210K |

- 📄 Paper: [LegoACE @ SIGGRAPH Asia 2025](https://doi.org/10.1145/3757377.3763881)
- 💻 Code: [VAST-AI-Research/LegoACE](https://github.com/VAST-AI-Research/LegoACE)
- 📊 Architecture: 32-layer Llama-style transformer, hidden size 768, vocab ~16K

---

## Quick start

> Full inference pipeline (LDR tokenizer, multi-view rendering, LDR → GLB
> conversion) lives in the [GitHub repository](https://github.com/VAST-AI-Research/LegoACE).
> The snippets below show only how to load the weights.

```bash
git clone https://github.com/VAST-AI-Research/LegoACE.git
cd LegoACE
pip install -e .
```

### Multi-view image conditioned (recommended)

```python
from model.llama_image_condition import ImageConditionModel

model = ImageConditionModel.from_pretrained("VAST-AI/LegoACE", subfolder="mv").to("cuda")
```

End-to-end usage with the `dataset/MVNpzDataset.py` loader and Blender-based
GLB export is documented in the GitHub README:

```bash
python inference/inference_multi_view.py \
    --ckpt_dir VAST-AI/LegoACE \
    --dataset_name <your_dataset> \
    --dataset_class dataset.MVNpzDataset.MVNpzDataset \
    --save_dir ./outputs/inference \
    --save_name mv-demo \
    --infer_number 100 --batch_size 4 --repeat 4 --dataset_split val
```

### Text conditioned

```python
from model.llama_text_condition import TextConditionModel

model = TextConditionModel.from_pretrained("VAST-AI/LegoACE", subfolder="text").to("cuda")
```

```bash
python inference/inference_text_condition.py \
    --ckpt_dir VAST-AI/LegoACE \
    --dataset_name <your_dataset> \
    --save_dir ./outputs/inference --save_name text-demo \
    --prompts "A red sports car" "A modern brick bed" "A bridge over a river"
```

---

## Outputs

Each generation step emits a quintuple `(x, y, z, rotation_id, brick_type_id)`.
The full pipeline converts those token sequences into:

1. **LDR** — text-format LEGO instructions (LDraw)
2. **GLB** — 3D mesh via Blender + [ImportLDraw](https://github.com/TobyLobster/ImportLDraw)
3. **Normal maps** — pyrender renderings of the assembled model

LegoACE supports an LDR vocabulary covering 28 common brick types and 20
discrete rotation classes; see [`utils/brick_ids.py`](https://github.com/VAST-AI-Research/LegoACE/blob/main/utils/brick_ids.py).

---

## Intended uses & limitations

**Intended uses**
- Research on autoregressive 3D / LEGO® generative models.
- Generating LEGO assemblies for academic and creative exploration.

**Limitations**
- Outputs are restricted to the 28-brick vocabulary used in training.
- Quality depends on prompt phrasing (text) or image quality (multi-view).
- The model has been trained primarily on small/medium-scale assemblies and
  may produce structurally unstable or non-buildable arrangements.
- Generation requires the LDR tokenizer files (`*_dat_dict.json`,
  `*_rot_dict.json`) that ship with the dataset, not with these weights.

---

## Citation

```bibtex
@inproceedings{xu2025legoace,
  author    = {Hao Xu and Yuqing Zhang and Yiqian Wu and Xinyang Zheng and
               Yutao Liu and Xiangjun Tang and Yunhan Yang and Ding Liang and
               Yingtian Liu and Yuanchen Guo and Yanpei Cao and Xiaogang Jin},
  title     = {LegoACE: Autoregressive Construction Engine for Expressive LEGO{\textregistered}
               Assemblies},
  booktitle = {Proceedings of the {SIGGRAPH} Asia 2025 Conference Papers},
  publisher = {{ACM}},
  year      = {2025},
  pages     = {40:1--40:11},
  doi       = {10.1145/3757377.3763881},
  url       = {https://doi.org/10.1145/3757377.3763881}
}
```

---

## License

Released under the [MIT License](https://github.com/VAST-AI-Research/LegoACE/blob/main/LICENSE).

LEGO® is a trademark of the LEGO Group, which does not sponsor, authorize, or
endorse this project.