---
license: apache-2.0
tags:
- cad
- brep
- 3d-generation
- diffusion
model_name: HiFi-BRep
---

# HiFi-BRep

Checkpoints for **HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation**.

[GitHub](https://github.com/1nnoh/HiFi-BRep) · [CVPR 2026](https://openaccess.thecvf.com/content/CVPR2026/papers/Hou_HiFi-BRep_High-Fidelity_Latent_Representation_for_Robust_B-Rep_Generation_CVPR_2026_paper.pdf) · [arXiv](https://arxiv.org/abs/2608.16485) · [HuggingFace](https://huggingface.co/1nnoh/HiFi-BRep) · [ModelScope](https://www.modelscope.cn/models/innohou/HiFi-BRep)

## Inference Checkpoints

| Variant | Dataset | Faces | Decoder | Diffusion | State |
| --- | --- | ---: | --- | --- | --- |
| `4-30` | DeepCAD | 4–30 | `deepcad/hifi-brep-decoder-4-30.pt` | `deepcad/hifi-brep-diffusion-4-30.pt` | online |
| `7-30` | DeepCAD | 7–30 | `deepcad/hifi-brep-decoder-7-30.pt` | `deepcad/hifi-brep-diffusion-7-30.pt` | online |
| `abc-4-50` | ABC | 4–50 | `abc/hifi-brep-decoder-4-50.pt` | `abc/hifi-brep-diffusion-4-50.pt` | EMA |

HuggingFace:

```bash
hf download 1nnoh/HiFi-BRep \
  --include "deepcad/*.pt" "abc/*.pt" \
  --local-dir checkpoints
```

ModelScope:

```bash
python -m pip install modelscope-hub==0.1.8
modelscope download innohou/HiFi-BRep \
  --include "deepcad/*.pt" "abc/*.pt" \
  --local-dir checkpoints
```

## Full-VAE Checkpoints

| Variant | Dataset | State | Checkpoint |
| --- | --- | --- | --- |
| `4-30` | DeepCAD | online | `training/deepcad/hifi-brep-vae-4-30-online.pt` |
| `7-30` | DeepCAD | online | `training/deepcad/hifi-brep-vae-7-30-online.pt` |
| `abc-4-50` | ABC | EMA | `training/abc/hifi-brep-vae-4-50-ema.pt` |

HuggingFace:

```bash
hf download 1nnoh/HiFi-BRep \
  --include "training/*/*.pt" \
  --local-dir checkpoints
```

ModelScope:

```bash
python -m pip install modelscope-hub==0.1.8
modelscope download innohou/HiFi-BRep \
  --include "training/*/*.pt" \
  --local-dir checkpoints
```

## Usage

See the repository documentation for [inference](https://github.com/1nnoh/HiFi-BRep#inference-demo), [evaluation](https://github.com/1nnoh/HiFi-BRep/blob/main/docs/EVALUATION.md), [training](https://github.com/1nnoh/HiFi-BRep/blob/main/docs/TRAINING.md), and [data preparation](https://github.com/1nnoh/HiFi-BRep/blob/main/docs/DATA.md).

## Citation

```bibtex
@inproceedings{hou2026hifi,
  title={HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation},
  author={Hou, Junhao and Luo, Chenqi and Wang, Pufan and Lu, Jiaying and Liu, Yusheng and Qin, Feiwei and Fang, Meie and Zhou, Kun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={27199--27208},
  year={2026}
}
```

## License

The code and model artifacts use Apache-2.0. Datasets and third-party dependencies remain subject to their own terms.