---
license: mit
tags:
  - medical-imaging
  - mri
  - image-translation
  - cut
  - cyclegan
  - stargan-v2
  - mrixfields
  - baseline
pipeline_tag: image-to-image
library_name: pytorch
---

# MRIxFields2026 Baseline Checkpoints

Pre-trained baseline checkpoints for the [MRIxFields 2026](https://mrixfields.chihucloud.com/2026/) challenge (MICCAI 2026): cross-field brain MRI translation.

## Tasks

| Task | Goal | Input | Output |
|------|------|-------|--------|
| **Task 1** | Ultra-high field synthesis | 0.1T / 1.5T / 3T / 5T | 7T |
| **Task 2** | Low-field enhancement | 0.1T | 1.5T / 3T / 5T / 7T |
| **Task 3** | Any-to-any translation (single unified model) | Any field | Any field |

## Methods

| Method | Paper | Tasks | # sub-tasks |
|--------|-------|-------|-------------|
| **CUT** | Park et al., ECCV 2020 | 1, 2 | 24 |
| **CycleGAN** | Zhu et al., ICCV 2017 | 1, 2 | 24 |
| **StarGAN v2** | Choi et al., CVPR 2020 | 3 | 1 |

Total: **49** checkpoints (one per sub-task × method), **~7 GB**.

## Training

All checkpoints are from the recommended `pro_pretrained` setting: **unpaired pretrain on the retrospective split + paired fine-tune on the prospective split**. This is the strongest baseline; ablation modes (`pro_scratch`, `retro_scratch`) are kept for internal use and are not released.

## Directory Layout

```
.
├── README.md
├── LICENSE
├── task1_<src>_to_7T_<modality>/
│   ├── cut/pro_pretrained/weights/checkpoint_epoch100.pth
│   └── cyclegan/pro_pretrained/weights/checkpoint_epoch100.pth
├── task2_0.1T_to_<tgt>_<modality>/
│   └── (same as task1)
└── task3_any_to_any_multimodal/
    └── stargan_v2/pro_pretrained/weights/checkpoint_epoch50.pth
```

> Training-time configs (`config_actual.yaml`) are intentionally not included. Use the config templates in the upstream Baseline repo and override the data paths for your own environment.

## Download

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mrixfields/MRIxFields2026-Baseline",
    repo_type="model",
    local_dir="./MRIxFields2026-Baseline",
)
```

Single checkpoint:

```python
from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="mrixfields/MRIxFields2026-Baseline",
    filename="task1_0.1T_to_7T_T1W/cut/pro_pretrained/weights/checkpoint_epoch100.pth",
    repo_type="model",
)
```

## Training & Inference

See the upstream Baseline code: [`MRIxFields2026/Baseline`](https://github.com/MRIxFields/MRIxFields2026/tree/main/Baseline).

## License

MIT — see [LICENSE](LICENSE). Copyright © 2026 MRIxFields.

## Citation

If you use these baselines, please cite the original method papers:

```bibtex
@inproceedings{park2020cut,
  title={Contrastive Learning for Unpaired Image-to-Image Translation},
  author={Taesung Park and Alexei A. Efros and Richard Zhang and Jun-Yan Zhu},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2020}
}

@inproceedings{CycleGAN2017,
  title={Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks},
  author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A},
  booktitle={IEEE International Conference on Computer Vision (ICCV)},
  year={2017}
}

@inproceedings{choi2020starganv2,
  title={StarGAN v2: Diverse Image Synthesis for Multiple Domains},
  author={Yunjey Choi and Youngjung Uh and Jaejun Yoo and Jung-Woo Ha},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2020}
}
```

## Links

- Challenge website: https://mrixfields.chihucloud.com/2026/
- Dataset (Synapse): https://www.synapse.org/Synapse:syn72060672/datasets/
- GitHub: https://github.com/MRIxFields/MRIxFields2026