---
title: conflux
canonical_url: "https://www.modelscope.cn/models/gevaertlab/conflux"
md_url: "https://www.modelscope.cn/models/gevaertlab/conflux.md"
repository: gevaertlab/conflux
last_updated: 2026-08-27
license: cc-by-nc-sa-4.0
parameters: 355.5M
tensor_type:
  - F32
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 8
stars: 0
tags:
  - medical
  - medical-imaging
  - computed-tomography
  - mri
  - generative
  - 3d
---

# conflux

> conflux - gevaertlab 在 ModelScope 开源的模型。Conditional 3D latent generative models for medical imaging.

gevaertlab/conflux 是 ModelScope 魔搭社区上的 355.5M 参数机器学习模型，采用 cc-by-nc-sa-4.0 许可。

- **Repository**: gevaertlab/conflux
- **License**: cc-by-nc-sa-4.0
- **Parameters**: 355.5M
- **Tags**: medical, medical-imaging, computed-tomography, mri, generative, 3d
- **Downloads**: 8
- **Stars**: 0
- **Last updated**: 2026-08-27

Source: https://www.modelscope.cn/models/gevaertlab/conflux

---

[![arXiv](https://img.shields.io/badge/arXiv-2607.02998-b31b1b.svg)](https://arxiv.org/abs/2607.02998)
[![Dataset](https://img.shields.io/badge/🤗%20Dataset-conflux--chest--ct-yellow.svg)](https://huggingface.co/datasets/gevaertlab/conflux-chest-ct)
[![Interactive Demo](https://img.shields.io/badge/🤗%20Demo-Spaces-blue.svg)](https://huggingface.co/spaces/mxvp/conflux-chest-ct-demo)
![CONFLUX](preview.gif)

# CONFLUX

**Conditional 3D latent generative models for medical imaging.**

CONFLUX synthesizes full 3D medical volumes from structured clinical metadata: a
VAE tokenizer compresses a volume into a compact latent, a single-stream
rectified-flow transformer generates in that latent space, and a Flow-GRPO
reinforcement-learning stage sharpens label faithfulness. This repository holds
the released checkpoints, one self-contained folder per modality.

<p>
  <a href="https://arxiv.org/abs/2607.02998">Paper (arXiv)</a> &nbsp;•&nbsp;
  <a href="https://huggingface.co/datasets/gevaertlab/conflux-chest-ct">Dataset</a> &nbsp;•&nbsp;
  Code — <em>coming soon</em>
</p>

## Available checkpoints

| Folder | Modality | Resolution | Conditioning | Dataset |
| --- | --- | --- | --- | --- |
| [`chest-ct/`](chest-ct) | Chest CT | 216 × 176 × 200 | 18 findings + sex + age + kernel | [conflux-chest-ct](https://huggingface.co/datasets/gevaertlab/conflux-chest-ct) |

*More modalities (e.g. brain MRI, abdominal CT) will be added as they are trained — each as a new self-contained folder.*

Each modality folder contains:

```
<modality>/
├── vae.safetensors     3D VAE (encoder + decoder)
├── dit.safetensors     rectified-flow transformer (final, RL-post-trained)
└── config.json         architecture + latent normalization for this modality
```

Both weight files are needed: the transformer generates a **latent**, and the VAE
**decoder** turns it into the volume.

## Usage

Point `MODALITY` at the folder you want.

```python
import json, torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from models import build_vae, build_dit, flow_sample   # from github.com/mxvp/CONFLUX

REPO = "gevaertlab/conflux"
index = json.load(open(hf_hub_download(REPO, "config.json")))   # repo index (lists modalities)
MODALITY = index["default_modality"]                            # "chest-ct"
cfg = json.load(open(hf_hub_download(REPO, f"{MODALITY}/config.json")))
vae = build_vae({"vae": cfg["vae"]}); vae.load_state_dict(load_file(hf_hub_download(REPO, f"{MODALITY}/vae.safetensors"))); vae.eval().cuda()
dit = build_dit({"dit": cfg["dit"]}); dit.load_state_dict(load_file(hf_hub_download(REPO, f"{MODALITY}/dit.safetensors"))); dit.eval().cuda()

# conditioning vector layout is in cfg["cond_layout"]; for chest-ct:
# [findings(18), sex(1), age one-hot(7), kernel one-hot(16)] = 42
cond = torch.zeros(1, cfg["dit"]["cond_dim"], device="cuda")
cond[0, 2] = 1.0                                    # e.g. Cardiomegaly (finding index 2)
sc, sh = cfg["latent"]["scale"], cfg["latent"]["shift"]
with torch.no_grad():
    z = flow_sample(dit, (1, cfg["dit"]["latent_channels"], *cfg["latent"]["spatial"]),
                    steps=50, cond=cond, device="cuda")
    vol = vae.decode(z / sc + sh)                  # (1,1,*spatial_size), model units ~ HU/1000
```

## Intended use

Research use — synthetic medical-volume generation, augmentation, and method
development. **Not for clinical use.** Outputs are synthetic and must not inform
any patient-facing decision.

## Citation

```bibtex
@article{vanpuyvelde2026conflux,
  title   = {CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training},
  author  = {Van Puyvelde, Max and Gulluk, Halil Ibrahim and Van Criekinge, Wim and Gevaert, Olivier},
  journal = {arXiv preprint arXiv:2607.02998},
  year    = {2026}
}
```

## License

CC BY-NC-SA 4.0 — non-commercial research use.