---
license: apache-2.0
datasets:
- ibrahimhamamci/CT-RATE
language:
- en
pipeline_tag: text-to-3d
tags:
- medical
- ct
- diffusion
- controlnet
- retrieval-augmented-generation
---

# RAGText2CT Weights

Weights for **RAGText2CT: Retrieval-Augmented Anatomical Guidance for Text-to-CT Generation**.

This release is independent from `dmolino/text2ct-weights` and contains the full checkpoint set needed by the `RAGText2CT-Release` codebase.

## Included Files

Under `models/`:

- `autoencoder_epoch273.pt`
- `unet_rflow_200ep.pt`
- `CLIP3D_Finding_Impression_30ep.pt`
- `controlnet_rag_best.pt`

Under `configs/`:

- `config_rag_rflow.json`

## What Each Weight Does

- `autoencoder_epoch273.pt`: 3D VAE for latent compression and decoding.
- `unet_rflow_200ep.pt`: text-conditioned latent diffusion UNet from the Text2CT backbone.
- `CLIP3D_Finding_Impression_30ep.pt`: CLIP3D report encoder checkpoint.
- `controlnet_rag_best.pt`: retrieval-guided anatomical ControlNet checkpoint for RAGText2CT.

## Intended Use

These checkpoints are intended for research on text-conditioned 3D CT generation and retrieval-augmented anatomical guidance.

They are not intended for clinical use or diagnostic decision making.

## Code

Use these weights with the companion repository:

- `RAGText2CT-Release`

The code release expects the files to live under `models/` with the names above.

## Notes

- The first three checkpoints are shared with the original Text2CT pipeline.
- `controlnet_rag_best.pt` is the additional checkpoint specific to the retrieval-augmented extension.
- Retrieval-bank artifacts such as `impression_embeddings.npy` and `impression_paths.json` are not included in this weights repo.

## Citation

```bibtex
@article{Molino2026RAGText2CT,
  title={Retrieval-Augmented Anatomical Guidance for Text-to-CT Generation},
  author={Molino, Daniele and Caruso, Camillo Maria and Soda, Paolo and Guarrasi, Valerio},
  year={2026},
  journal={arXiv preprint arXiv:2603.08305}
}
```