ReCo-Diff

ReCo-Diff: Residual-Conditioned Deterministic Sampling for Cold Diffusion in Sparse-View CT

---

🔷 Highlights — ReCo-Diff

• Residual-Conditioned Deterministic Diffusion
  ReCo-Diff introduces residual-conditioned self-guided sampling that explicitly enforces residual consistency during cold diffusion. By directly conditioning on observation residuals, the model stabilizes multi-step reconstruction without relying on heuristic reset strategies.

• Stable Reconstruction Under Severe Sparsity
  The proposed method suppresses error accumulation across sampling steps and maintains consistently lower timestep-wise error trajectories, particularly in extremely sparse-view settings (e.g., 18 views).

• Unified Model Without Fine-Tuning
  A single trained model supports multiple target view configurations (18 / 36 / 72 views) without additional fine-tuning, demonstrating robustness and practical applicability across varying sparsity levels.

📦 Requirements

• Create a new virtual environment and install the required libraries using requirements.txt.

git clone https://github.com/choiyoungeunn/ReCo-Diff.git
cd ReCo-Diff
conda create -n recodiff python=3.7
conda activate recodiff
pip install -r ./requirements.txt

🔧 Install torch-radon

• torch-radon is required for simulating DRRs and geometry utils. Install torch-radon by:

1. Download torch-radon from torch-radon

   git clone https://github.com/matteo-ronchetti/torch-radon.git

2. Due to some out-dated Pytorch function in torch-radon, modify code by running

   cd torch-radon
   patch -p1 < path/to/ReCo-Diff/torch-radon_fix/torch-radon_fix.patch

3. Install torch-radon by running

   python setup.py install

---

📊 Dataset

• Download the AAPM dataset.
• After downloading, run preprocessing with ./datasets/preprocess_aapm.py.
• Provide dataset paths directly in each script before running.

python ./datasets/preprocess_aapm.py

• Provide dataset paths directly in each script before running.

Dataset Structure

Organize the dataset directory as follows:

<aapm16 data root>/
├── train_img/
│   ├── L067_FD_1_1.CT.0001.0001.2015.12.22.18.09.40.840353.358074219.npy
│   ├── L067_FD_1_1.CT.0001.0002.2015.12.22.18.09.40.840353.358074243.npy
│   └── ...
└── test_img/
    ├── L506_FD_1_1.CT.0002.0001.2015.12.22.20.19.52.894480.358589814.npy
    ├── L506_FD_1_1.CT.0002.0002.2015.12.22.20.19.52.894480.358589838.npy
    └── ...

---

💾 Pretrained Checkpoint

Download a checkpoint and use it for testing.

Checkpoint Structure

ReCo-Diff ckpt/
├── ReCo-Diff-net-colddiff_179_epoch.pkl
└── ReCo-Diff-net-colddiff_ema_179_epoch.pkl  

---

🚀 Train (ReCo-Diff)

• Script: recodiff_train.sh
• Set res_dir and dataset_path at the top of the file, then run:

bash recodiff_train.sh

---

🧪 Test (ReCo-Diff)

Single checkpoint

• Script: recodiff_test.sh
• Set res_dir, dataset_path, and net_checkpath_default at the top of the file, then run:

bash recodiff_test.sh

All checkpoints in a directory

• Script: recodiff_test_ALL_search.sh
• Set ckpt_dir at the top of the file, then run:

bash recodiff_test_ALL_search.sh

---

📈 Results

---

📚 Citation

If you find our paper helps you, please kindly cite our paper in your publications.

@misc{choi2026recodiffresidualconditioneddeterministicsampling,
      title={ReCo-Diff: Residual-Conditioned Deterministic Sampling for Cold Diffusion in Sparse-View CT}, 
      author={Yong Eun Choi and Hyoung Suk Park and Kiwan Jeon and Hyun-Cheol Park and Sung Ho Kang},
      year={2026},
      eprint={2603.02691},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.02691}, 
}

---

🙏 Acknowledge

Our code is built upon CvG-Diff.
We sincerely thank the authors for generously sharing their code.