Noise2Change: Generating Any Changes in the Noise Domain

Generating Any Changes in the Noise Domain
Q. Liu, Y. Kuang, J. Yue, P. Ghamisi, W. Xie, L. Fang
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

---

Installation

conda create -n noise2change python=3.10 -y
conda activate noise2change
pip install -r requirements.txt

---

Pretrained models

• EDM2: NVIDIA EDM2
  edm2-img512-xs-2147483-0.135.pkl
• VAE: StabilityAI
  sd-vae-ft-mse

---

Data preparation

bash preprocessing/run.sh

---

Training

Mask generator:

bash scripts/train/mask/train_mask_generator.sh

Image generator:

bash scripts/train/mask/train_image_generator.sh

---

Generation

Real T1 → synthetic T2:

# Encode T1 latents before T2
bash scripts/generate/mask/T2/generate_mask_T2.sh
bash scripts/generate/image/T2/generate_image_T2.sh

Fully synthetic (optional):

bash scripts/generate/mask/T1/generate_mask_T1.sh
bash scripts/generate/image/T1/generate_image_T1.sh
# Encode T1 latents before T2
bash scripts/generate/mask/T2/generate_mask_T2.sh
bash scripts/generate/image/T2/generate_image_T2.sh

---

Citation

@ARTICLE{10825310,
  author={Liu, Qingsong and Kuang, Yidan and Yue, Jun and Ghamisi, Pedram and Xie, Weiying and Fang, Leyuan},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  title={Generating Any Changes in the Noise Domain},
  year={2025},
  doi={10.1109/TPAMI.2025.3643733}
}