Prompt-SID: Learning Structural Representation Prompt via Latent Diffusion for Single-Image Denoising

Huaqiu Li, Wang Zhang, Xiaowan Hu, Tao Jiang, Zikang Chen, Haoqian Wang

Abstract

Prompt-SID is a self-supervised single-image denoising framework designed to preserve structural detail. It learns an original-scale structural representation through latent diffusion, injects that representation into a transformer denoiser through structural attention, and uses scale replay training to reduce the resolution gap. Here, “prompt” means a latent structural representation extracted from the noisy image, not a natural-language prompt.

Key results

On the SIDD validation and benchmark sets, Prompt-SID improves PSNR over Neighbor2Neighbor by 0.55 dB and 0.49 dB, and over Blind2Unblind by 0.23 dB and 0.19 dB. The paper also evaluates synthetic and fluorescence imaging denoising.

Citation

@article{li2025promptsid,
  title={Prompt-SID: Learning Structural Representation Prompt via Latent Diffusion for Single-Image Denoising},
  author={Li, Huaqiu and Zhang, Wang and Hu, Xiaowan and Jiang, Tao and Chen, Zikang and Wang, Haoqian},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39}, number={5}, pages={4734--4742},
  year={2025}, doi={10.1609/aaai.v39i5.32500}
}