CVPR Findings2026
PGDM: Physics-Guided Noise-Free Diffusion Model Based on Point Spread Function for Light-Scattering Removal in Unpaired Biomedical Images
Jinze Zhao, Keyi Han, Qiushi Huang, Jie Tian, Zhenhua Hu
Abstract
Effectively mitigating light scattering caused by biological tissues and other intrinsic factors remains a significant challenge in enhancing the structural clarity of optical biomedical images. Although diffusion models have been a promising approach for image improvement, the addition of Gaussian noise during the forward process destroys the original structural information, making it challenging to recover details. Using anatomically paired images to train model can provide direct supervision for recovering structural details. However, acquiring such paired optical biomedical images is both challenging and costly. To address this, we propose a novel physics-guided noise-free diffusion model (PGDM) that incorporates the point spread function (PSF) to suppress light scattering while preserving structural details. Instead of adding noise, PGDM employs PSF-based optical forward model to simulate optical degradation in real world, which enabes a physically interpretable stepwise de-scattering in reverse process. To further improve image fidelity, we propose a shared frequencybased dark channel prior (Fre-DCP) to guide the reverse process and introduce a Fre-DCP loss to preserve structural details during PGDM training. Additionally, we design a new reverse sampling process to improve both fine detail and perceptual quality by leveraging complementary advantages of residual and convolutional features. Experimental results demonstrate that PGDM outperforms state-of-the-art methods, including diffusion-, GAN-and dehazing-based models, across various optical biomedical datasets. PGDM exhibits strong zero-shot ability and generalizability in suppressing light scattering across diverse optical biomedical imaging modalities without retraining.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- 4Diffusion: Multi-view Video Diffusion Model for 4D Generation cites High-Resolution Image Synthesis with Latent Diffusion Models
- Unpaired Image-to-Image Translation via Neural Schrödinger Bridge cites High-Resolution Image Synthesis with Latent Diffusion Models
- Unpaired Image-to-Image Translation via Neural Schrödinger Bridge cites Tackling the Generative Learning Trilemma with Denoising Diffusion GANs
- Diffusion Models Without Attention cites High-Resolution Image Synthesis with Latent Diffusion Models
Builds on6
- High-Resolution Image Synthesis with Latent Diffusion ModelsRobin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser et al.CVPR 2022 · 13,123 citations
- Tackling the Generative Learning Trilemma with Denoising Diffusion GANsZhisheng Xiao, Karsten Kreis, Arash VahdatICLR 2022 · 726 citations
- Unpaired Image-to-Image Translation via Neural Schrödinger BridgeBeomsu Kim, Gihyun Kwon, Kwanyoung Kim, Jong Chul YeICLR 2024 · 131 citations
- 4Diffusion: Multi-view Video Diffusion Model for 4D GenerationHaiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu et al.NeurIPS 2024 · 119 citations
- Diffusion Models Without AttentionJing Nathan Yan, Jiatao Gu, Alexander M. RushCVPR 2024 · 26 citations
Related papers
- Frequency-Guided Posterior Sampling for Diffusion-Based Image RestorationDarshan Thaker, Abhishek Goyal, René VidalICCV 2025 · 2 citations
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and DeblurringXiaoqian Lv, Shengping Zhang, Chenyang Wang, Yichen Zheng et al.CVPR 2024 · 48 citations
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired TrainingYunwei Lan, Zhigao Cui, Chang Liu, Jialun Peng et al.AAAI 2025 · 39 citations
- Fully Zero-Shot Image DehazingShuocheng Wang, Ruoxi Zhu, Jiaming Liu, Zhengyang Cao et al.ICML 2026
- Frequency Domain-Based Diffusion Model for Unpaired Image DehazingChengxu Liu, Lu Qi, Jinshan Pan, Xueming Qian et al.ICCV 2025 · 13 citations
