IDF: Iterative Dynamic Filtering Networks for Generalizable Image Denoising
1CS, Hanyang University    2CS, Friedrich-Alexander-Universität Erlangen-Nürnberg   
*Equal Contribution    †Corresponding Author   

ICCV 2025

🔥 One Model. Many Noises. Clean Results. 🔥

Compact 0.04M parameters with strong Out-Of-Distribution (OOD) performance!


Even though IDF is trained on a single Gaussian level (σ = 15), It generalizes to both Synthetic (Gaussian, Spatial Gaussian, Poisson, Speckle,
Salt & Pepper) and Real-World (SmartPhone, DSLR, Monte Carlo, Medical Imaging) noise.

🔎 Drag to Compare: From Noisy to Clean

Compared with prior SOTA image denoisers trained at a single Gaussian noise level (σ = 15), IDF removes OOD noise more effectively across diverse images while using far fewer parameters and requiring no external priors such as CLIP.

🎚️ Control Noise Removal and Preserve Detail

IDF allows users to tune denoising strength by adjusting the number of iterations,
balancing fidelity versus realism, computational cost, and inference speed.

💡 Generate Pixel-Wise Varying Averaging Filters

IDF generates pixel-wise varying denoising kernels and regularizes them by enforcing their elements to sum to one.
This mechanism guides IDF toward content-adaptive averaging rather than memorizing specific noise patterns.

BibTeX 

                    
                    @inproceedings{kim2025idf,
                        title={IDF: Iterative Dynamic Filtering Networks for Generalizable Image Denoising},
                        author={Dongjin Kim and Jaekyun Ko and Muhammad Kashif Ali and Tae Hyun Kim},
                        booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
                        year={2025}
                    }

Visitor Count

several