Train Diffusion model for Image denoise

[abdelkareemkobo]

I have multiple grayscale images and their noisy versions. I want to train a diffusion model that, given a noisy image, returns the denoised version for various types of noise (e.g., Gaussian, Poisson, etc.). Some noise types have fewer images than others; for example, one type (Poisson) has 128 images, while another has only 28.

I attempted to build data augmentation to add noise during training for the DnCNN, which splits the images into patches (50x50) and trains the model by calculating the difference between the original image and the denoised image, then computing the loss. Unfortunately, the model did not learn effectively, suffered from overfitting, and produced blurry results.