Using diffusers to create denoising models

Dear Community,

I’ve successfully utilized the stable diffusion technique for creating synthetic images, as guided by the tutorial at Train a diffusion model.

My current research interest lies in exploring the potential for noise reduction in images, rather than generating entirely new synthetic images. I’m curious if this objective can be achieved using the existing capabilities of the diffusion package.

Specifically, I’m considering an approach where the added noise to an image is minimized, so as not to completely transform the original image into noise. During the diffusion reversal process, I’m thinking of using a noisy version of the image instead of initiating the process with Gaussian noise.

Could anyone provide insights or suggestions on whether this approach is feasible with the diffusion package, and if so, what might be the best method to proceed?

Thank you in advance for your assistance.