So the idea is to iterate and adapt the model() to fit an image. To make the whole system smaller. Insert np.sort() for the model weights. This way you can ?lossy compress the weight matrix and get a smaller overall file. Since a many sorted data under some denoising does look continous.