Iterative Denoising for the Campanile
The noisy image at every 5th loop of denoising. From left to right, we start with a noisy image at t = 690, moving towards cleaner images at t = 90. The list of images below are at timesteps \( [690,540,390,240,90] \).
Next, below is a comparison of the original image with all three methods of denoising. We see that Gaussian Blur denoising is not at all effective. One-Step denoising, while effective, produces blurry images. Furthermore, the final result is quite "far away" from the original image of the campanile and retains a few original features. On the other hand, iterative denoising produces a sharper image, which reflects more of the details of the original image. The intuition here is that one-step denoising only gets to add one big increment of noise, while iterative-denoising makes several smaller but more accurate approximations at the time-steps. Below we see the original image, gaussian blur denoised, one-step denoised, and iterative denoised, Left-to-Right.
