Need some help to clear my doubts

[Shutterbug2109]

trafficstars

First of all thanks a lot for this implementation. I am a student trying to understand the concept of Diffusion models and their applications. I am confused about the inputs the model takes to train and test. I am trying to adapt this to the dataset I have. Please validate my understanding As mentioned there are 3 folders : LR_imgs = contains the LR images (either you have it or you downscale it to the required dimension) In my case I have 128x128 LR images, which are obtained from some method. HR_Images = contains Ground truth Images of HR ( I have 256x256 HR GT images which are obtained from Super-resolution methods) SR_Images = contains Upscaled LR images to the size of GT HR images using bicubic interpolation (In my case it will be 128x128 -->256x256)

I need to train everything from scratch. My questions are :

1. When we say posterior q for reverse diffusion process is conditioned Gaussian distribution q(x_{t-1}|x_t, x_0), we need the original image x_0 in ideal. So is it a reconstructed o/p of our diffusion model or any Image that we are providing?

2. When are we using the LR images? Just to check the performance of our model?

3. Is it necessary to have images in the same order in LR, HR & SR folders? As I have different folders for everything and the preprocessing is done outside the code, do I need to load the images in the same order from each folder?

If anyone can answer my questions, it will be really helpful. Thanks a lot