I am working on diffusion models and I came across your paper, very impressive and interesting work, congrats! I was just curious about the dynamic range that you used to clip the medical CT images. In the image_datasets.py L103 you clip the images from 0 to 0.1 and then divide by 0.35.
What was the rationale to clip the images by those values but the FBP and RLS using a dynamic range of [0, 1]? I've read your lines on the paper where you say "We rescale each dataset globally to make them have the same intensity range, but we do not perform any normalization on those images", but I am still a bit confused about the coneection between this and your implementation.
ignaciobartol
commented
2 days ago
Hi,
I am working on diffusion models and I came across your paper, very impressive and interesting work, congrats! I was just curious about the dynamic range that you used to clip the medical CT images. In the image_datasets.py L103 you clip the images from 0 to 0.1 and then divide by 0.35.
What was the rationale to clip the images by those values but the FBP and RLS using a dynamic range of [0, 1]? I've read your lines on the paper where you say "We rescale each dataset globally to make them have the same intensity range, but we do not perform any normalization on those images", but I am still a bit confused about the coneection between this and your implementation.
Thanks a lot!