search

FirasGit / medicaldiffusion

Medical Diffusion: This repository contains the code to our paper Medical Diffusion: Denoising Diffusion Probabilistic Models for 3D Medical Image Synthesis
353 stars 62 forks source link

About vertically flipping #18

Open Ryann-Ran opened 9 months ago

Ryann-Ran commented 9 months ago

Why doesn't vertically flipping influence the generation? If the model receives the flipped image as input, why doesn't it learn to generate the flipped one?

malonzia commented 6 months ago

Because the target images are not flipped.And the loss function is based on the difference between the target image and the generated image, rather than the difference between the generated image and the flipped input image.

Ryann-Ran commented 6 months ago

@malonzia Thanks for your answer. So, does it mean that any type of data augmentation can be applied to the input image?

malonzia commented 6 months ago

@malonzia Thanks for your answer. So, does it mean that any type of data augmentation can be applied to the input image?

Yes, any type of data augmentation can indeed be applied to the input image. This assertion is based on the papers and code I have read.