search

VinAIResearch / WaveDiff

Official Pytorch Implementation of the paper: Wavelet Diffusion Models are fast and scalable Image Generators (CVPR'23)
GNU Affero General Public License v3.0
366 stars 28 forks source link

DWT and IWT #4

Closed Mdahao closed 1 year ago

Mdahao commented 1 year ago

Hi sir，Thanks for your work. What is the difference between the DWT and IWT in WaveCNet and pytorch_wavelets? Which package you are using?

hao-pt commented 1 year ago

I apologize for the confusion earlier. Initially, during our research, we used the wavelet transformation from pytorch_wavelets to transform the input image into wavelet subbands. However, we found that this lib is not supported for gradient propagation. Therefore, when designing the Wavelet-embedded generator, we opted to use the wavelet transformations from WaveCNet instead.

To this end, most of our experiments utilized wavelet transformations from pytorch_wavelets to process input images, while the wavelet-embedded generator used the ones from WaveCNet. However, only two experiments, namely CelebA 256 & 1024, used wavelet layers from WaveCNet for both the input image and the network.

Mdahao commented 1 year ago

I apologize for the confusion earlier. Initially, during our research, we used the wavelet transformation from pytorch_wavelets to transform the input image into wavelet subbands. However, we found that this lib is not supported for gradient propagation. Therefore, when designing the Wavelet-embedded generator, we opted to use the wavelet transformations from WaveCNet instead.

To this end, most of our experiments utilized wavelet transformations from pytorch_wavelets to process input images, while the wavelet-embedded generator used the ones from WaveCNet. However, only two experiments, namely CelebA 256 & 1024, used wavelet layers from WaveCNet for both the input image and the network.

Thanks for your reply.