The number of images in celeba-hq dataset? 30k or 200k?

autonomousvision / giraffe

This repository contains the code for the CVPR 2021 paper "GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields"

https://m-niemeyer.github.io/project-pages/giraffe/index.html

MIT License

1.23k stars 160 forks source link

The number of images in celeba-hq dataset? 30k or 200k? #38

Closed Tianhang-Cheng closed 2 years ago

Tianhang-Cheng commented 2 years ago

Hi, sorry to bother you, I meet some problem about which folder I should choose.

I'd like to use images of 128*128 resolution, and I notice the celeba-128 folders contains 30k images, but the img_celeba images also has 200k images, so I got little confused about it. Could you tell me which folder should I choose? Really thanks!!! : )

m-niemeyer commented 2 years ago

Hi @Tianhang-Cheng , thanks for your question! I'm not totally sure what you mean: We have trained our model on CelebA at 64x64 pixels, and CelebA-HQ at 256x256 pixels. Hence, we have never trained a model for 128x128 pixels! If you want to do this, and you question is which dataset to choose, then it's probably up to you, but the more traditional choice would be CelebA (not CelebA-HQ) because the dataset is significantly bigger.

Hope this helps a little!

Tianhang-Cheng commented 2 years ago

Ok, thank you very much for your reply! I've got it now. For low-resolution tasks, such as 64×64, it is better to use CelebA, which has 200k training images. But for high-resolution tasks, such as 256×256 or 512×512, it's better to use CelebA-HQ, which has 30k images. Thanks a lot!