1

Using Open Ai's CLIP with BigGAN, VQGAN etc. are there image databases other than:

ImageNet 1024, ImageNet 16384, COCO,S-FLCKR, WikiArt, FacesHQ

that can be used? If so what others?

Rubén
  • 34,714
  • 9
  • 70
  • 166
garrettlynchirl
  • 790
  • 8
  • 23

1 Answers1

2

ImageNet 1024 and ImageNet 16384 is two different modifications of VQGAN trained on the same ImageNet dataset. Use may find all models trained by authors on their github: https://github.com/CompVis/taming-transformers

skabbit
  • 96
  • 1
  • 6
  • 1
    A stupid question, but since I'm only fiddling around with it, which one does create more realistic art? And why 1024 vs 16384, what is this? – Fusseldieb Oct 07 '21 at 17:21
  • 2
    16384 means it has more neurons ("memory" to catch details), and it could give more detailed images than 1024. But it doesn't necessarily lead to artistic realism in terms of CLIP usage. – skabbit Oct 09 '21 at 10:22