search

timothybrooks / instruct-pix2pix

Other
6.1k stars 527 forks source link

Trained on 256px but inference on 512px? #97

Open andreemic opened 1 year ago

andreemic commented 1 year ago

As far as I understand the training config, IP2P was trained on 256x256 images but images in the paper are generated at 512px.

  1. Doesn't training at a different resolution than inference impact performance?
  2. If so, why did you train on 256px? Did you try training on 512px?
timothybrooks commented 1 year ago

Yes! I trained at lower resolution because I did not have compute available to me to train at 512x512. I tried at 128x128 as well, and things definitely improved at 256x256, so it might improve more if trained at 512x512.

kosmels commented 2 weeks ago

Hello guys! @timothybrooks thank you for your answer. I have 1 question - do I need to change anything directly in source code (or train config) if I want to train on larger resolution, e.g. 512x512?