search

hila-chefer / TargetCLIP

[ECCV 2022] Official PyTorch implementation of the paper Image-Based CLIP-Guided Essence Transfer.
232 stars 27 forks source link

What is the parameters of training #2

Closed ajundo closed 2 years ago

ajundo commented 2 years ago

Great work! I tried to train my own dirs, but failed in some case. Are you using the default parameters to get them?

hila-chefer commented 2 years ago

Hi @wangsiheaaaa, thanks for your interest in our work! The hyperaprameters are consistent through all our experiments. The only hyperparam that we may change is the number of source images to train with (due to limited resources we didn’t always have the option to train with 8 images, so some of the directions had to settle for 4). If it’s ok with you, can you please share your target image? Also, it is highly recommended to train with images that were inverted. Where did you get the training images from? Thanks.

ajundo commented 2 years ago

Following your Elsa examples, i tried some cartoon faces. The inverted image by the pretrained e2e-ffhq model is poor. Did you use the e2e-ffhq to invert the Elsa image too? source: 3 inverted version: 00002

hila-chefer commented 2 years ago

@wangsiheaaaa I see :) The elsa example is part of our non-inverted targets examples- so the elsa target was actually not inverted :) I've updated our README with training instructions, please feel free to ask any questions if anything is unclear :) For your target, I did a very quick sanity check to train and see that the results make sense, here are the results I got for just 4 training images (results are presented on our training set, so I didn't play with the coefficients at all): example