style transfer
stone100010 opened this issue · 2 comments
When I was training the '. NPY' file, '-- num steps' I used 50000 steps and' -- no tiled '. The training picture was close to the original picture. I started to carry out style transfer, The result is not good. I think it may be that the '-- num steps' in the training'. NPY 'file is not enough. Now I am continuing training with the goal of 300000' - num steps'. Once again, I would like to express my most sincere thanks!
Issue-Label Bot is automatically applying the label question to this issue, with a confidence of 0.75. Please mark this comment with 👍 or 👎 to give our bot feedback!
Links: app homepage, dashboard and code for this bot.
There is a trade-off between image quality and semantic quality, and a large number of steps improve image fidelity at the cost of semantic fidelity. I think --num_steps is too large (like 1 order of magnitude too large), especially with --no_tiled set to True. You should check this discussion.
For questions related to StyleGAN2, I would suggest to check the popular repositories:
- https://github.com/rolux/stylegan2encoder/issues
- https://github.com/Puzer/stylegan-encoder/issues
- https://github.com/pbaylies/stylegan-encoder/issues
Because in my repository with a few stars, barely nobody will check the question.