search

nolan-dev / GANInterface

Tool to interface with a StyleGAN model
Apache License 2.0
217 stars 25 forks source link

How to find a stable combination of feature maps to edit specified locations? #11

Open qzq2514 opened 4 years ago

qzq2514 commented 4 years ago

really appreciate your nice work！ But I have a small trouble. you have found the stable combination of feature map to edit anime faces. but when I edit the non-anime face images on your v0.2.0-alpha using ffhq.pb . I can't find a stable combination of feature maps to edit specified area of those face photos，such as mouth-smiling. How to get those combination? looking forward to your reply？Thanks

nolan-dev commented 4 years ago

Hey sorry for the late reply. It's possible it may be more difficult on natural images of faces, I haven't tried enough to be sure. I was planning on doing another blog on the subject because v0.2.0-alpha was setting up for better scripting tools for doing just that. One application I was interested in was finding the feature maps that led to a feature like smiling/non-smiling, creating examples, and then using that to train a classifier without need for any prior labels. However, I decided to prioritize some other projects (very interested in combining 3D software with generative models because it reflects the actual world better). If you've tried the techniques I listed in https://towardsdatascience.com/animating-ganime-with-stylegan-the-tool-c5a2c31379d for various layers and didn't get any good results, it might turn out that anime faces are more conducive to this type of modification than other types of images. I'd definitely like to hear if you manage to get it to work.

Thanks, Nolan

qzq2514 commented 4 years ago

Thanks for your reply! I did follow the steps you show in the introduction of tools,and I did really find some seemingly stable combination of feature map to edit face photos,like mouth-opening/mouth-closing, curly hair/non-curly hair,but to be honest,it is really time consuming. Beside, I am really happy to have the similar idea about training a classifier or a discriminator but the goal is somewhat different. By the way,I am also insterested in your next project about 3D software. Thanks for your reply again~

qzq2514 commented 3 years ago

Thanks for your work, based on your excellent work, I conduct some more investigation and experiment, it all shows a great effect on facial attribute editing, so I write a paper and want to put you in my acknowledgement to really appreciate your work and your help to me. I try to connect you from medium and twitter but I am a newer to these two tools, so I connect you here. I am looking forward to your reply. Thanks

nolan-dev commented 3 years ago

Happy to hear that you've been able to expand on the work. I'd certainly appreciate any acknowledgments, if you want you can email me at nkent.dev@gmail.com