Page 1 of 1

Does the old "cnn" extractor still exist? Need help with updating commands from a very old version of this program!

Posted: Wed Aug 07, 2024 8:26 pm
by norm_l1

I messed around a bit with this program quite a while ago (mid-2018). It seems the software has changed and improved a lot since then, which is great! Back then I used the following commands, which gave me really solid results, although they did take quite a while to run:

Code: Select all

python faceswap.py extract -i photo/trump -o data/trump -D cnn
python faceswap.py train -A data/trump -B data/cage -m trained_models/trump_cage
python faceswap.py convert --converter Masked --seamless -mh --blur-size 0 --erosion-kernel-size 1 -i photo/trump -o output/trump_cage -m trained_models/trump_cage

Not surprisingly, these commands do not work as-is any more. I opened the current GUI and there is a dizzying variety of options available so I'm a little lost. Is the extractor that used to be called "cnn" as of mid-2018 still a part of this program, or has it been completely superseded? To be honest, I don't remember exactly what the flags --converter Masked --seamless -mh --blur-size 0 --erosion-kernel-size 1 were for, except that I had done some trial-and-error and it seemed that these flags gave a very nice result. If anyone could point me in the right direction, as far as getting comparable results, I would be very grateful. Thank you very much.


Re: Does the old "cnn" extractor still exist? Need help with updating commands from a very old version of this program!

Posted: Tue Aug 13, 2024 9:37 am
by torzdf

Welcome back :)

That was probably the DLIB extractor which, yes, has been completely removed from the project (due to being inferior to current offerings, and constantly causing issues for end users as it needed to be compiled to use the GPU).

The convert guide is kept up to date, and faceswap has a 'preview tool' to check on settings before running a full convert. You can look into it here:
viewtopic.php?t=1083

The other guides should help you with wading through a lot of the other options: