How to convert after training

Converting source frames into swapped frames
Forum rules
Read the FAQs and search the forum before posting a new topic.

Please mark any answers that fixed your problems so others can find the solutions.
Post Reply
User avatar
rokh
Posts: 2
Joined: Sat Nov 09, 2019 8:40 pm

How to convert after training

Post by rokh » Sat Nov 09, 2019 9:16 pm

OK so I trained the models, using extracted frames one set (A) for source Video and another set (B) for the swap. Training took longer than expected (over 20 hrs.) and I had to terminate early and then tried again, and again I had to terminate after several hours. 1th question: Is it OK to use the data from a terminated early training or do you have to have a completely finished training which could take days?
So I went ahead and used the early terminated training data to try to convert my project. So I put in my input directory path and my output directory path and specified my model directory (which again contained logs from the early terminated training) and I used all the default settings except I changed the Writer to Ffmpeg and I checked the box next to Swap model. The resulted converted video was the original video with the faces blured and the second try was even worse.
Unfortunately there isn't a thoroughly explained Convert process explanation on the forum like the other processes such as extraction and training which are clearly explained in their own chapters. And it also puzzles me what happens to my B folder containing the pictures that I like to replace the originals with because during conversion I don't see a file path for those and I am just guessing that the Model directory would lead the program to that file.
Having said all that my 2nd question is: "what exactly am I not doing right?"

Tags:

User avatar
bryanlyon
Site Admin
Posts: 96
Joined: Fri Jul 12, 2019 12:49 am
Answers: 10
Location: San Francisco
Has thanked: 3 times
Been thanked: 19 times
Contact:

Re: How to convert after training

Post by bryanlyon » Tue Nov 12, 2019 5:19 pm

As long as training doesn't corrupt (for example by the PC crashing during saving) you are fine. In fact, clicking "terminate" is the proper way of stopping training and continuing it is the recommended way of running faceswap over multiple training sessions.

We are working on a full Convert guide, but it is a pretty self explanatory process compared to the others. My recommendation is to start with the Preview tool, make sure the results look good, then read the tooltips in the convert tab. The process is quite simple and I think you'll be able to figure that out (and you can ask questions if you run into a problem).

You do not need to put the source faces into convert at all. You need to provide the frames you want the faces swapped, the alignments for that source video, and the trained model. The model does the swap from "memory" (this is the point of the training) so it doesn't need to see the separated faces in order to convert.

User avatar
rokh
Posts: 2
Joined: Sat Nov 09, 2019 8:40 pm

Re: How to convert after training

Post by rokh » Tue Nov 12, 2019 8:10 pm

Thank you for replying. It just gets confusing when you have checked or unchecked a box here and there that might get you in trouble along the way. I'm still having problems with the conversion process and my next step is to go back to the beginning and try to do everything by the book and by reading the forums posts to find out where the problem is. Meanwhile I look forward to the new convert guide when it becomes available. Thanks again.

Post Reply