How much better can I expect?

Want to understand the training process better? Got tips for which model to use and when? This is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Training a Faceswap model.

If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
smonaghan
Posts: 4
Joined: Wed Sep 18, 2019 3:10 pm

How much better can I expect?

Post by smonaghan »

This is my first deepfake attempt. I'm trying to put the FLOTUS' face into a super silly old music video. The music video is a VHS rip and unfortunately that's the best quality available. That said, I'm going for humor more than photo-realism here, so I want it to look good enough to be recognizable and funny.

I've been training for a few days using the villain model. For the A>A and B>B should I expect much more improvement? For example, smiling teeth are just sort of a white line. Do teeth get better fidelity? Does the airbrushed smoothness of the skin eventually resolve to sharper? And then I assume once the fidelity of those models increase, the A>B and B>A will begin to improve as well.

I'm definitely open to any tips!

Image

by bryanlyon » Mon Sep 23, 2019 9:13 pm

Unfortunately you've chosen a poor source. Your best bet here is going to be to find more sources for A. Ideally they shouldn't be VHS quality Even if your final swap will be. Right now you're training the model that A is full of noise and poor quality. That simply wont work well even with a good quality B.

Go to full post
Last edited by smonaghan on Mon Sep 23, 2019 9:17 pm, edited 1 time in total.
Top
User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:
Contact bryanlyon

Re: How much better can I expect?

Post by bryanlyon »

Unfortunately you've chosen a poor source. Your best bet here is going to be to find more sources for A. Ideally they shouldn't be VHS quality Even if your final swap will be. Right now you're training the model that A is full of noise and poor quality. That simply wont work well even with a good quality B.

Top
User avatar
smonaghan
Posts: 4
Joined: Wed Sep 18, 2019 3:10 pm

Re: How much better can I expect?

Post by smonaghan »

Thanks Bryan! I updated the image above with the current state of my model from my google drive. Let me know if you think I can expect any improvement or if I should cut my losses and do my best with what I've got.

So far it's doing better than I originally expected.

Top
Locked

Return to “Training Discussion”