Would someone mind taking a look at my data set?

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.

Post Reply
User avatar
tbailey
Posts: 2
Joined: Tue Mar 02, 2021 10:34 am

Would someone mind taking a look at my data set?

Post by tbailey »

HI there,
I'm a new user and have recently trained my first model for over a week on a standard CPU (100K iterations) and got a big blurry mess as a result. I've learned a lot, and know a few of the issues for another run, but I'm wondering if someone who knows about this stuff could take a quick look at my data set and tell me if I'm wasting my time?
It would be a massive help to know whether I'm likely to get any good results with these source files or not.

Source A: is Britney spears from the music video for toxic - I haven't used any other footage of her for the training, only this video. Meticulously sorted through faces in manual to make sure they were all correct.

Source B: https://1drv.ms/v/s!Aixt82g3hO24rn_6Cbw ... c?e=kz0EaK

Obviously the faces are very very different, and I'm not attempting to get a convincing swap - just looking to get his face on there and mimic the movements, in a disturbing way.

I guess my questions are:
Is this input data going to be good enough?
Why was my first result so blurry? Do I need a higher resolution B source?
Is training on a CPU a waste of time - I could ask around to borrow a GPU perhaps?

Any help you could give me would be really useful - I think I've got as far as I can go on my own, and a week at a time of training is not something I can experiment with!

Thanks in advance


User avatar
bryanlyon
Site Admin
Posts: 596
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 154 times
Contact:

Re: Would someone mind taking a look at my data set?

Post by bryanlyon »

Using one video per side is usually a way to guarantee poor quality. There simply isn't enough variety in either side of this dataset. You need to focus on a large number of different lighting conditions, expressions, and poses. In your case you have some variety in expression and pose, but none in lighting. Even the Toxic video is pretty uniquely lit throughtout.

Variety is KEY to a quality swap.

Also, training a a CPU is a REAL pain, expect weeks if not months of training. You may want to examine cloud solutions to get access to a GPU.


User avatar
tbailey
Posts: 2
Joined: Tue Mar 02, 2021 10:34 am

Re: Would someone mind taking a look at my data set?

Post by tbailey »

Thanks for the reply. I'll look into providing more source material


Post Reply