Page 1 of 1

Converting a Video/Images [Help with blurred faces]

Posted: Tue Jul 30, 2019 12:36 am
by Brookcie

Hello everyone,

I have been working on some face swaps previously and they have come out very low res.

This is the following troubleshooting I've done with little improvement (if at all):

1) Increase face pics to couple thousand
2) Following tutorial on fixing alignments file for multi faces, no faces, etc.
3) Run up to 70 000 iterations

This is my dilemma of WHY I think I must be missing something obvious when converting:

During the training process the previews look quite decent (with some blur, but decent) however when using this model on converting my faces come NO WHERE NEAR as good as the preview - very blurred and complete mess. I have attached my settings and image examples to articulate my point better.

Am I missing something when converting?
Image
Above is the Preview in Training

Image
Finished Convert Pic

Convert Settings are as follows:

Input Dir: The original video I want to face swap
Output Dir: Some random folder I want to output it
Alignments: The alignments files (cleaned and edited manually in tools) for the input video

Model Dir: The trained model folder that was the output for the training section

All other settings are blank/default

Any help would be appreciated, thanks.


Re: Converting a Video/Images [Help with blurred faces]

Posted: Tue Jul 30, 2019 1:12 pm
by torzdf

Which model are you using?

You should try using the preview tool to set your convert settings prior to convert.


Re: Converting a Video/Images [Help with blurred faces]

Posted: Tue Jul 30, 2019 5:49 pm
by bryanlyon

Remember that only the last column there in your screenshot is actually the swap. I can see that even that is quite blurry and indistinct.

Also, 70k iterations is not really enough for most models (especially the more intensive ones like Realface and Villain) and will generally benefit from a lot more training. Focus on those previews and train long enough until those swap columns stop looking better.


Re: Converting a Video/Images [Help with blurred faces]

Posted: Tue Jul 30, 2019 11:59 pm
by Brookcie

Thank you all for your advice.

I am up to about 170 000 iterations and it doesn't look a whole lot better?

I feel like I am missing something important when I compare to deepfakes you see on YouTube.

Do the really good ones simply have 500 000 iterations or something? Because I feel like I am using some high quality faces to train with many different expressions and it doesn't look great at all.


Re: Converting a Video/Images [Help with blurred faces]

Posted: Wed Jul 31, 2019 12:38 am
by torzdf

It will depend a lot on the model.

The most important thing is quality of data.