Faceswap Forum

That looks like a problem with (at least) one of your training images not being an extracted faceswap face

The solid colors mean model corruption. You can try lowering the learning rate. That preset is particularly complex, so can be susceptible to collapse at higher learning rate.

Also, if using mixed precision, I highly recommend setting the Epsilon Exponent to -5 for that preset.

This happens for 1 of a number of reasons....

1) There are special characters within the meta data of the video file. You can use something like MP3Tag

2) It's a false positive derived from some kind of memory error. If this happens, normally just running the process again remedies the problem.

It's possible to do any of the Faceswap processes on Google Colab. I don't use it, personally, so can't help you with how, but it shouldn't be too involved.

ianstephens wrote: ↑Mon Jul 12, 2021 8:31 pm

Quick question @torzdf -

Did you use "no warp" at the end of training for this video? Or was it simply a run of _2M iterations with the StoJo preset model?

Yes. Can't remember when I disabled. Probably about halfway through

ianstephens wrote: ↑Tue Jul 06, 2021 1:52 pm

Is this simple change below all it takes to train at 384px for example (or is there more to it)?

Short answer. You could just do that, but there's more to it.

Long answer: A lot of maths.... I will come back and articulate at some point in the (hopefully) not too distant future.

That speed is about correct for StoJo. It is a heavy and complex model.

Roudrunner_404 wrote: ↑Thu Jul 08, 2021 1:37 pm

Ok so. I might not be that smart but. Must I swap face in the video I used for training?

No. It's best to follow the guides to understand how it all works (it's easier than me just re-explaining it all here)

The GTX 710 is too old to use the latest Faceswap version (requires Compute Capability 3.5+). Also, 2GB means you may not be able to do very much. You can try V1 of Faceswap though, which should work on your card:

https://github.com/deepfakes/faceswap/r ... tag/v1.0.0

Just use the preview tool to adjust the size of the mask to exclude the eyebrows.

babala wrote: ↑Sun Jul 11, 2021 10:22 pm

so I my 3080 start running now.

with batch size 4 and 12.8 EGs/sec. and batch size 5 also working but 1 EGs/sec more so...

is this normal speed for Phaze-A ?

That very much depends on the settings you choose.

FWIW convert is not particularly well optimized.

Either way, It should never use all of your RAM. It's specifically built to try not to use it by only holding onto those frames that it is processing at any given time.

Moved to general, as I'm not sure what you're asking for here.

How much GPU/CPU to use for what?

I will need to look into + test this. My initial assumption would be memory copies though