Faceswap Forum

I have a problem, my result after converting is not so good after 1000000 iterations. The strange thing is that, during the training, it looks really good in my monitor. What could be happening? One thing that is bad in after converting is that the mask does not seem to cover the face properly, the lower lip looks like the original, its not changed. The eyes look halfgood though.

Try the preview tool, it helps you see the results of your convert settings on the video itself.

I do use the preview tool, that looks bad to. I try to change the parameters as well but very little is changed and it does not improve.

Preview image is tiny, convert can be huge.

Save a copy of the preview image during training and zoom into it. It should give you a better idea of why you are seeing the differences you are.

I can clearly see on the training image that the lower lip looks much thinner, like it should. On the convert its just as thick as on the original face. Also overall the training image looks really promising even considering its so small. I should also say that if I swap the model its really OK.

The training previews are generated in exactly the same way as the converted image.

Thats my thought exactly but the result is not the same. I will do some more work and if I dont get it right I can try to post some pictures to show how it looks. I have a 32" monitor so I se clearly how it looks after/during traning. Can it be the alignements that somehow is not the same? Its a mystery.

While training it looks good while converting it become blur.

I'm having a similar issue after updating. Training preview looks fine

But when I use the preview tool or convert the mask is smaller than before

Here's the preview tool with manual color brightness turned all the way down(VGG clear mask, this is what I use with training)

Extended mask

I think something happened with the alignments and conversion. The masks look correct when I look at them with the manual tool

VGG clear

Extended

Do you know if there is a way I can undo the latest update(s) to get this working again with my data? I haven't tried making a new model on the latest update but have hundreds of hours invested in a few training models so I don't want to start over if this is an issue with when the model was created.

Thanks for the heads up. This should be fixed in latest commit.

Thanks!

FYI I started up 2 new models and the face centering appears to be working correctly but the new legacy model had the same issue. So it seems to only be with conversion on models with legacy centering.