Training seems to not progress. Or atleast thats what the analytics show.

[kingrun2]

So read through the training guide. I know about the quality of images and everything and the time it needs. But it also mentions to check the chart to see if its going down. Mine just seems to be in one place I think I don't know how much to rely on the chart as the images are still quite blurry on the SWAP face.
Chart:

Preview:

This is in total I think of 16hours and 170000 iterations at 16 batches. I understand that picture quality matters and theres a lot of nuances my only questions is will it get better given more time. Like say a few days? Or is this the best it will get?

[torzdf]

Honestly? Depends on the data. If you can post a screenshot of your training preview window we may be able to offer more guidance....

Also, the rate the graph drops slows DRAMATICALLY the further it gets, so you really need to use the graphs in the Analysis tab to be able to zoom right in.

[kingrun2]

Yeah this is the first time I am doing this. Used mostly default settings. Footage is not that great. But just wondering if it will get better over time

Preview

Total graph in analytics.

I am not expecting this to look amazing or anything. I was just wondering will the face blurriness would subside over time. I know I should probably choose a video with less motion, more quality in it and there is a lot of lightning changes where as my face is supplied in the same lighting. But mostly is the facial blur that worries me.

[torzdf]

The good news is that it looks like it is still learning (you can zoom the graph if you so wish, also).

The bad news is that your data (particularly on the B side) is no way near varied enough, so you're unlikely to ever get great quality from this particular swap.

[kingrun2]

So basically I would need to add more varied facial footage of myself. (B) side I mean. And maybe in different light exposures for it be more accurate? Also Is there a possibility to add this additional footage to the current training?

Also is batch at 16 good for this card which is a GTX 1060 6GB card. Or should I change something ? I am using the Original Trainer preset.

[torzdf]

Yes, you can add more faces (stop training first, add more faces, merge alignments files, continue training).

Original model will only take you so far, but it's a good starting point.

More data in the A set wouldn't hurt either, but B needs more attention.

That batch size is fine.

Everyone's first deepfake is terrible