I see many people asking about how long to train, iterations, batch sizes ect.
Thought I would share to give people an idea of results.
For this I was testing DFL-SAE.
This was input size 192, and batch of 6. Roughly 7K images on both for various sources.
I updated the extracted faces and alignments before the last one and you can see it get really clear. Teeth, eyebrows, eye color and shine, ect.
I've not tweaked any settings in the convert process as you can tell. Just defaults.
The best, bottom right, was about 300K iterations. Took about 5 days total.
Likely would have been much sooner if I had used better sources and paid attention to the garbage I had put in there.
I've also done swaps with less iterations, and marginal source images.