Read the FAQs and search the forum before posting a new topic.
This forum is for discussing tips and understanding the process involved with Training a Faceswap model.
If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.
Please mark any answers that fixed your problems so others can find the solutions.
Hi.
When searching for it, I found that the realface model is relatively rarely mentioned in the forum. Can somebody eleborate on how it compares to the others in training time and quality? I have seen (subjective) rankings such as dlight > dfaker == villain > original, so I wonder how it would compare.
Does it make sense to raise the input size with this model or would it not improve quality very much, as is the case with some other models?
Thank you in advance.
It's a long time since I've used that model, but I seem to remember I got good results from it.
I would put Villain at the top of the list, as long as you're prepares to wait a week or so for it to train.
My word is final
Realface seems to be pretty dang good.
I've been meaning to do a comparison between the two models and take them to completion I just and I had the time maybe you should run it for a week or two and see what the results are
I dunno what I'm doing 
2X RTX 3090 : RTX 3080 : RTX: 2060 : 2x RTX 2080 Super : Ghetto 1060
I've always personally viewed it as villian>realface>unbalanced>dlight>dflsae>dflh128>dfaker>original>IAE>lightweight
That being said, I never use villian or realface, dlight, because even though I think they are "more realistic"
what ever realistic bonus is not worth on average taking on average 3-6x times longer to train.
Unless you are the person who has a powerful computer and you are willing to just let it train for several weeks,
I'd have to guess you might get a better product if you had a medium or low quality model
and let it train to its true completion rather than a high quality model that's not trained enough.
Personally I'm finding dflh-128 to give almost perfect realistic results on 720p and 1080p videos and good results
for 2k and 3k VR videos with 60 hours of training on a Geforce RTX 2060, using high quality data
Oh and for context, I'm calling viallian, realface, unbalanced, as high quality models from high to low, dlight, dflsae, dflh128 as medium quality from high to low, and dfaker, original, IAE, Lightweight as low quality from high to low
