Comparing model types with examples

[abigflea]

Thank you I'm currently working on some more but it takes absolute ages

[algeron]

This post has seriously helped me alot. I've got a DFLH128 model I'm super excited about now thats 93% perfect with in 50 hours of training.

When starting new models/projects, now I refer to those gifs I made of your videos all the time.

I agree, this is a very insightful thread. Thanks abigflea & cosmico! May your losses never spike.

[torzdf]

May your losses never spike.

[Al1d]

Thank you
this was very helpful

i'd really appreciate it if you gave us information about the hardware doing the job (GPU)

[Bluecolorbanana]

So can we have a hierarchy of models from best to worst. For better understanding.

[EW20]

After about a year I've discovered there isn't really one "best" model. Just like there isn't a best hammer, wrench, or screwdriver. Different tools for different jobs, and you also have to consider the time aspect and your hardware.

Will Realface at 192px give you a better result than a DFL model at 128px output? If you train them to the same point, most certainly. But is that difference in quality worth training for 200 hours instead of 75? And does that better quality even matter for the end swap?

I've adopted the following workflow and it seems to strike a balance between time and quality, but it's just what works for me. I have a 16GB Nvidia GPU and a really fast cpu, just to give you a ballpark of my hardware.

First, spend the time to get a really good source set of what you want to swap in. Try and nail everything and get as many angles as you can with good lighting.

For the target:
Extract at 512px
Sort and eliminate unwanted faces, manually check alignments and fix the obviously poor ones. Then delete the original extracted faces and re-extract from the alignments file.
Train using lightweight with a batch size of around 100. I'll get around 500eg/sec and will let this go for a few hours or until the previews are acceptable.
I convert, then see what the quick swap looks like. If the faces just don't work at all I won't bother going further. But if it looks promising, I note where the faces are in the frame and their size.
If the face isn't close up and smaller in size, I'll usually choose one of the DFL models at 128px output. I've not really noticed a difference between these and a high-resolution models in these situations.
If the face is larger with more close-ups, I'll move towards Villain or Realface as I find I get better details with these in bigger faces. With realface I go 128 in and 256 out to teach some upscaling. Villian is easier but takes a loooong time.

More often than not I gravitate towards the DFL models because I find that I get too impatient with the heavier models and won't always train to completion. In my view a lighter model trained to completion always looks better than a heavier one that's still got a lot of work left.