Faceswap Forum

Why do we need to add a new model whenever the target changes?

If you use a model trained on a different set of faces It wont work well.

The model learns Original(A) to Target (B) .
If you change the A face, the trained model will not exactly understand.
The resulting face will look not great, really blurry, and sometimes monstrous.

abigflea wrote: ↑Tue Aug 04, 2020 9:24 am

If you use a model trained on a different set of faces It wont work well.

The model learns Original(A) to Target (B) .
If you change the A face, the trained model will not exactly understand.
The resulting face will look not great, really blurry, and sometimes monstrous.

If keep using one Model dir, the quality problem gets better after some time.
But if I add a new model, there's a difference that takes a lot of time.
Isn't it a good idea to keep using the old one?

Are you asking about resuming a new model?
Or starting a new one like Dfaker , Lightweight, ect, in the same model folder?

abigflea wrote: ↑Tue Aug 04, 2020 11:59 am

Are you asking about resuming a new model?
Or starting a new one like Dfaker , Lightweight, ect, in the same model folder?

Same model / based on original train options.
Leave A(base) alone and change B image only.
From what I feel, this method is faster than training from the beginning.
Then can I consider not making a new model dir?

I don't think that will work very well.
Typically you will want to start everything anew.

Give it a try, let's see the results.

abigflea wrote: ↑Tue Aug 04, 2020 12:18 pm

I don't think that will work very well.
Typically you will want to start everything anew.

Give it a try, let's see the results.

I guess there's no case like me in the forum.
Thank you for your time.

I've seen people talking about trying what you are proposing, but never really seen the results.
This suggests, as I would guess, it didn't work well.

I'm aware some people use pre-trained models but expect they are unhappy with the results and end up spending the same amount of time making it look right.

Good luck!
Let me know how it goes!