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.
Can you use different size faces for the original and swap face sets, e.g. 256 for original and 512 for swap?
What actually happens when a higher resolution face set is used to train a lower resolution model, e.g. using 256 faces for dfl-h128 or 512 faces on sae set to 256? Does the face set get scaled down prior to being used for training?
Thanks.
Don't worry about different sizes. The model only takes in set sizes anyway.
swapration wrote: ↑Thu Sep 10, 2020 1:34 am
- Can you use different size faces for the original and swap face sets, e.g. 256 for original and 512 for swap?
No. All training images must be the same size.
- What actually happens when a higher resolution face set is used to train a lower resolution model, e.g. using 256 faces for dfl-h128 or 512 faces on sae set to 256? Does the face set get scaled down prior to being used for training?
They get downscaled. All images get resized for the model.
My word is final
Oh, okay so I guess there's no point in using 512 resolution faces on dfl-sae set to 256? Might even slow it down a bit initially because of resize?
It depends. If the coverage is 100%, then 256 extract will be fine. If the coverage is anything below 100%, then you'll want to extract at a higher resolution (maybe 384?)
My word is final
Thanks! I had initially thought that the model's resolution was 1:1 with the faces used - that the enter face was saved regardless of the % coverage set, with the level of detail based on the mask...
But it makes so much more sense now! That's why in the training guide it says higher coverage means less quality!
Hello, in other word if i extract at 256px and train using "Orginal", i just lost time in extracting procedure?
Thats all?
No, different extract size doesn't take different amounts of time.
