Extraction to exact input resolution.

Want to know about the Faceswap's Face Extraction process? Got tips, ideas or just want to learn about how it all works? Then this is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved for Extracting and preparing face sets for training a model in Faceswap.

If you have found a bug are having issues with the Extraction process not working, then you should post in the Extract Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Post Reply
User avatar
tokafondo
Posts: 31
Joined: Mon Dec 16, 2019 1:43 pm
Has thanked: 10 times
Been thanked: 4 times

Extraction to exact input resolution.

Post by tokafondo »

When (re)extracting faces before training, a resolution can be set.

Does extraction to the exact resolution the model will be using when training, is different than extracting to a higher resolution?

If the model is set to have input resolution 192x192, will I save time (because of skipping the resize step) or enhance quality (because the source image won't be modified from the original, possibly worsening its quality) if I have my dataset also at that resolution?

Or it's the same, anyway, to have my dataset extracted at 256x256 and have my model set as input resolution something like 160x160 or 192x192?

Thanks.


User avatar
bryanlyon
Site Admin
Posts: 596
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 153 times
Contact:

Re: Extraction to exact input resolution.

Post by bryanlyon »

No, leave it at the default unless you have a very good reason to change it. See the extract guide on "coverage" if you want an explanation of why.


User avatar
torzdf
Posts: 1266
Joined: Fri Jul 12, 2019 12:53 am
Answers: 120
Has thanked: 39 times
Been thanked: 251 times

Re: Extraction to exact input resolution.

Post by torzdf »

Incidentally, default should now be 512x512.

Unless you are training a model with an insanely high input size, you should leave it as this. The cropping and resizing code for feeding the model is well optimized, so there are no savings to be made here, and you are more likely to get it wrong (thus forcing an upscale) than right, as there are numerous calculations that need to be made to work out the "correct" image size to input into a model.

My word is final


Post Reply