The place to discuss Faceswap and Deepfakes
https://forum.faceswap.dev/
During the step "Extracting a training set from an alignments file"
I went from around 600 extracted images to 30. Is that normal ?
Is 30 enough to train my model ? (I put the Batch Size of the training to 16) If not, should I use some of my 600 extracted images ? Or I should only use extracted training sets from alignement files ?
You will need to gather more images for training.
30 is not enough for training. 600 at c.30fps will be too similar/not have enough variation
Okay so I record myself in every angles for 20 seconds, that gave me 600 extracted faces. And with the process to "extract a training set" that reduce it to 30. (I don't really understand this step by the way, is it to delete similar faces that can't be used in the training set?). So should I record myself for a longer time ? In different environments / lights also ? To get more faces for training set ?
20 seconds at about 1 frame per second wouldn't give nearly enough data, either in variety or expression. You'll either need to greatly increase the extraction rate or record more data. I'm betting that a single 20 second video will never be sufficient and instead you'll want at least a couple minutes in a variety of lighting and pose conditions. Selfie style videos will almost never give appropriate data.
Hello everyone
I have a beginners question, which I think belongs to the topic.
So far extraction went fine, thanks to the guide. While sorting extracted faces I saw a few upside down ones, which is the same position the actor is in the scene.
My question is: do upside down extracted faces matter for training and converting? And while I'm at it, do extracted images where the actor is mirrored matter? Like if he sees himself in a mirror and the extractor rightly extracts a face.
Cheers
Upside down faces do matter. All extracted faces are aligned to be orientated upright, so if you see upside down faces in your extracted faceset, then these have been 'misaligned' and require fixing
Mirrored is fine. The training process actually randomly flips faces left to right anyway.