Faceswap Forum

If there is only one subject in each video, is reducing the batch size meaningless for accurate person differentiation?

The guide says, "Lower batch sizes will train slower, but will distinguish differences between faces better."

Since there is only one person in each video, I'm wondering if it makes any difference.

If this is the case, then I plan to increase the batch size as much as possible.

It might make more sense to think of it as a "forest for the trees" sort of thing. Low batch sizes will get a single tree looking better, while high batch sizes will get the whole forest looking better. Larger batch sizes do train faster, but they can't necessarily train better. I did a test at BS=512 and found it basically didn't train at all. Best general advice is to start with as high of a batch size as you can (within reason: stick below a maximum of 128, if you can go higher, perhaps use a larger model instead) and then lower it as you train further. You'll get the hang of it as you train models, as it becomes intuitive.