Closed JZArray closed 2 months ago
i think the training strategy is erroneous - i'm limited to my single gpu - 3090 the idea i read in another more recent paper - which makes more sense to me just to skip the frames ahead - and have the neural net work out the warping. overfit to one model. get to rotate head / face / eyes - completely disentangle to one subject. i put a bunch of updates here - https://github.com/johndpope/MegaPortrait-hack/pull/47 but I'm more interested to create new architecture. the game is to disentangle - but seems like there were perhaps better ways from many years ago. https://arxiv.org/pdf/2008.02401
also the es - leakage - i haven't solved.
i solved id leakage in this repo - https://github.com/johndpope/IMF
there can be a problem with the batch normalization where many images in batch contaminated things. https://wandb.ai/snoozie/IMF/runs/fyk2cfti?nw=nwusersnoozie
i create a seperate ticket.
same as the title