DF Community model tests- test #1 Batch size, all 5 clips comparison/compilation.
Once you read this and watch all videos please vote for which one you think looks best here:
Here is what we did, we’ve used the same model, src and dst datasets, we’ve applied the same workflow to how we trained those models, only difference between them is batch size of the model. To ensure each model had equal chance to learn properly we didn’t train each model for the same amount of time but rather trained it based on epoch targets.
One epoch in simple words is done when a model processes all of the faces from a training dataset, in this case an average of SRC and DST dataset image count, since each model used different batch size it would process different amount of faces in a given amount of iterations.
– model trained at batch size 6 would do 6000 faces in 1000 iterations
– and at batch size 10 it would do 10.000 faces in 1000 iterations.
We’ve trained highest batch size model on each stage until we saw loss values stop improving and then we’ve calculated epoch/iteration targets for other models.
Here are full versions of all clips so you can watch them in full quality and resolution too.
Clip 1 and 3 by me:
Clip 2 by @almoosedalice https://msdeepfakes.com/user/145482
Clip 4 by @swing123 https://msdeepfakes.com/user/82429
and Clip 5 by @threecaptain https://msdeepfakes.com/user/38333
Here you can vote for which one you think came out looking best: