We have temporarily disabled registrations. They will be re-enabled within a few weeks.

Experiment - SAE vs H128

Which has better results?
I think the results of the SAE training looks better.
The eyes seem to be more accurate on the SAE model.
Pause at 0:37 - you might notice how the Face of the H128 is more at an angle while SAE the face is looking straight at the camera which is more accurate to the original video.
The left video uses a model trained in SAE for 270k iterations.
The right video uses a model trained in H128 for 400k iterations.
They both use Overlay and Learned Mask to convert them.

Notes:
- SAE seems to take longer per iteration
- 270k iterations of SAE to the about the same time as 400k of H128
- I'm not to sure what caused the color differences with both videos, maybe
it was color transfer