Jonathan Fly

A lot of progress! I'm a bit disappointed in V1 RGB quality but I knew the resolution was low. The background removal is impressive though. > It allows to change...

I only tried V2 for a few minutes. No bugs or crashes. The main thing I noticed is my green screen error is actually pretty consistent, in my case shifted...

The lack of exposure controls are why I stopped using this for the moment. It looks pretty good in an evenly lit room though.

That works, though what should really happen is that training is _resumed_ if the model already exists. Not sure how difficult continuing from an existing checkpoint would be to add...

It shouldn't happen if you've never trained the model? If you see a /TrainedModels directory just rename it for now (or whatever directory it is storing the models in if...

There's a big gap between typical CycleGAN datasets and a dataset of 1, so adapting this architecture to cover some of the ground in between would still be super cool!

Your examples used 250px size training images, though in some cases it was 1000px wide by 250px high. If I want to maximize quality at the expense of training time,...

With 16GB of GPU memory, the highest resolution output I have achieved is 667 x 413 (scale 9 or 10). Does this seem about right? I'd like to fit more...

Same, and also get it with tinyshakespeare and other data sets that should be plenty large. Happens in CPU and GPU mode.

I frankenstein'd an absurd way to do this. Restream the stream to localhost. Install OBS Studio Install NVIDIA Audio Effects SDK: https://www.nvidia.com/en-us/geforce/broadcasting/broadcast-sdk/resources/ Open OBS setup your stream to stream your...