> no backprop, no autodiff, no training framework

OP says it like it's a good thing, but doesn't explain why. To not rely on Pytorch? What's the benefit?

Extremely cool stuff. I’m excited for neural approaches to further improve the bitrate frontier of “visually lossless” compression