See wiki for more details and further results.
cae_32x32x32_zero_pad_bin model, after roughly 5.8 millions of optimization steps;
left: original, right: reconstructed.
A dataset can be constructed by downloading frames using the scripts provided here. For the above results, I have randomly selected and downloaded 121,827 frames.
[1] Lossy Image Compression with Compressive Autoencoders, Theis et al.
[2] Variable Rate Image Compression with Recurrent Neural Networks, Toderici et al.