Scale-Agnostic VQGAN Take a pre-trained VQGAN and replace the CNN kernels with CKConv's fitted to generate the kernel's they're replacing Fine tune Should now be able to encode/decode arbitrary resolution images?
