Comment by a-dub

this is interesting. would be cool to explore something like integrating a vlm to add a "semantic" term to the loss function. looking through the comparisons, some of the baseline codecs create meaningfully different details (as could be described by text) in the images.