Hacker News Reader

Comment by jedberg | Hacker News Reader

I'd say I'm most impressed by the compression. Being able to compress an image 42x is huge for portable devices or bad internet connectivity (or both!).

View on HN

That is 42x spatial compression, but it needs 16 channels instead of 3 for RGB.

Furthermore, each of those 16 channels would typically be mutibyte floats as opposed to single byte RGB channels. (speaking generally, haven't read the paper)

Even assuming 32 bit floats (the extra 4 on the end):

4*16*24*24*4 = 147,456

vs (removing the alpha channel as it's unused here)

3*3*1024*1024 = 9,437,184

Or 1/64 raw size, assuming I haven't fucked up the math/understanding somewhere (very possible at the moment).

It is actually just 2/4 bytes x 16 latent channels x 24 x 24, but the comparison to raw data needs to be taken with a grain of salt, as there is quite a bit of hallucination involved in reconstruction.

I have to imagine at this point someone is working toward a fast AI based video codec that comes with a small pretrained model and can operate in a limited memory environment like a tv to offer 8k resolution with low bandwidth.

I would be shocked if Netflix was not working on that.

I am 65% sure this is already extremely similar to LGs upscaling approach in their most recent flagship

a 42x compression is also impressive as it matches the answer to the ultimate question of life, the universe, and everything, maybe there is some deep universal truth within this model.

Reply on news.ycombinator.com