New way to encode a profile picture dropped:

https://x.com/Ethan_smith_20/status/1801493585155526675

320 bits is basically a hash. Small enough to go on chain for every user.

weird thing is that the 256 token row looks worse than the 32 token row

It seems logical that prettiness is higher with fewer tokens. After all, if you do down to 0 tokens you're basically running a random image generator based on the training data, and pretty images are disproportionately more likely to make it into datasets. The more tokens, the more you compromise with reality.

never thought there could be an Uncanny Valley of image compression, but yeah this makes sense