Evaluating text-to-image models is hard, but here are some potentially interesting dimensions:

Depth and breadth of knowledge
- easier: Icelandic wedding
- harder: 1940's wedding in Reykjavík's oldest theater

Compositionality for rare or unseen combinations
- easier: a scarecrow juggling purple birds
- harder: remorseful blue lasagna on a leather pentagonal table with furry elephant legs

Coherence of complex outputs
- easier: hands, letters
- harder: full website mockups, wiring diagrams

Migrating over from another server!

To inaugurate this account, here are some puns I got GPT-3 to write

"He found a cure for his alement at the local brewery."

"To be a master baker, you knead to know your bread."

"Keeping the royalty in check can be a real reign it in situation."

"Gardening is a great way to reseed into the background."