my main thought after 100 DALL-E 2 attempts is i bet they would look less impressive if we had personally seen every image in the training data. it's hard to know how much of it is cribbing off of pretty close matches in the training data + some remixing, style transfer, etc.
Conversation
like i think the reason "digital art" gets generically good results is that human artists have just made a shitton of really good digital art on deviantart etc. and that's obviously in there somewhere even if i don't know any of the details of how DALL-E 2 was trained
1
1
32
"less impressive" is a relative term, it's still very impressive, and i think i'm learning some things about my personal visual aesthetic this way which is cool. i really like big-ass trees
Quote Tweet
"the world-tree yggdrasil towering over a beautiful futuristic solarpunk city full of lights with the milky way in the sky, digital art"
Show this thread
1
2
39
i feel like i have a relatively underdeveloped visual imagination and i like the idea of using DALL-E etc. to kind of compensate for that. find out what kinds of things i could imagine if i had a better imagination, maybe jumpstart my imagination that way
1
33
sooner or later there are going to be similar questions about DALL-E etc. about the extent to which these algorithms are laundering the work of human artists
Quote Tweet
github copilot is incredible. it just sells code other people wrote, but because it's an "AI" it is apparently allowed to launder that code without it being a "derivative work". lol. lmao. what an amazing grift.
Show this thread
Replying to
The extent is 100%, but isn’t that what all human artists do too? What would a book be like if its author have never read anything by anyone else? Probably pretty primitive.
4
