Hacker News new | ask | show | jobs
by caterama 1443 days ago
I would love to see what comes out with certain aspects of the prompts negated.

- "lemon gelato that’s been shaped to look like a heart, on a handmade waffle cone being held up to the camera in a cobblestone courtyard somewhere in italy" ... what about "somewhere not in italy"?

- "Diorama made of clay of a group of computer programmers looking disapprovingly at their CMO who has just given them diet pepsi instead of mountain dew" ... "looking approvingly"?

- "friends gathering around a tabletop “shichirin” grill where an assortment of meats and seafoods are being grilled over glowing binchotan charcoal; everyone is happy." ... "everyone is unhappy"?

1 comments

Ok, here are some results:

1: "not in italy": https://ipfs.io/ipfs/QmWPbZnZL7mHazzMmGxx6wQjcbC2DdKtgpiYYUY...

2: "looking approvingly": https://ipfs.io/ipfs/QmNgD9niZy1n4KWSXS3HEFeZm2qxQ2SzESfA6z7...

3: "everyone is unhappy": https://ipfs.io/ipfs/QmcTktpFQGeGDAp7e3MwEPwwFMVMbDaEgW6PuSA...

I think getting good Dall-E results wil end up being an "art" in its own. Dall-E is like a broad brush and honestly I've never been good at getting great results. I think figuring out how to push Dall-E in a way that aligns with what you want with the right descriptors really goes a long way.

I think to get there, we need a good dictionary or wall of examples that tell you what you can even do. I didn't even know you could have it create clay dioramas.

It's interesting that the gelato's colour seems to have become the wall colour as well. What happens if you ask for strawberry gelato, or lime?
Looks like a big weakness of DALL-E 2 is mixing up the properties of every object and of the background/setting.

https://twitter.com/david_madras/status/1512573390896480267

https://www.lesswrong.com/posts/uKp6tBFStnsvrot5t/what-dall-...

Mew