Hacker News Reader

Comment by qwertox | Hacker News Reader

I've been trying to get some sensible images out of my descriptions, but I fail miserably.

In this case I had the prompt "cow chewing bone" with 4 squares representing the two pair of feet, the body and the head. None cared about chewing on a bone.

With DALL·E 2 I tried to get an image of a little girl building sandcastles and a monster threatening her:

"little scared girl building a sandcastle and a big angry monster is looking at her."

"little scared girl building a sandcastle six damaged sandcastles are to her side. a big angry monster is threatening her. it is dark." https://imgur.com/a/f5FFKOi

"little scared girl building a sandcastle with six damaged sandcastles to her side and a big angry monster threatening her"

Is there some kind of structure the sentences should follow?

View on HN

Yes, checkout examples on lexica or use a prompt builder to help, like promptmania.

Also, most of the good ones you see online are cherry picked from hundreds of runs, so set your batch size too 1000 and go to bed! After that, people then tend to run some of the good results through img2img, also with a lot of variations produced from a single image. Finally, some people also run them at higher resolutions if they have enough VRAM, as smaller resolution can distort or generate rubbish. For the messed up faces, they run it through gfpgan a few times to get prettier faces. Other than that, it is pure luck (using random seeds) to figure out what works and what doesn't. Use the 2 sites above to help you improve your prompts.

(meant in the context of stable diffusion)

Just know that if you let it run over night often, you will see it on your electricity bill. My GTX 1660 runs at max while rendering, which is 125W. Leaving it running over night can easily eat 2 to 6 kw's, depending on your system.

I managed to get one that was correct with "A little scared girl is building a sandcastle, while a monster is looking at her. Award-winning photograph.", but I couldn't figure out a phrasing where it wouldn't most of the time get confused thinking that the sandcastle is the monster, or that the girl is the monster.

Dalle is bad at being instructed to have an exact count of items in the picture. Ask for 6 kittens and you get 7 & each kitten will be much more "wrong" than a piture of a single kitten.

Dalle is is bad at positional prompts. Ask for somethi g to be in the top rightbhand corner and it will appear bottom centre

Reply on news.ycombinator.com