coyotecraft
Sponsor
Scratch all that. Forget Dall-E 2.
Dall-E 3 is out; see posts below.
I thought I'd share my experience.
$15 for 115 credits = 460 images.
Sounds like a lot. But nothing it generated was practical. It was all crap. I think Dall-e is better at realistic photo manipulations and impressionistic oil paintings. I guess there are other AI models out there that are better for anime and digital art styles.
I knew after the first 15 generations that it was a waste of money. It was a matter of trial and error to figure out text prompts to use to get desirable results. I'd try "single point perspective", "no contour lines" but I don't think it has a grip on what that means.
It's completely hopeless at pixel art. I mean, nice pixel art. After a couple attempts it produced something that might have been drawn in MSPaint with basic colors. Which is probably the kind of data it was trained on. It looked like it was trying to generate gibberish signatures along with it. Like "creadtaL: Ciyitoe". So yeah. I think it was just sampling image search results.
I was skeptical about the Dall-E's "out painting" features. I tried it with my own art. I had an unfinished airship drawing, so I thought I could get it to add some kind of metal texture or pattern to it. But that didn't work. It sort of just distorted it with some filters. Idk.
I experimented with prompts like "isometric" "Top Down 3/4th Projection". And started getting results that were almost-game assets made from clay or something. It's eye-catching if you zoom out and squint. Lol. So for generating game assets, it might be good for making out-of-focus parallax backgrounds and textures. But again, I think it's got to be trained on some kind of existing library of images. I bet it could generate a decent forest landscape, because there are images like that in abundance. But if I asked for something more unique like the "inside of circus tent" I'm going to get something that's made of melted plastic and spider legs. In one case, I did get the inside of a circus tent, but it also had a bunch of smaller tents inside it.
Some of the images were not even close. Like, I asked for a "Gym with treadmills" and I got an array of bizzare objects I could probably turn into sci-fi ships or something.
After a couple hours of looking at stuff that almost looks like something, but isn't quite anything, it felt weird to look at actual stuff again.
Dall-E 3 is out; see posts below.
$15 for 115 credits = 460 images.
Sounds like a lot. But nothing it generated was practical. It was all crap. I think Dall-e is better at realistic photo manipulations and impressionistic oil paintings. I guess there are other AI models out there that are better for anime and digital art styles.
I knew after the first 15 generations that it was a waste of money. It was a matter of trial and error to figure out text prompts to use to get desirable results. I'd try "single point perspective", "no contour lines" but I don't think it has a grip on what that means.
It's completely hopeless at pixel art. I mean, nice pixel art. After a couple attempts it produced something that might have been drawn in MSPaint with basic colors. Which is probably the kind of data it was trained on. It looked like it was trying to generate gibberish signatures along with it. Like "creadtaL: Ciyitoe". So yeah. I think it was just sampling image search results.
I was skeptical about the Dall-E's "out painting" features. I tried it with my own art. I had an unfinished airship drawing, so I thought I could get it to add some kind of metal texture or pattern to it. But that didn't work. It sort of just distorted it with some filters. Idk.
I experimented with prompts like "isometric" "Top Down 3/4th Projection". And started getting results that were almost-game assets made from clay or something. It's eye-catching if you zoom out and squint. Lol. So for generating game assets, it might be good for making out-of-focus parallax backgrounds and textures. But again, I think it's got to be trained on some kind of existing library of images. I bet it could generate a decent forest landscape, because there are images like that in abundance. But if I asked for something more unique like the "inside of circus tent" I'm going to get something that's made of melted plastic and spider legs. In one case, I did get the inside of a circus tent, but it also had a bunch of smaller tents inside it.
Some of the images were not even close. Like, I asked for a "Gym with treadmills" and I got an array of bizzare objects I could probably turn into sci-fi ships or something.
After a couple hours of looking at stuff that almost looks like something, but isn't quite anything, it felt weird to look at actual stuff again.
Last edited:
