5 ms·

My son is an artist and we have conversations around AI generated art almost every week these days. Our general consensus has been: 1. Artists that use SD etc.

by nvln 4y ago

My son is an artist and we have conversations around AI generated art almost every week these days. Our general consensus has been: 1. Artists that use SD etc., will save a lot of time and have a big edge. Example: Generating references instead of searching for them, Generate random inspiration(s) for abstract ideas, Variations of a particular idea to compare and contrast. 2. Non artists who have powerful imagination but lack the artistic skill will stand to benefit as well. For example, I look forward to generating visual companions for my poetry, essays or fiction. I have the choice of commissioning an artist if my piece is successful. 3. Artists that are purists and masters of their craft with a very unique style will continue to thrive in their niche. These people have to lobby and come up with a license that explicitly prohibits feeding their images for learning. But a skilled artist can replicate the style and feed it. There is an ethical line here that may require creative tweaks to existing laws to protect people that fall in this category. Super exciting tech and I can't wait for this to move over to 3D models. I know 2D image to 3D model is already pretty close to being real(nvidia ominiverse, nerf etc) and SD can be the starting point of that pipeline. Prompt -> 3D world that can be decomposed to meshes, textures and USD will change the game development landscape quite a bit.

mattmanser 4y ago
So I just tried beta.dreamstudio.ai, and am a bit underwhelmed.. It's compositions generally look silly or completely ignore the instructions, even if I turn cfg way up. I was generally generating 9 compositions at a time, here's some of the things I've tried: "An elf fighting an orc in a forest" Half the images had christmas elves so it's obviously bad at context, half looked like miniatures (models like games workshop), so it's obviously got a load of its data from pictures of miniature but doesn't realize and thinks that's what it should draw, and the "fighting" always looked pathetic. "A human berserker with an axe fighting 3 orcs in the desert" That had problems with actually putting a human in. Also no axe. Orcs designs were obviously stolen from LotR films orc design, the actual breath of orc designs in fantasy didn't seem to be represented at all. Often the weapons are for some inexplicable reason only part drawn. "A space elevator leading up from new york" Was a little more promising, but looks pretty amateurish, just a picture of a space elevator transposed over a picture of new york with terrible lighting, though it was choosing half decent perspectives. "A 12th century castle on a hill in a hellscape surrounded by demon hordes" Pretty much ignored everything after "castle", just a bunch of 12th century looking castles. "A castle on a hill in a hellscape surrounded by demon hordes" None of the 9 images generated has any sort of demons, it just coloured some castle a bit red. I think artists are safe for another few years.
- hbn 4y ago
  I'll link to my other recent comment here: https://news.ycombinator.com/item?id=32634074#32661992 https://news.ycombinator.com/item?id=32634074#32661992 This isn't supposed to be used the same as Dall-E mini. You're supposed to be very detailed in your prompts. Also the online demos tend to not let you tweak parameters. You get more options if you run the model yourself
- fleddr 4y ago
  Your experience is quite common, especially when new to prompting. You can get more control by checking prompt guides, but this still leaves a lot to be desired. You seem to have missed the main point of the article this thread is based-on though. With img2img, you can input a kid-level MS paint input and control composition in detail. With inpainting you fix individual parts, with outpainting the surroundings/background. With text inversion, you can reuse successful parts in any context. People are even making the first videos and animations. All of this is available today, not years from now. The workflows are still clumsy, but it's been a week. Many people are working on the UI to streamline all of this. You're severely underestimating progress.

by nvln 4y ago

mattmanser 4y ago

So I just tried beta.dreamstudio.ai, and am a bit underwhelmed.. It's compositions generally look silly or completely ignore the instructions, even if I turn cfg way up. I was generally generating 9 compositions at a time, here's some of the things I've tried: "An elf fighting an orc in a forest" Half the images had christmas elves so it's obviously bad at context, half looked like miniatures (models like games workshop), so it's obviously got a load of its data from pictures of miniature but doesn't realize and thinks that's what it should draw, and the "fighting" always looked pathetic. "A human berserker with an axe fighting 3 orcs in the desert" That had problems with actually putting a human in. Also no axe. Orcs designs were obviously stolen from LotR films orc design, the actual breath of orc designs in fantasy didn't seem to be represented at all. Often the weapons are for some inexplicable reason only part drawn. "A space elevator leading up from new york" Was a little more promising, but looks pretty amateurish, just a picture of a space elevator transposed over a picture of new york with terrible lighting, though it was choosing half decent perspectives. "A 12th century castle on a hill in a hellscape surrounded by demon hordes" Pretty much ignored everything after "castle", just a bunch of 12th century looking castles. "A castle on a hill in a hellscape surrounded by demon hordes" None of the 9 images generated has any sort of demons, it just coloured some castle a bit red. I think artists are safe for another few years.

hbn 4y ago
I'll link to my other recent comment here: https://news.ycombinator.com/item?id=32634074#32661992 https://news.ycombinator.com/item?id=32634074#32661992 This isn't supposed to be used the same as Dall-E mini. You're supposed to be very detailed in your prompts. Also the online demos tend to not let you tweak parameters. You get more options if you run the model yourself
fleddr 4y ago
Your experience is quite common, especially when new to prompting. You can get more control by checking prompt guides, but this still leaves a lot to be desired. You seem to have missed the main point of the article this thread is based-on though. With img2img, you can input a kid-level MS paint input and control composition in detail. With inpainting you fix individual parts, with outpainting the surroundings/background. With text inversion, you can reuse successful parts in any context. People are even making the first videos and animations. All of this is available today, not years from now. The workflows are still clumsy, but it's been a week. Many people are working on the UI to streamline all of this. You're severely underestimating progress.