Two different one shot WMF cd cover art generative image resynthesis experiments above and below for the
"Over the Edge" cd release. Above is using the retrained CLIP image embedding guidance for the Stability model, below is using the Versatile Diffusion mode just released this week. Fascinating the differences between the 2 different approaches. You really need a dial to adjust between the tightness vs the abstraction, depending on what you want to achieve using this kind of model. At the rate things are moving in this area we'll probably see that in a few days.
The original WMF "Over the Edge" cd cover artwork used as the one shot input is show below for comparison.
5 comments:
john: supercool, i repost "the edges" in my "on the other hand"-blog, max
Dear john,
It is probably no longer justified to develop software that draws by AI, especially with the arrival of giant companies like Google and ...
Have you ever thought that maybe it is possible to create a software that recreates artistic behavior with artificial intelligence, so that it doesn't create paintings like others, but by utilizing the knowledge of artists' behavior, it makes its own creations. Inventing new drawing behaviors, for example, can enable it to create something that was previously unknown.
A.I video:
https://makeavideo.studio/
Make-A-Video with text
Bring your imagination to life and create one-of-a-kind videos
Max, thanks for the repost.
Re: 'no longer justified to develop software that draws by AI'.
Well, that is certainly one way to view it. However, here are a few things to think about.
1: The whole point of the Stability Ai push is to open source large models that can then be fine tuned for specific use cases, or repurposed into novel uses. So that large closed companies like Google or OpenAi/Microsoft don't have a lock in on the uses of large models.
2: 'draws by Ai'? Studio Artist actually does draw using Ai, at least the paint synthesizer part. None of these multi-modal neural net generative image synthesis models have anything to do with drawing. They are the next generative of image compression-decompression systems. There is a big distinction in my mind. I'm happy to elaborate, but think through how they work, what is going on under the hood.
We could certainly make Studio Artist better at it (drawing), using the latest and greatest deep learning ideas. But how that would play out is very different than what is going on under the hood of a latent diffusion system.
So maybe that is a lead in to what you then bring up. Creating software that recreates artistic behavior with AI. Yes, i have been thinking about that a lot. I have some ideas, but i'm very open to hearing other peoples options on it.
My post on the Studio Artist User Forum about 'gesture' and 'movement' is one small part of that thinking.
https://studioartist.ning.com/forum/topics/the-role-of-gesture-and-movement-in-digital-art
Ai systems that generative build art processes is another angle on it. Gallery Show is a primitive attempt at doing that kind of thing in some respects. But think about how Codex works, then think about building art processes.
I should have typed 'next generation of image compression-decompression systems' above.
Post a Comment