Stable Whaaaaat?!
Yes I promise, it is stable.. ish
Artificial Intelegence?
So after some playing around and testing over a couple of weeks I think I may have (somewhat) found a way to more or less get what I want out of the AI.
The Fun Stuff
So as you start to understand(more or less) how the diffusion and text models interpret your words it will all come together and you can start writing prompts and tweak some settings to make some super cool things...Not So Fun Stuff
While playing around with a lot of settings and models it can get quite annoying and tiring when the produced result is just a tinnyyy bit off. Which leads to some wacky creations...Whatcha usiiing?
As many of my previous AI posts, I use Visions Of Chaos with a bunch of AI models; for text to image generation I typically use StableDiffusion but in the past I had used VQGan of DALL-E.
Okay, but gimme the settings bro...
Well, that is easier said than done, for the most part its not just a one-size fits all solution, its more a duct-tape kind of solution (for now). But here are some tips:
Big Yesses
For the most part, any prompt will create something more or less decent, but just like regular art you need to carve it out. When giving it prompt try to be as specific as you want, you want the image to look like a digital painting, cinematic, have a dark ambience etc. - then specify it.Big NO NO's
AI is not some magical tool that just knows what you want it produce, it can only do what you tell it to do. So the more vague you are the more it will struggle and the results will look poor.Some Examples
The images below are prime examples, the images on the left showcase barely any descriptors while the images on the right contain far more descriptiors. All the artworks are generated over 500 iterations*NOTE FOR THE FIRST TWO I FORGOT TO NOTE THE SEED*
a fantasy landscape
thresh combined with pyke from league of legends hooking a soul with his hook seed:296945706
a fantasy landscape featuring mystical plants area landscape hyper realistic detailed fantasy dramatic cinematic
thresh combined with pyke hooking a soul from league of legends, league of legends, detailed, cinematic, dark, digital artwork, concept art, ghostly hands holding the hook, armoured skeleton, thresh, pyke,cinematic lighting, highly detailed, illustration, artstation, digital painting, art by Bob Kehl, Terry Moore, and Greg Rutkowski seed: 1664105712
Okay then..
Yep, as you can see adding more details and descriptors can really take an image to the next level. But ofcourse as it stands, AI is a tool not a magic one fit solution.As for my settings, in VoC this would be it:
I just have quite a lot of HD passes on the image to take it from a blurry 512x512 to a nice crisp and detail 4096x4096
Conclusion
After using VQGAN and a couple of other models such as disco I can definitely say StableDiffusion is the current best model for getting more natural and coherent results.. for the most part sometimes you are left questioning where you went wrong like this:
What?!