Kinky Desires wrote:OMG wow! Thank you very much for that info. I was wondering how Viridian had used some of AcidTester's art to make AI photos, and when I looked into it all I could find for using a reference photo were special 3rd party AI plugins for software like photoshop, which I don't have - I'm very oldschool and still use PaintShop Pro
I use the Stable Diffusion models on Mage.Space.
Acidtester's art is challenging to work with because of the complexity in its layers and content. "Safari GIrl" was almost a 1:1 render using the illustration, photorealistic prompts and a photorealistic fine-tuned Stable Diffusion model. Others, like "Officers Down" and the Supergirl ones took a lot more to put together.
Often with these source images, I need an intermediate step to get the AI to not only up-scale the quality of the image so that it isn't pixelated, but also get the shapes and colours clear enough for the AI to recognise with prompts. One of the benefits of Mage.Space is that it allows you to switch between fine-tuned models, so I use Lyriel (or DucHaitenAIArt, or DreamShaper) to turn the drawing into a flatter digital illustration. I may have to use NovelAI if Stable Diffusion can't recognise the image at all. I just _something_ that looks like the desired output and I can render it through SD from there. Then I change my prompt to use more photograph and photorealistic prompts and refine my render with Lyriel and/or Realistic Vision models
Taking his drawing of Shaun Safari:
My first attempt at rendering it as a photograph through Realistic Vision doesn't generate a good result because of the low resolution and scratchy detail.
Hence I need some intermediate steps to transform the source image into a clearer source. Using the Lyriel fine-tuned model, I use a digital painting prompt on low strength to get the AI to work with the drawing.
Good enough. I then slot in my photorealistic / photographic prompts:
Then I switch to Realistic Vision and run the image several times to get it looking more like the photorealistic output I want. I may further refine the graphic by using inpainting to run specific section through other fine-tuned models, such as changing the face or mud textures, until I get something like this:
The prompt used for this, using Stable Diffusion (Realistic Vision V2):
(extremely detailed 8k), (ultra high resolution), ((photorealistic)), ((masterpiece)), sharp focus, beautiful and detailed lighting, shadows, ((skin details, high detailed skin texture, 8k hdr)), cinematic lighting, digital photograph of jungle explorer in quicksand, brown hat, detailed brown braided hair, detailed grey ((dried mud)), jungle background, scared shocked concerned face open mouth shouting, muddy khaki blouse