sixgunzloaded wrote:And I don't recognize who 5ece7fbe1e6245a7b734bad3555267af.png is supposed to be, but she's gorgeous and that scene gets more perfect every time you update it.
It's gone through a storied history now. It's originally my drawing based on QSA's "Entombed", but the pose and frame is completely different. I've used it as a benchmark for what I can do with AI. Now that it's reached a new peak, I've actually repurposed it as a tribute to Acidtester's Safari Girl, as the theme fits. It's a very appropriate blended tribute to two cornerstones of the QS Art scene, with my original art, put through the modernisation of AI.
sixgunzloaded wrote:When you're asking it to muddy your subjects up, do you have to specify exactly where to put the muddy bits or can it figure it out on its own? How much of the fine details do you really have to be extra clear about when asking?
The spooky thing is that all the details are AI-generated. I can't really specify exactly where details are, as the AI can't recognise them (e.g. I can't say "torn left sleeve" or "muddy thigh"). I can edit the source at any point to paint details on to give the AI more clues as to what it should generate. However, everything in the last day's sets is 100% AI generated.
The big game-changer is the LoRA "Detail Enhancer". LoRAs are basically "masks" that add specific details to a generation from a much narrower set. The Detail Enhancer gets the AI to add in fine details, such as creases, for almost anything. With this, it's able to get things like buttons and uniform patches correctly detailed. You can set the Detail to be as high or as low as you want. Compare the following:
The top image is the plain output in Realistic Vision 5 (with my secret blend of prompts, of course). RV5 is already a huge step above RV2 and Lifelike I've been using last month, and no doubt it looks amazing.
But with Detail Enhancer set to 2, it's a glow-up. Everything in it is more detailed. Hair, skin texture and glow, the mud, the shirt creases, even the leaves in the background. That's been my secret to get the photographic faces and lighting. The returns diminish as you go higher - the bottom one "overdoes" the detail by making the shirt crinkly rather than creased. I normally set it to 1-2, which is how I've been able to get some of the intricate lace designs in dresses and lingerie.
Even on top of this. Mage Space's last update introduced a HD Fix function. There was already an AI upscale function to improve resolution, but the HD algorithm also automatically adds in details to make the upscaled image "better". What I'm seeing now with Realistic Vision 5 and the HD Fixer is that the AI is better able to "understand" the right look and expression. For many of these, when I say "worried" and "sinking in quicksand", it seems to have a better training set that associates the worried expressions with the peril scene. Very often, despite me specifying an eye colour, the AI _closes_ their eyes as it recognises the struggling pose and expression. In fact, I used to struggle with where the subject was looking, since it tended to look at the viewer. Now, it has more creativity in looking at something else on its own.
