Okay, I lied. Somewhat...or is it just stretching the truth, though? Just think of this as a mere continuation of my previous post: following my recent breakthrough with Stable Diffusion's Regional Prompter extension and the Hires.fix upscaler, I tried different LoRAs, applying each of them on a single
(NSFW) scene/setting. Each image was generated using the same, basic prompt, but each time adapted to reflect the subject character's design quirks (hair length and style; eye, skin and hair color; chest size, possible accessories; facial expressions) - unfortunately SD didn't obey all orders (some characters should be sitting but are actually standing) and in certain cases I even had to overly emphasize certain traits to get more accurate results (Rakshata Chawla's skin tone, which I had to correct afterwards using Photoshop, in the interest of full disclosure; I also had to add weights to Villetta's skin color, too).
You'll note how difficult it is to generate pics that follow CLAMP's/Takahiro Kimura's style to a T or come even close to it. Some LoRAs fare better, others worse - your mileage will also vary according to the diffusion model used (Yesmix, meinamix, aamAnyloraAnimeMixAnime, etc.). I understand that better results can be gotten through networks but that's just extra workload for a program that is already crash-prone when upscaling to certain image sizes and is working with...not the most amazing amount of VRAM around, to be honest. I found more LoRAs to play with (Monica, Dorothea, Nunnally and Marianne), but at this moment I have not yet decided whether I'll generate something...
but if I ever do, just check this post for edits and new additions.
EDIT:
Note: some LoRAs (Low-Rank Adaptation assets, essentially skinning files) are trained on visual material so specific it actually constrains the range of available results. The Marianne vi Britannia LoRA, for one, was extremely insistent on placing her
indoors, most of the time inside a wooden corridor. I wanted her either in a garden or inside a palace. You can see in the prompts just how much weight (in the form of parentheses, which indicates how much importance the software must assign to these criteria) I had to give to "outdoors" and "French gardens", before the LoRA and the diffusion model finally deigned to indulge me and move the character to an outdoors setting. Same-o for the Julius Kingsley LoRA:
absolutely impossible to get a side profile of Julius, only him facing the viewer - then I decided to add some chaotic/style/atmosphere LoRAs to experiment a little and vent my frustration. In both cases, the upscalers' eldritch-archwizard-magic did the rest (you can see the mind-boggling difference in detail between the upscaled, final versions and their original starting points:
#1 and
#2). Who the hell ever said AI image generation was a walk in the park? Certainly not for those who have a very clear idea of what the output
should be compared to what you actually are served with.
Feel free to use the images as you please (save commercial purposes, of course). Toodles.