So...I hope you'll pardon the thread necromancy, but I figured I could just post these here as they are exclusively Code Geass pictures and this sub-forum got defibrillated back to consciousness a couple weeks ago with the
Rozé of the Recapture trailer.
But...a disclaimer first?
These are all AI-generated by yours truly.
I'm aware of the controversy surrounding such software and how it can cheapen the hard work and creativity of blood-and-bone artists, but I promise this is a one-time thing. For those interested, I have appended the generation prompts and settings. All assets (models, LoRAs, upscalers, textual inversion embeddings) can be found at Civitai.
Click on every image to see their full version
This one here has perfect LoRA outputs (the characters are distinct from one another and do not suffer from accidental dimorphism), but the background (namely the window) shows a skewed perspective that is not shared by both Kallen and Milly. Bah, let's just pretend they're standing on an out-of-frame stairway, shall we?
So here we come to the reason why I'm doing this post at all: I'm celebrating a milestone that I've been trying to reach for a couple of months now: creating an output image with 1°) TWO different LoRA characters whose designs don't bleed into one another and 2°) successfully making Hires.fix upscale the base image, 3°) physically isolating the characters with Regional Prompter, 4°)
all of that without crashing Stable Diffusion, consarnit. Condition #4 was the hardest of them to fulfill, as Hires.fix liked to crash before I started fiddling with venv startup parameters. #3 took me a lot of time to properly master and was the true secret to prevent Stable Diffusion from mixing the character models in a digital blender. I will spare you the endless cortege of eldritch...
things SD produced before this.
Base image, in 768x512px. Note the blurriness (esp. the eyes) and note Kallen's right (left side for us) elbow: a blurry hand wrapped around it. Also take note of her
waist right beneath it - it'll be important for the next picture.
Same image as above, but processed through Hires.fix and upscaled to 1536x1024px, showcasing how Stable Diffusion adds details and smoothens flaws and other artifacts, even correcting contextual errors. Look at Kallen's right elbow: the hand has now become a textile fold in her sleeve - and her waist has been altered to show that missing hand! Also, Milly's smile has been improved, as have the eyes, too. The decorative patterns across the glass windows have been altered and given better definition. Very pleasant result. But it took me a lot of time to get this.
Same prompt as above, of course.
"But, Renegade - why just Milly, Kallen and Lelouch? Why not others?"
The reason is simple: their character LoRAs (character skinning models, if you will) were the best-looking and
most reliable ones I've found so far. It's a bit too easy to find character LoRAs on Civitai, Pixai or Huggingface that seem more fit for hentai material (e.g. gigantic breast sizes) than normal/vanilla scenes. I recently tried a LoRA for Viletta Nu that only spat oversaturated and blurry stuff - thus earmarking it for deletion; my Lelouch LoRAs also tend to be rather temperamental, often spitting out illustrations whose subjects are barely 75% Lelouch-like or feel like they belong in a different anime from a different era (like Amuro Ray). The mileage also varies depending on the diffusion model used and any (ambiance|style|detailing) LoRAs you pair together. There is a LOT of trial and error involved in this. I might try other characters like C.C., or others from different series, though. Someday, though; not today.
Before you leave, I bid you remember one fact: for every beautiful or, should I say "seemingly perfect", picture the AI generator produces, there are quite likely
dozens of very unsatisfying or outright failed ones hiding in the background or immediately banished to your operating systems' recycle bin. What you see posted here are essentially...
the happy accidents - I honestly have no other way to describe them. It's just that - beautiful coincidences where the stars aligned perfectly. There is a whole art to
harnessing roughly guiding Stable Diffusion's in-built randomness and making the best of it. There is a
stupendous amount of trial and error involved in image generation (finding the right settings, the right prompt, the right negative prompt and the right assets!) and, given that I'm quite a bit short on free time nowadays, I'll be putting Stable Diffusion and both its wonders and horrors on the backburner. So...please: don't forward me requests like "can you generate a pic for me that shows yadda yadda yadda?". It's very time-consuming (especially for my aging PC) and sometimes even frustrating. But when you hit that rare unicorn? Yeah, makes it
almost worth all the effort. The rest of the time, your prompt is more likely to accidentally summon Nyarlathotep or Y'golonac through your screen.