Judge Verity Gavèlle from Ace Attorney Investigations 2: Prosecutor's Gambit, aka Justine Courtney in the fan dub.
For Illustrious, gk2verity is the trigger word. For anima, use @gk2verity. The Illustrious LoRA was primarily trained using official art and HD sprites from the game, so official art is a good tag to use to replicate the art style.
For the Anima version, the official art tag was removed as the LoRA was trained solely on official art and HD sprites.
General Appearance Tags: brown hair, hair over one eye, crown braid, brown eyes
Outfit Tags: white cape, collared cape, black gown, red sash, long sleeves, wide sleeves,brown shoes
Due to the nature of the sprites from the Investigations games, outputs have a bias of having her body angled to the side. straight-on is good to use in prompts to get her to face forward.
Illustrious sample images were generated using WaiNSFW with a CFG scale of 2, 30 steps, and with the Euler A sampling method.
Anima sample images were generated using Anima-turbov1 with a CFG scale of 1, 12 steps, and with the ER SDE sampling method.
Description
FAQ
Comments (3)
Is Anima that much better? I tried before, but it seems a lot less flexible to me. And it takes over double the time to generate an image with the same sampler, scheduler and steps as Illustrious. Maybe I'm missing something that makes it the 'next best thing' though.
This is the first time I've tried Anima myself. I've been asked to train something in Anima before, but the buzz cost to train it on civitai was so high that I never bothered. Now that I can train locally, I figured I'd give it a shot and see what it's all about.
From the little testing I've done so far, I'd say Anima is by far the best looking in terms of visual fidelity. I think these results are pretty damn close to being one-to-one to the actual graphics of the game. I do agree that results can be a bit stiff and there were a few things it struggled to generate, but at the same time, I had better results prompting some concepts that Illustrious struggled to generate as well.
Prompting with the same settings I used in Illustrious did take longer, but I followed the recommended settings on the Anima checkpoint page and they came out significantly faster. I used anima_turbov1.0 with the ER SDE sample, 12 steps, CFG 1, and the Beta scheduler and those came out faster than my usual Illustrious results.
I definitely need to do some more testing with other anima checkpoints and settings to see how things work, but for my first shot at anima, I'm pretty damn happy with what I got. I don't see myself phasing Illustrious out in favor of it, but I'm definitely interested in re-training my older LoRAs in anima in-between making new Illustrious ones.
@DiegoUmejuarez Excellent breakdown. Thank you kindly.
And you're definitely right about the visual fidelity. Even with my less-than-spectacular experience with Anima in the past, I noticed the accuracy of the character's appearance, especially the clothing, was on point. And I noticed the eyes looked incredibly detailed compared to Illustrious-- which I found often had a weird, smudged inky look to them when the character was at a distance. But again, I always found the stiffness a bit... hard to get used to. It was especially weird for NSFW stuff for me, as male and female genitalia looked very unrealistic. It sort of felt reminiscent of using FLUX, if you ever used that all those years ago...
Also, that might explain the generation speed difference. I typically generate at 15 steps on Illustrious, but if you can get away with using less than that on Anima, then I can definitely see how it could be faster. You've certainly inspired me to play around with Anima again and give it a second chance, so thanks.
And hey, man, more variety/options is always a good thing. So, as usual, I appreciate everything you're doing by also giving people Anima.