➼Obligatory self-promotion: consider supporting or commissioning me at https://www.fiverr.com/metalchromex/make-you-a-stable-diffusion-anime-character-model or https://www.patreon.com/MetalChromeX so I can continue making cool crap for you all to enjoy for free!
Two separate Patreon members recently requested that I do some Makoto Niijima images within about a week of one another. I thought about just borrowing someone else's model of her, but I thought it'd be more fun/interesting to try making a general-purpose style model. I'm glad I did, as the results are pretty good (especially the in-game 2D cutscene version)!
There are three versions:
An Illustrious LoRA only trained on Persona 5/5 Royal and Person 5 Strikers in-game 2D cutscenes (this is my personal choice/recommendation).
Update (3-17-25): I tweaked the Persona 5/5 Royal and Person 5 Strikers model to add additional images to the dataset and make it slightly less overtrained. This should, hopefully, lead to even higher-quality images!
A LoRA based only on the 3D in-game models (experimental, but it works and looks better than you might think!)
Update (5-29-25): Just for fun, because I heard a larger dim/alpha helps with 3D-based models, I added a 64/32 dim/alpha version. It isn't dramatically better, but it certainly doesn't seem any worse. I haven't extensively tested it, but I'd probably recommend it over the 32/16 version.
Update (5-11-25): I added an Illustrious version. Generally speaking, Pony does a better job of replicating 3D-based models. Still, with the right checkpoint, you may be able to get close like the Pony version. It'll still produce some decent 2.5D-like images, if nothing else.
An Illustrious LoRA that is a hybrid of screenshots of Persona 5/5 Royal's in-game 2D cutscenes and Persona 5: The Animation. This model was trained on over 150 images, about 70ish in-game and 80ish from the anime. It works fairly well, but I'd recommend the other two Illustrious versions due to their more consistent art styles.
Description
FAQ
Comments (7)
I've been toying around with this Lora for a bit now using low strength (0.4-0.6) on the iLustMix checkpoint, which is already very good at capturing the Persona style without Loras. Visually, there's a noticeable change in style that brings the art closer to Persona's like advertised, but it comes at a cost. I find the prompts far less responsive, requiring a much heavier hand to produce the expected results.
This Lora overrides a lot of my prompting decisions and makes it harder to set up images, at least in my tests so far. Prompts that worked smoothly using the base checkpoint refuse to cooperate at all when applying a light touch of this Lora, and that can be very frustrating. It also introduces new concepts that I struggle to remove. For example, the Haru images I posted introduce a badge with the number 5 for some reason, despite Haru in the games not having any badges on her chest, from what I can recall. This consistently showed up in all poses, and I found no way to get rid of it, even using negative prompts like "badge" or "number" and turning up their strength.
So it's a bit of a mixed bag so far. The Lora improves the quality of the final images but makes prompting far more frustrating. I don't know if this feedback is useful in any way, but there it is anyway. I really like what you've done here and hope you can refine it further in the future.
I'll continue playing around with this Lora and see if I can find ways to tame it a little. When it works, it works.
Interesting, thank you for your detailed feedback. I think that, at this point, I would probably update the Royal/Strikers model if I were to come back to this model. I also may have done too many steps on this model; a V2 would probably have about half the amount of steps. Let me put a V2 of the model on my to-do list.
As a quick aside, I usually use WAI-NSFW-Illustrious; I noticed I did not put it as a recommended resource on this page, so I have gone ahead and done so.
Done! I made some tweaks; please let me know how V2 performs compared to V1.
@metalchromex Oh wow, that was quick! Thank you very much. I'll give it a go ASAP and let you know how things turn out.
I've been toying around with this one in the early hours of the morning now, and I'm quite pleased with the results. The new version doesn't fix any of the issues I mentioned, but it does soften them somewhat and makes it less frustrating to work with. Here are a couple of my random notes and observations:
- My favorite thing about this Lora is that the style is consistent, which I couldn't quite nail with just the base checkpoint. This was always the most problematic when showing a lot of skin in an image, but this Lora has no trouble fixing that. This is probably the most important reason why I will continue using it. That, and the fact that the consistent style is so appealing.
- Trying to add "scoop neck" makes it add actual scoops (like the ones you'd use with ice cream) to the background, which the base model has never done in my experience. I find this funny and annoying at the same time.
- The base iLustMix checkpoint I use (I like its looks more than the recommended model, which I know is probably causing me more frustrations, but hey) has very strong knowledge of the major Persona characters, but the Lora overrides some of this knowledge with incorrect data. It has some weird, apparently in-built assumptions about character looks, such as Futaba's t-shirt having a black splatter pattern instead of the canonical red one (the base model gets it right 90% of the time. I've yet to get it right when adding the Lora), Chie from Persona 4 frequently getting a houndstooth pattern on the neck of her jacket, or Makoto wearing a black turtleneck with straps for no reason I can understand. It can be hard to counter these, and I can't really see why it does it in the first place. This is a less pronounced problem with the new version but still present.
- Prompting for this Lora requires stronger keywords, often strengthened several steps, to get the expected results. Illustrious is usually good about not needing this, so this feels like a return to Pony in some ways.
- My impression is that I will probably have to use other Loras with specific character looks in them with some of these characters (Haru and Futaba in particular) if I want to use your Lora, but that comes with its own share of problems.
Also, I posted a bunch of images generated with v2 of the cutscene model just now. They may appear under a different name than this comment since Civitai seems to be struggling to catch up with my username change from yesterday.
Thanks for the continued feedback. At this point, I'm unsure how/if all those issues can be corrected; there may be a way, but I am not sure. Again, I'd recommend trying a different checkpoint and see if that makes any difference.