A style lora trained on the small amount of footage from Danganronpa 2x2 that's been made available as of December 2025. Starting out with a Chroma version but I should have it working with other models over time. Next up is a Z-Image version. For some of the technical issues, potential usage problems, and how I decided to go about putting the lora together see the Technical Issues section below.<h1 id="usage">Usage</h1>It should work with either tags, narrative prompts, or a combination of both. I tend to go with the combination route.You should be able to use this to trigger it: danganronpa_2x2, danganronpa_2x2_(style), The image uses the danganronpa 2x2 sprite style.It's heavy redundancy in the context of just this 2x2 dataset, but I wanted to maintain finer versioning options for later inclusion of this lora's dataset into my danganronpa locations lora.You can also try to use the Danganronpa 2x2 GUI elements with something like this: danganronpa_2x2, danganronpa_2x2_(style), The image uses the danganronpa 2x2 sprite style. The image uses the danganronpa 2x2 GUI style. Game GUI elements are present including a black bar at the top with white text reading CHAPTER 1 BGM Chroma Beats LEVEL 1. A blue text box is at the bottom containing the text "Testing Monokuma generation". A smaller blue box overlaps the text box on the upper left with text reading "Monokuma".<h1 id="characters">Characters</h1>I was going for style over being able to generate specific characters. But in the Chroma build of this lora Monokuma can be prompted for surprisingly well. Not always perfect, but more often than I would have expected. Chiaki can more or less be prompted for as well if you don't mind some inconsistencies. Examples of both should be in this lora's gallery. However, I haven't done much testing in terms of how flexible they might be when moved too far away from the limited samples present in the training data.<h1 id="lora-stacking">Lora Stacking</h1>So far I've only tested using this in combination with other loras using Chroma. And the testing has been somewhat limited. But for the most part it seemed to play nicely with some of my Danganronpa character loras using this lora at a strength of 1 and the character lora at a strength of 1.<h1 id="technical-issues">Technical Issues</h1>The biggest challenge is just dealing with how little footage of Danganronpa 2x2 has been released. In terms of promo video runtime it looks fine at first glance. But there's very little actual variety in what they show. A small number of characters within a limited number of environments and not a lot of different presentations of those elements with variation in context, day/night cycle, or scale.I thought about using various tricks to help compensate for the problem but ultimately decided to keep this lora as pure as possible with only authentic Danganronpa 2x2 data. The plan is to then merge this dataset with some of my other existing stuff, in particular the Danganronpa location lora, for usage cases when a more generalized 'danganronpaish' idea is more important than going with what's found in a specific game.The big issue with this approach is convergence on a more limited representation of some basic concepts. For example, a male character with an ahoge will have natural drift into looking like Hajime. Just because Hajime has a comparitely larger amount of representation in the existing Danganronpa 2x2 footage.With the Chroma lora at least I noticed what looks like possible signs of lighter concept bleed. Like solid clothing colors having a higher chance of additional red accents. But it seems like it's at an acceptable level given the payoff of better overall styling. I might upload an earlier checkpoint if that issue becomes more blatant with further testing. But in general I'm pleasantly surprised at how well this worked out given the constraints. It's not perfect by any means. But it turned out far better than I'd expected. So far with the Chroma version at least it seems to drift into more of a Rui Komatsuzaki art style over Danganronpa 2x2 specifically if the prompts move too far from something present in the Danganronpa 2x2 promo footage. But again, that's still better than what I expected to get.

Files with SHA256: 129f0d68e33ce9512476d431adfc1577f13a8c8505edcbe762781c821e91cfbe