<p>Short gist: Lora trained on the various Danganronpa anime. Version 1 of the lora is more averaged out between entries. Version 2's training data has Danganronpa 2.5: Nagito Komaeda and the World Vanquisher as the majority of the data and sprinkles in others at around 10 to 15 percent of the dataset. With 2.2 the ratio is about 70% Danganronpa 2.5. The 2.2 ratio seems like a good balance but not enough of an improvement to be worth retraining the previous models with. </p><p></p><h1 id="important-for-z-image-turbo">Important for Z-Image Turbo</h1><p>If you're seeing poor quality images when using this lora try increasing the amount of steps you're using. For whatever reason I've noticed that this lora seems to need much higher step counts to get acceptable results. I'm also in the process of slowly creating z-image versions of my existing character loras. In particular the danganronpa ones. While tweaking and updating the datasets and releasing them for chroma as well. So that should be coming as a steady trickle.</p><p></p><h1 id="important-all">Important - All</h1><p>Read the version/platform usage information. I train on the same images for all of them, but differences in models and prompting formats means that using them differs tremendously on a case-by-case basis. For example, you can pretty much just toss the Illustrious lora in with the keywords and you'll be good to go. But Chroma works best by mixing and matching other loras at specific strengths.</p><p></p><h1 id="what-is-it-in-theory:">What is it in theory?:</h1><p>An attempt to get an even blend of styles between the anime Danganronpa the Animation, Danganronpa 3's Despair Arc and Danganronpa 3's Future Arc. But with special emphasis on Danganronpa 2.5: Nagito Komaeda and the World Vanquisher as it's got the most visually "normal" Danganronpa feel along with a larger range of characters from the games. I mean it has Sayaka doing a toast run, Ryoko, and even Aiko in outdoor lighting (for about three frames)! While the Future arc is weighted down by odd lighting...almost everywhere. And Dannganronpa the Animation for obvious reasons is centered heavily on specific indoor areas. This is very much a work in progress. My initial release was pretty underwhelming to me at first but as I got used to the quirks it grew on me. The 2.0 version is a nice leap forward, especially with the Illustrious version, but I'm positive there's still a 'lot' of untapped potential. This is more a good start or foundation that I want to build on than anything close to a finished concept.</p><p></p><h1 id="what-is-it-in-practice:">What is it in practice?:</h1><p>The Flux lora's fine as a general anime style booster with strong thematic elements from the Danganronpa anime. The 2.0 for Illustrious seems to be a LOT closer to reaching my original intent even if it's not as close as I'd like.</p><p></p><h1 id="which-loramodel-combination-should-you-use">Which Lora/Model combination should you use?</h1><p>Illustrious: The best jack of all trades and easiest to use</p><p>Chroma: Strong style and instruction following but more difficult to produce Danganronpa characters with due to limited in-model understanding and finicky compatibility with Flux Danganronpa character loras.</p><p>Lumina: Lowered anime style quality but high built-in compatibility with Danganronpa character tags and strong instruction following.</p><p>Flux: Similar to Chroma but full compatibility with the existing Flux lora ecosystem including my Flux Danganronpa character loras.</p><p></p><h1 id="usage:">Usage:</h1><p></p><h2 id="z-image:">Z-Image:</h2><p>I suggest using a higher number of steps than usual when using this lora. For prompting, narrative captions work best but I've gotten surprisingly good results with just tags as well. As with the other models the results of combining a narrative caption prompt with tags can be helpful as well. Though it doesn't seem to have as drastic a difference as with some of the others.</p><p></p><p>You should always use a tag of "danganronpa_3_(anime)_(style)" with the z-image lora. And I generally like to follow that up with "This is an anime screencap from Danganronpa 3. This is an anime screencap from Danganronpa 2.5." to sample from both the main danganronpa 3 anime and the 2.5 special. While the huge list of 2.5 specific tags do work I've found that this more concise way of applying the styles works just as well with z-image.</p><p></p><p>I also find it helpful to give z-image a little more help with the specifics by adding something like this to the prompt: "The art style is a highly stylized anime aesthetic characterized by dramatic high-contrast cell-shading and distinctive hyper-detailed character design."</p><p></p><p>Like with the other models some of the characters that were highly represented in the anime are 'mostly' promptable with this lora. There's usually going to be some issues. Seiko's mask not being quite right, Miaya's wheelchair, Chiaki's hairclip, etc. The Hope's Peak uniform does seem to consistently work however. Though even with characters that came out pretty well in the lora it can be a throw of the dice. Sometimes it might work out well, other times they'll be hardly recognizable compared to the official character design. Also like the other models it's better to try combining an existing character lora with this style lora rather than just hope training for the character stuck in this one. But the gallery for this lora should have some examples of successful prompts for some of the trained in characters that don't rely on any other loras but this one.</p><p></p><h2 id="lumina:"><strong>Lumina</strong>:</h2><p>The Lumina version was my first attempt at a lora for the platform. It's trained on NetaYume Lumina and should work best with it or variants like Neta Cat Tower. The lora's usable, but I'd classify it as overly finicky and unpredictable. I've seen results with the lora range from negligible to great. Consider it a work in progress. On the plus side NetaYume Lumina already has a broad knowledge of Danganronpa. Really, the main change there brought by the lora is just that characters in Danganronpa 3 but not well represented on Danbooru come out more faithful to their intended design.</p><p></p><p>Lumina has a distinctive prompting style that you should aim to use. With this lora, try something along these lines:</p><p>You are an assistant designed to generate anime images based on textual prompts. &lt;Prompt Start&gt; Tags: anime_screenshot, official_artwork, official_art, modern_anime, danganronpa_(series),super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono , anime screencap, danganronpa_3_(anime), photoshop_(medium), detailed_background,danganronpa_3_(anime)_(style), Detailed Prompt: INSERT NARRATIVE DESCRIPTION OF IMAGE Style: INSERT NARRATIVE DESCRIPTION OF STYLE</p><p></p><p>You can add your own tags after the tags in the "tags" section of the prompt. Write out a narrative description to replace the "INSERT NARRATIVE DESCRIPTION OF IMAGE" text. And write a more detailed narrative description of the style to replace "INSERT NARRATIVE DESCRIPTION OF STYLE". Though the style in particular, but really any of those structures, isn't an absolute necessity. The gallery should include some examples where I mostly just used tags or where the description part carries most of the heavy lifting. The model uses gemma for text processing so it's pretty flexible there.</p><p></p><p>You can also include this sentence to push to a more in-game style: "The image is in danganronpa sprite style." along with an extra tag of "komatsuzaki_rui_(style)" In theory at least.</p><p></p><p>In general Lumina as a whole seems very sensitive to style prompting so try playing around with the descriptions and pushing it in different directions using artist tags. The gallery should have some different examples of style prompts but my tentative goto right now is: " The image is rendered in the high-definition cel-shaded animation style of Danganronpa 3 featuring sharp clean linework and dramatic lighting with strong highlights and deep shadows. It uses saturated yet muted color palettes for atmosphere and meticulous attention to texture on clothing hair and glassware. The overall aesthetic is polished studio-quality official anime artwork that blends psychological thriller ambiance with danganronpa's signature visual flair."</p><p></p><p>Again though. This was my first Lumina lora and I consider this lora more proof of concept than a finished product. I'm hoping my next attempt at iterating on it will have more consistent results.</p><p></p><h2 id="flux:">Flux:</h2><p>The 1.0 Flux lora wasn't trained to recognize specific danganronpa anime, just the full style in general. But if you're using this with one of my danganronpa character loras, you'll generally be able to push to something more specific with either "This is an anime screencap from Danganronpa 3." or "This is an anime screencap from Danganronpa the Animation." due to those loras having more precise captioning. Also, with the 1.0 flux version, it's worth pushing up the strength of the lora a little higher than normal, possibly up to a strength of 1.3. I've mainly been testing it with the beta builds of my Tsumugi lora, but with that at least lowering the Tsumugi lora's strength to around 0.7 and raising this style lora to 1.3 seems like a good ratio to push her past the standard Danganronpa V3 sprite look. I expect that the same would hold true for other character loras. Of course the downside is that high lora strength tends to throw away some of what makes Flux unique. Higher the strength, the worse the hands and more garbeled any text will be. You'll notice both issues showing up in some of the images in the flux lora's gallery.</p><p></p><p>With the 2.01 Flux lora...I honestly don't even know. I'm still trying to figure out what works best, though honestly the tag experiment wound up being more trouble than it was worth and it's probably best to just wait for a 2.1 build of the Flux lora that should leverage the captioning approach I used with the experimental Tsumugi lora. That seemed to play better with both tags and narrative descriptions without forcing either. When using the 2.01 Flux lora with other character loras I'm seeing preliminary good results with this style lora at 0.7 strength and the character lora at 0.5. In general the rule is to keep this style lora significantly higher than the character lora. If a character 'is' in this lora's training data then there's a good chance you can really push the strength of a character lora down too. Miaya for example looks nice, but not 'quite' the way she should, with her character lora at 0.1 and this style lora at 1.0 strength. But at 0.3 for the character lora she looks pretty much as she should. With the 2.01 Flux lora you can also push for a more authentic anime style by going with a 16:9 aspect ratio. So something like 1920 width and 1080 height should push the results further to the anime data the lora was trained on. The Flux 2.01 lora is also VERY sensitive to styles given in the prompt. By default I use the tags " danganronpa_3_(anime), danganronpa_(series), super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono, anime_screenshot , anime screencap, anime coloring". I've found that I get pretty good results with further descriptive text of "Shallow depth of field. Cinematic composition reminiscent of Makoto Shinkai's vivid color palettes. Hyper-detailed rendering, Unreal Engine 5 realism, 8K resolution with subsurface scattering on skin." So combined it's "danganronpa_3_(anime), danganronpa_(series), super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono, anime_screenshot , anime screencap, anime coloring, Shallow depth of field. Cinematic composition reminiscent of Makoto Shinkai's vivid color palettes. Hyper-detailed rendering, Unreal Engine 5 realism, 8K resolution with subsurface scattering on skin. "</p><p></p><p>I've also found that the lyhAnimeFlux_v4Niji Flux model seems to compliment this style really nicely. It takes it further away from the specifics of the Danganronpa anime but I think it's a nice look in and of itself. Especially with a mix of the Danganronpa anime style, the lyh base, and the sprite style from the games.</p><p></p><h2 id="chroma:">Chroma:</h2><p>The chroma lora is very flexable and you can use the activation keywords followed by tags, a narrative prompt, or a mix of both. Personally, my preference is a somewhat even mix of narrative prompt and tags sprinkled in.</p><p></p><p><strong>Recomended tags for chroma:</strong> danganronpa_3_(anime), danganronpa_3_(anime)_(style), danganronpa_(series),super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono_(style), super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono, anime_screenshot , anime screencap, anime coloring, detailed_background, this is a screenshot from danganronpa the animation, this is a screenshot from danganronpa 3</p><p></p><h2 id="illustrious:">Illustrious:</h2><p>This version is temperamental in a different way than the flux lora. If you're working with loras of danganronpa characters I'd advise lowering the strength of them to about 0.3 while keeping this lora at 1.0 strength. I figured I'd be on the verge of overtraining on this but it seems to be the opposite so next time around that'll hopefully be a bit less erratic.</p><p></p><p>Your prompt should probably contain these tags if you're using Illustrious: danganronpa_3_(anime), danganronpa_3_(anime)_(style), danganronpa_(series),super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono_(style), super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono, anime_screenshot , anime screencap, anime coloring</p><p></p><p>You can try to move it closer to the despair and future arcs in DR3 with: danganronpa_3_(anime), danganronpa_3_(anime)_(style), danganronpa_(series),anime_screenshot , anime screencap, anime coloring</p><p></p><p>Or closer to DR 2.5 with: super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono_(style), super_danganronpa_2.5:_komaeda_nagito_to_sekai_no_hakaimono, danganronpa_(series),anime_screenshot, anime screencap, anime coloring</p><p></p><h1 id="chroma:">Chroma:</h1><h3 id="danganronpa-characters-and-chroma">Danganronpa Characters and Chroma</h3><p>AKA, why some chroma gallery images 'almost' look right but are off in some significant ways, like Chiaki having a black hairclip. TLDR; You'll want to pair this with a danganronpa character lora to produce images of danganronpa characters in chroma.</p><p>I was really hoping that this lora would make it easy to generate a large chunk of Danganronpa characters in Chroma. Sadly, with a few exceptions, you're going to still need to pair a character lora with this style lora if you want to use it for specific characters. There's a few exceptions. Chiaki is a little over represented in the training data and you can usually generate her with this lora. I used some danganonrpa v3 sprites for Tsumugi just to see what would happen and I think hers in particular wound up meshing pretty well with this lora. Miaya, the ultimate therapist from the future arc, for whatever reason wound up pretty promptable too. But in general with the chroma lora your options are, by order of results: This style lora and a chroma chracter lora &gt; style lora and flux character lora for a character in the anime &gt; style lora and flux character lora for a character not in the anime&gt; just the style lora. Of course that's all in regards to trying to make an image featuring a specific danganronpa character. If you're just going for the style then this lora all on its own is great.</p><p></p><h3 id="danganronpa-flux-loras-and-the-chroma-style-lora">Danganronpa Flux loras and the Chroma style lora</h3><p>I've tried testing this with a lot of my flux danganronpa character loras and have found a few trends. In particular you should base the strength of the flux character lora on whether they were in the danganronpa anime or not. If a character was in the anime you're generally trying for more of a light nudge in the right direction with the flux lora rather than trying to really push for the lora's full data. For exmple, Chisa is in the anime to an extreme degree. As such you can generally get away with having her flux character lora at about 0.1 to 0.3 strength. While Miu Iruma wasn't in this lora's training data at all so her flux character lora needs to be much higher - generally around 0.8. The problem comes from Flux loras not being fully compatible with chroma. Push a flux lora too hard and you're getting a lot of bad noise along with the good. In practice this means keeping my flux loras at below 0.8 and this style lora at 1.0 if you mix them together for a prompt.</p><p></p><p><strong>Flux:</strong> version 1.0 doesn't require keywords, Illustrious version does. And the Flux 2.01 'should' work with no keywords or the same ones as Illustrious. Likewise Flux 2.01 should work with both narrative prompts and to a lesser extent with tags. In practice...I don't really know. Honestly the Illustrious lora came out a lot better than the Flux one. But I've had 'some' good results from the Flux 2.01 lora so figured I should share it.</p><p></p><p>While not the intent, there's enough of some of the more obscure characters in the training data, like MIaya, that it's strengthened Illustrious' recognition of them to the point where they might be more or less usable without an additional lora. But in general the best bet with the Illustrious style lora is to just combine it with a danganronpa character lora which is used at a lower strength. See usage for more details on that. And you might be able to get some Danganronpa characters in the Flux 2.01 with tags and no additional loras as well. In general the more a character is in the Danganronpa 2.5 anime the better your chances of getting them with tags in the Flux lora. Tsumugi's in there too and as a result you 'might' be able to push for something closer to game style while keeping anime elements by using a tag of komatsuzaki_rui_(style) in your prompt.</p><p></p><h2></h2><p></p>

Danganronpa Anime Style Mix

Danganronpa_Anime_Style.safetensors

FLux 1.0

Anima (preview) v2.2

Z-Image Base 2.2

Z-Image Turbo v2.1

Lumina 2 v2.01

Chroma v2.01

Flux 2.01

Illustrious 2.0

Files with SHA256: b8ffbc0cb83ad02f0803eb373ce405f337559288a736642d254dc8fd6643a5f9