Danganronpa Anime Style Mix

What is it in practice?:

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.

Which Lora/Model combination should you use?

Krea 2: The style part of this style lora is great, and it 'almost' handles a fair amount of Danganronpa characters without needing a dedicated character lora. Generally with one significat issue for each. Chiaki's hairclip is wrong and her Hope's Peak Uniform often has Danganronpa 2 elements, Junko's hairclips are generic, Seikio's mask isn't quite right, etc etc. So in general the Krea 2 lora benefits from an additional character lora when possible.

Illustrious: The best jack of all trades and easiest to use

Anima: Solid understanding of many Danganronpa characters even without character lora but the lora strength tends to run a bit overly strong and might need some tinkering to find the best results.

Z-Image Base: I think the Z-Image base lora might be the highest overall quality. Biggest downside is that Z-Image has no built in danganronpa knowledge and character loras aren't as plentiful.

Z-Image Turbo: I feel like it didn't come out as well as the z-image base lora. It's fine, but the z-image base results were just considerably better.

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.

Lumina: Sadly just didn't come out that great.

Flux 1: Similar to Chroma but full compatibility with the existing Flux 1 lora ecosystem including my Flux 1 Danganronpa character loras.

Important for Z-Image Turbo

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.

Important - All

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.

Usage:

Krea 2:

While the lora can use tags, it generally works best with narrative style prompts.

Z-Image:

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.

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.

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."

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.

Lumina :

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.

Lumina has a distinctive prompting style that you should aim to use. With this lora, try something along these lines:

You are an assistant designed to generate anime images based on textual prompts. 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

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.

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.

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."

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.

Flux:

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.

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...