Obsidian Anise
Obsidian Anise is a manga-driven SDXL merge built for artists who want strong linework, expressive character styling, and high-fidelity rendering across multiple aesthetics. The model carries a clear manga influence in its contrast, silhouettes, and shading approach, while still supporting a wide range of finishes — from clean anime to semi-real and full realistic rendering.
What sets this merge apart is how consistent and structured it stays across styles. Lineart comes out crisp and intentional. Anime illustrations land with bold forms and clear expressions. Semi-real and realistic prompts maintain their detail without losing the underlying stylization. The model naturally leans toward strong contrast, defined edges, and deeper blacks, giving every output a confident visual weight.
Backgrounds and environments are a major strength. Architectural detail, forests, interiors, industrial scenes — all generate with convincing depth, texture, and atmospheric lighting. Surfaces like metal, fabric, fur, and stone are rendered with clarity and texture breakup, which helps blend stylized characters into richer, more realistic worlds.
Lighting performance is also strong. Rimlight, bloom, dramatic shadows, spell effects, and warm/cool shifts all resolve cleanly, giving images a dynamic finish without washing out the linework.
This merge suits:
Manga-style illustrations
Clean anime art
Lineart-focused compositions
Semi-real and stylized realistic portraits
Fantasy, sci-fi, modern, and anthro
Anyone who wants stylization + texture + depth in one model
🔧 Recommended Settings:
A well-rounded setup for general SDXL prompting.
Steps: 24–32
Sampler: Euler a or DPM++ 2M Karras
CFG: 4–5.5
Resolution: 832×1216 or higher
VAE: Default / model-embedded
🏷️ Recommended Quality Tags
anime screencap, anime style, thin lineart, cel shading, cinematic lighting, dramatic shadows, cinematic lighting,
american comic style, bold outlines, sharp inking, vibrant solid colors, cinematic lighting, dramatic shadows, cinematic lighting,
clean lineart, sharp outlines, precise inking, controlled shadows, high-contrast
semi-realistic details, refined textures, natural lighting, atmospheric depth,
masterwork, masterpiece, best quality, (masterpiece:1.2), (best quality:1.2),(very aesthetic:1.2), (absurdres:1.2), newest,
anime_art_style, 2d, line_art_style, limited_pallette,
Description
FAQ
Comments (16)
Looks like a really cool checkpoint. In your fist sample images (leather suit and stone cross), I see this "in the style of ckncore". What lora is that?
I think I accidentally left in a trigger word when testing. I didn't use a Lora on that generation.
@Sexiam thank you. That was gonna bug me.
@DeRivitiveKraftwork My image still have metadata in them. Dragging them into Comfyui would show the workflow
@Sexiam yeah, no Comfy because I do everything from me phone. Anything I do that's not onsite gen is in A1111 because a phone can handle the interface.
This is a really interesting model! Love it!
Very nice and clean style, I love it , Thank you
A very SLICK, SMOOTH, POLISHED, LUBRICIOUS checkpoint.
been testing it for a bit, and i gotta say, really loving the output
Amazing checkpoint with a style like no others, good job. 👍
A couple questions, tho. You suggest using some specific quality tags to force generating specific styles using your checkpoint.
For exemple "anime_art_style, 2d, line_art_style, limited_pallette,".
I ve been testing it using other quality tags such as "masterpiece, best quality, absurdres, ultra detailed" and i also ve been testing it using no quality tag at all but "anime_art_style, 2d, line_art_style, limited_pallette," and it made zero difference regarding the generated image style.
Also your suggested tags are for the most part either deleted Danbooru tags or tags that aren't Danbooru ones. This is a merge, not a trained checkpoint so could you explain from where those tags come from? I m suspecting that they have no impact at all on generated pics and that any change that comes after adding them is just the result of the prompt token weight change.
Could you explain?
Sure, if you could point at which tags you're referring to. The tags were a suggestion based on a result achieved over multiple generations, and not offish standards.
@Sexiam Their are 2 categories among tags that i'm pointing out.
The first category is the one related to quality. Tags such as "masterpiece", "best quality", etc... were trained tags on the original Illustrious release. Nowadays, because of the fact that merges are build upon other merged checkpoints instead of trained ones, those tags have almost no impact at all, especially in prompts that are above 75 tokens.
The second category is the one related to tags that aren't part of Danbooru's tags database. So tags such as "line_art_style", "anime_art_style", "anime screencap", "thin lineart", "clean lineart", "sharp outlines", "cinematic lighting", "natural lighting", "dramatic shadows", "controlled shadows", "american comic style", "precise inking", "bold outlines", "sharp inking", "vibrant solid colors", "high-contrast" (the correct syntax would have been "high contrast"), "semi-realistic details", "refined textures", "atmospheric depth", "masterwork" aren't part of Danbooru's tags database. If, and this is a HUGE if, any of those tags produce a consistent result then this would either be because of one of the checkpoint that you used for this merge having been trained on that specific tag with that specific syntax or because of Illustrator natural lanaguage understanding... which is extremely limited and unpredictable/irreliable.
@antarek That's kind of the nature of merging. No one is spending thousands on a custom dataset and hundreds of hours of compute. Loras come with their own clip and expand on the vocabulary of a checkpoint when merged. On top of that most of the complaining about prompting doesn't take into the fact that Illustrious is a finetune of SDXL which takes natural lanauge.
Not to seem confrontational but if you're having success with a prompt feel free to use whatever works. If you're looking for the next Illustrious base then this merge isn't it and I never claimed it was. The example of diffrent prompt styles that are most common on the plateform. I don't expect people to prompt the exact same way.
Not really suer what you're expecting from me or complaining about. Gen however you want.
@Sexiam I apologize if you perceived this as negative criticism. That wasn't my intention. I simply wanted to clarify which suggested tags are functional and not just placebo tags. As mentioned earlier, your checkpoint is truly excellent. However, regardless of the tool used, i prefer to use it after integrating its actual parameters. You also allude to this when you point out that LoRa integrated to merges add tags they have been trained on to the merged checkpoint's lexicon. Therefore, in my opinion, it's important for checkpoint creators to maintain a list of the functional tags they respond to when they aren't Danbooru tags. This is just my opinion, tho.
@antarek My applogies then. A lot of what I do is just testing. I've tried to eleminate some of the information that's not acturate, but AI as a hobby is scattered all over the place. I'm sure I'll get better over time but I'm just presenting information for direct testing.
I'll keep what you said in mind and try to refine what I do the best I can.
@Sexiam It makes sense, thanks a lot for your work on this fantastic checkpoint. 👍
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.