For all of you who, like me, sometimes need a bit of inspiration or just want an easy way to generate prompts for Anima, I’m sharing this LLM. It’s a fine-tune of the Huihui-gemma-4-12B-it-qat-q4_0-unquantized-abliterated model. While its core capabilities are largely the same, it has a better understanding of prompt structure and specific Danbooru tags than the original.
The model comes with 3 system prompts:
NSFW Gay
NSFW Straight
SFW
Of course feel free to edit them and share in the comments if you have some tips to improve them!
Tip: It’s a vision model, so feel free to have some fun by giving it images to see what kind of prompts it generates!
Just so you know: I didn't originally plan on sharing this model because I fine-tuned it for myself using exclusively masculine and mostly NSFW prompts. Because of this, it might tend to lean towards male-centric and NSFW content. However, I want to reassure all our straight friends (and those looking for SFW results) that after testing, the model is still very versatile and perfectly capable of generating female-focused and SFW prompts as well. (But still be sure to double-check your prompt before hitting generate! :p)
Description
Comments (4)
Excuse me being a semi-beginner in text-gen.
So i suppose i can load teh model in Koboldcpp.
but how much VRAM does it use and how does it work?
Like if i sent it a text, it can imagine the prompt?
No need to apologize; that’s what the comment section is for!
I don't use Koboldcpp (I use LM Studio), but yes, the model works just like a standard LLM. There might be some Koboldcpp-specific nuances regarding vision, though (assuming you can actually use images in the software).
In terms of VRAM/RAM, the model uses 8.57GB for an 8,000-token context window.
That's right, you just need to use the correct system prompt or tweak it to match what you’re looking for. Basically, you could almost replace the model itself with any other LLM because most of the work is done by the system prompt. The advantage of this specific LLM over others is that it’s uncensored and fine-tuned to handle Danbooru tags better. For example, the base model tends to write '2person, 1boy, 1girl' instead of just '1boy, 1girl,' or '1boy, 2boys' instead of '3boys,' whereas the fine-tuned version gets it right
@Aerem i gave it a spin since i can load it with also anima and still have enough VRAM. It is really interesting how it explain his reasonning for better prompting. Defitively gonna keep loading and see. I appreciate the vocabulary effort to get male variant too. I neef better dick description to keep the gang bang pictures interesting.
Also after rereading you, i notice it is an image to text model too, and not jusr text to text. I had no idea.
Koboldcpp is said to be able to use image as input but require a multi model vision model file mmproj. No idea what that is, i need to check later.
@PotatoesAI Thanks for the feedback!
The vision models have two parts: the main text part (the largest file, named after the model) and the smaller 'mmproj-model-bf16' vision part. You can find both in the downloads.