--- license: creativeml-openrail-m tags: - stable-diffusion - text-to-image - safetensors language: - en ---

HoloKuki

### Previews! To make the previews fair I applied no editing, no inpainting or such techniques. All the images have metadata. There might be more previews on the [civitai](https://civitai.com/models/17598/holokuki) page, since it's easier to upload them there.
### Holos! This is a big ass file, open it and look at the holos. Stare at them.
### Description TL;DR WHAT I TRIED TO ACHIEVE: A midpoint between being able to generate stunning scenery (dpep) and just cute anime girls (Anything) with some other models mixed in - to help improve other aspects. It's a model, made by mixing other models, the base were Anythingv4.5 and dpep-chillout mix. I've made many mixes before and I've applied the knowledge I've gotten from that here - a lot of BW merging was applied. I feel like this mix is good enough to contend with the most popular ones, and it's the one I use currently, for all my purposes. It can do SFW, NSFW, thanks to closer to death's dpep it has the capability for interesting backgrounds and I've merged in a VTuber model - all of Hololive can be prompted for. But I think the images speak for itself better than I can. ### What to grab I advise grabbing [this file](https://huggingface.co/Aotsuyu/Kukicha/blob/main/HoloKukiv2-fp16.safetensors) - this is version 2. ### Defaults that I use: ``` detailed background, masterpiece, best quality, prompty promptu Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000 Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1189140555, Size: 512x768, Clip skip: 2, ENSD: 31337, Discard penultimate sigma: True ``` This is with kl-f8-anime2 vae, probably use higher cfg if trying with the anything or whatever vae, hell, mess around. I also sometimes use 22ish cfg with the dynamic thresholding extension: ``` Dynamic thresholding enable#: True, Mimic scale: 7, Threshold percentile: 96, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4 ``` ### DARKNESS The darkKukiv1.safetensors LoRA included in the repo is my first version of a noise offset LoRA specifically for this model. I'm working on improved version as well, but this one isn't bad by any means. What it does is it makes the model able to output a wider range of lighting, example:
That's with the lora at 0.6 strength and the positive prompt: ```detailed background, masterpiece, best quality, 1girl, solo, hatsune miku, pov, bed, window, night, darkness, dark, blue hair, twintails, fake smile``` #### Horror This is an another source of noise offset so if that's an issue, disable the darkness LoRA. And just read the readme of the LoRA. https://huggingface.co/Aotsuyu/HorrorLora/ ### I want to know more about such and such Your best bet is this https://rentry.org/TohoAIFAQ I'm trying to slowly fill it out. ## Da recipe Recipe for V2, V1 was similar but also included a gape merge at the end. BW merge anything-v4.5-pruned.safetensors and dpep3-chillout.safetensors = temp1 at `1,0.9,0.7,0.5,0.3,0.1,0.4,0.4,1,1,1,1,0,0,0,0,0,0,0,0.1,0.3,0.5,0.7,0.9,1` temp1 and Cocoa = temp2 at `0.15,0.15,0.15,0.15,0.15,0.31,0.3,0.3,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15` temp2 and Tea = temp3 at `0.35,0,0,0,0,0,0.1,0.1,0.1,0,0,0,0,0,0,0,0,0,0,0,0,0.35,0.35,0.35,0.35` temp3 and 7th_anime_v3_A = Kukiv2 at `0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.4,0.4,0.4,0.4` Add diff merge of Kukiv2 + (merge of hll1 and hll3.1 at like 0.4 ratio (more of hll1) I think) - animefull-final-pruned = HoloKukiv2 ## Big Thanks to You should also check out their models! - [Closertodeath](https://huggingface.co/closertodeath) for being cute. - [andite](https://huggingface.co/andite) for being handsome. - [syaimu](https://huggingface.co/syaimu) for making a nice model that looks nice. Their model is a part of the mix. They don't know. Don't tell them. - [NoCrypt](https://huggingface.co/NoCrypt) for providing the necessary computing power. - Other members of the 東方Project AI community. - Anons responsbile for hololive/vtuber finetunes.