૮₍ ˶ᵔ ᵕ ᵔ˶ ₎ა A release of my "KuromiwaMix" illustrious model merges! ♡♡♡ Made for/with Illustrious SDXL.
➜ Highly Recommended to use V1.5, the finished model.
➜ Best settings are with CFG 7, 832x1216, Clip Skip 2, Denoising 0.4, ♥
➜ The base model for this collection is NTRMIX. ♥
➜ NO inpainting or controlnet/editing was used in any of this collections samples. ♥
➜ This collection will NOT continue to be updated because this is a completed model as of v1.5. ♥
( ╹ -╹)! The Kuromiwa collection is partially free-use under the illustrious license, however, please do not claim to have made the model yourself, sell the models or merges of it, ( unless the merge has such drastic change you cannot tell its KuromiwaMix, feel free to message me on discord to ask, or to ask for permission for selling images. ), sell generated images, or use it for illegal material. ♡♡
➜ Danbooru tags and artist styles/characters from noobai are available, the recommended quality tags are: masterpiece, best quality, highres, year 2024
➜ Negative prompts can vary but the recommended are: bad hands, low quality, worst quality, bad quality, error, fault, wrong, distorted, disfigured, grainy, censored, boring, bland, dull, english text, bad anatomy, sketch,
~ヾ(＾∇＾) Sharing the model or posting gallery images is highly appreciated! See you next time!
𐙚 Contact 𐙚:
waikitty on discord
waikitty on civitai
Description
v1.1, more details, a little loss of color-contrasting abilities
FAQ
Comments (7)
I noticed you don't use quality tags in your prompts for this model. Does it not need them? Also, does it know the same artist styles and characters as Illustrious/NoobAi?
It has the same knowledge as NTRmix, all artist styles and characters too, however I did not merge any NoobAi into this model, I do use quality tags, but I have removed them from the prompts to encourage creativity and versatility in images posted or created, I will add a section to the description for quality tags as this is a frequently asked question!
If the main line tends to be thick, you may be able to solve this by putting “sketch” in the negative prompt.
This was translated from Japanese using DeepL
i'll add that!
I really love the style of this checkpoint, but the Danbooru dataset is messed up. Even if you turn up the CFG a lot, the characters only appear after many tries. Maybe is a way to fix it, i hope so.
Hi! Can you tell me more specifics? Also PSST: Im working on remaking this completely at the moment :)
I really like the style of your checkpoint, it looks awesome. The thing is, when I try to generate several characters with it, I need tons of attempts just to get one good image. Normally, I use the Tiled Attention Couple node, and with other models it works like magic — multiple characters and LoRAs show up without problems. But with your checkpoint, it feels like something’s off with how CLIP and the Danbooru dataset are integrated.
It’s not that CLIP itself is broken, but the way the dataset was aligned during training might make it harder for the model to separate characters in prompts. That’s why it struggles with multi-character generations, even though single characters look amazing in your style.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.