Prim Fiorire (プリム・フィオリーレ) from H anime "Kuroinu ～The proud saint is stained with white filth～ 黒獣 〜気高き聖女は白濁に染まる〜".
This model offers a mini-sized version of the original art style.
You could use the prompts below to create characters:
mai_primk,very_long_hair,wavy_hair,pink_hair,parted_bangs,blue_eyes,large_breasts,hair_ribbon,tiara,choker,necklace,white_dress,center_opening,detached_sleeves,puffy_sleeves,white_gloves,pink_pantiesRecommend preferred Strength: 0.7~1.
A classic character that needs no further explanation.
Since some characters have already been created, let's continue to make some more. Compared to the excellent models created by other artists in the community, this set of models is a mini version.
This character is also a young-looking adult. Please visit my Pixiv community to view other example illustrations.
PS: Some sample illustrations cannot be posted in the community; you can visit my Pixiv community posts to view them.
https://www.pixiv.net/artworks/143678037
About Anima
This is a derivative model of Cosmos-Predict2-2B-Text2Image, with an architecture more geared towards model engineering, facilitating AI componentization tasks such as LLM model integration.
While its marketing emphasizes natural language understanding capabilities, it's actually still trained on Danbooru-style tags.
For creating favorite anime characters, diffuse methods rely more heavily on the underlying data the model carries.
Like many other models (such as IL), Anima still needs to accumulate more training data and adapt to better VAEs to enhance its performance and expressiveness.
For personal enthusiasts like myself who can't make model engineering our main profession, I'll wait for the Anima model ecosystem to mature before investing further :)
Description
FAQ
Comments (10)
They're the best characters in the world that I would love to create! But I'm getting less!
(´･ω･`)
Love Prim! I would highly appreciate it if you could make this Lora for Anima.
Anima currently requires fine-tuning. LoRA-level adjustments contribute very little to the overall model output, and using it might require combining a lot of data. Hopefully, this simple fine-tuning guide will be of some help to you :)
https://civitai.red/models/2551489/prim-fiorire-kuroinu-ova?modelVersionId=2984683
@00x09901 I love your stuff but I disagree that loras contribute very little to the output. Prodigy and adafactor are generally worse in picking up styles on anima in my experience, (which can work to your advantage) - CAME seems to be more aggressive on styles combined with more steps - might land it better for you. Just my thoughts anyways love your stuff!
@Sophorium if he does more steps and epochs, the only thing will do overtraining the Text encoder,wont frooze the images but enough to make any propting useless, unless the training methodes chances, gives more option to configurate that without crashing the whole process, there we can see training futher epochs than 10-30 depending on what you want.
@NobodyofInterest_Official epochs are irrelevant those are just save points, 480 steps might be a little low, you dont train the tenc in anima as it's not an SDXL-based CLIPL/G model. I'm not suggesting he they fry the model or anything, just suggestions from the things i've already noticed training anima =)
@Sophorium Thank you so much for the experts' suggestions; they were invaluable! :)
Image generation using diffusion methods is essentially pixel-based pattern prediction, and more base data is more conducive to generating beautiful results.
Anima uses the DiT method, which has higher computational costs but is better suited for handling intricate global images and data types with a timeline concept.
However, training and fine-tuning such models also places higher demands on the quality of the training set and the intensity of training. As a non-profit enthusiast, these costs are factors I must consider. :) Because the capacity of LoRA networks is relatively small compared to the base model, achieving better results for artistic effects that don't exist in the base model, without Anima's current large-scale fine-tuning checkpoints, requires larger network dimensions, higher-quality training data, and stronger training intensity. This is why I don't plan to invest too much in this direction at the moment.
I look forward to organizations with more financial resources and energy further refining and fine-tuning Anima through larger-scale data engineering :)
@00x09901 About anima cant argue anything beacuse i cant even use it myself, but from my experience using multiple epochs loras the overtrain is always a issue. My philosophy is that you shouldn't fix something that isn't broken; if it works, it works.
@NobodyofInterest_Official Haha, it is indeed a simple and plain truth o(^▽^)o