A style model based on Animagine XL V3.1, trained using AI-generated materials. It is a specialized, narrow-scoped model with a distinctive artistic style, inheriting numerous character knowledge from Animagine XL V3.1.
base on: https://huggingface.co/cagliostrolab/animagine-xl-3.1
在制作模型的过程中，感谢MIAOKA对我的及时支援！为我解决了燃眉之急。
感谢breadman为了我提供的大量图片，没有他就没有这个模型。
封面动作和角色借鉴自Animagine XL V3.1 的示例图，下面是cagliostrolab给出的推荐参数：
For higher quality outcomes, prepend prompts with:
masterpiece, best quality, very aesthetic, absurdresnegative prompts like:
nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]
it’s recommended to use a lower classifier-free guidance (CFG Scale) of around 5-7, sampling steps below 30, and to use Euler Ancestral (Euler a) as a sampler.
Description
FAQ
Comments (6)
wow an amazing model, damn you are fast :)
btw do you happen to know a good ratio for merging anime checkpoint such as this with sdxl lightning lora? this model works amazing standalone, but with 8 step lora at weight 1, degrades the quality a lot, even with cfg 1 and sgm uniform.
thanks
Interesting question. How to merge Lightning LoRA with a checkpoint is something I haven't personally tried yet. I trained using a dataset with a single art style for Animagine XL V3.1, so there's a possibility that it may reduce compatibility with certain LoRA checkpoints. It's also possible that Animagine XL V3.1 might not be fully compatible with certain specific LoRA models.
@Chenkin Thanks for replying, in my experience, lightning lora works really well in realistic/3d models, but on anime based models, the quality degrades a lot. a good lightning based anime lora/model could make high range upscaling process a lot faster.
Also your models are amazing, especially the loras, thanks for sharing them.
@Chenkin 听说novelai v3就是在XL的基础上融合了巨量的LORA和画师图集做出来的