/upscale
animetoreal,soft light, masterpiece, best quality, realistic,photorealistic,1girl,,hoshino_ai_(oshi_no_ko), eyelashes, eyes_visible_through_hair, fangs, floating_hair, glint, sparkle, star_(symbol), star_hair_ornamenthair_between_eyes, hair_ornament, hair_ribbon, heart-shaped_ornament, long_hair, purple_eyes, purple_hair, rabbit_hair_ornament, sidelocks, star-shaped_pupils, symbol-shaped_pupils, tsurime, very_long_hair, gloves, idol_clothes, pink_gloves, pink_ribbon, pink_shirt, shirt, idol,,in the mountain,covering mouth,floating_clothes
NekoCat
A style transformation lora, the greater the weight of the lora, the closer the picture is to the realisitc style.
Recommended Base Model: Animagine XL V3.1
Recommended Weight: 0.7
Trigger Word: animetoreal
一个风格转换lora,权重越大越接近真实风格。
推荐底模:Animagine XL V3.1
推荐权重:0.7
触发词:animetoreal
训练初衷:
Animagine XL V3.1等二次元大模型相比传统真实模型有许多优势,其人物姿势、服饰更加可控,泛化能力强,无需角色Lora即可识别诸多动漫角色,是否能将其优点转移到到真人图片上呢?最直接的方法便是图生图或者条件生图,但结果往往很差,五官比例、身体比例不协调,于是打算训练一个Lora帮助二次元大模型生成现实风格的图片。
Lora效果:
优点:五官比例、身体比例、皮肤纹理、头发细节、衣服材质、光影细节等方面更接近现实,但受限于二次元底模,风格转换并不彻底,只能算是3D的水平,但在此基础上图生图的效果非常好;选择较好的图片搭配本人的风格转化工作流(https://www.seaart.me/workFlowAppDetail/col9tite878c7397l90g)即可彻底转化为真人。
缺点:转化为现实风格后更容易触发平台的审查机制,导致NSFW标识;对男性角色和儿童角色的支持不够友好;比较依赖底模Animagine XL V3.1,泛化性不足。