Chtholly_nXL

Prompt

illustration,solo,best quality, amazing quality, very aesthetic, absurdres, chtholly, 1girl, solo, (blue hair), long hair, blue eyes, multicolored hair, two side up, streaked hair, (pink hair:0.8), very long hair, smile,cry,blush,high collar,open mouth,portrait, <lora:chtholly-000016:1>

Negative Prompt

dark skinned,dark skin,watermark,doll,word,words,bad feet,lowres,bad,error,missing,worst quality,jpeg artifacts,bad quality,watermark,unfinished,signature,extra digits,scan,bad anatomy,bad hands,text,missing fingers,fewer digits,cropped,low-resolution,normal quality,signature,animal,

Prompt

illustration,solo,best quality, amazing quality, very aesthetic, absurdres,film lighting,from_above, 1girl,cute,smile,blush,v,bracelet,collarbone,chtholly,solo, (blue hair), long hair, blue eyes, multicolored hair, two side up, streaked hair, (pink hair:0.9), very long hair, <lora:chtholly-000016:1>

Negative Prompt

dark skinned,dark skin,watermark,doll,word,words,bad feet,lowres,bad,error,missing,worst quality,jpeg artifacts,bad quality,watermark,unfinished,signature,extra digits,scan,bad anatomy,bad hands,text,missing fingers,fewer digits,cropped,low-resolution,normal quality,signature,blurry

Model description

模型基于noobai-XL-1.0微调，对于noobai-XL系列模型适配很好。数据集为约四百张高分辨率番剧画面截图，训练使用余弦退火+一次重启，在16轮(1568步)得到了较好的收敛。模型需要chtholly作为启动词，对于最初的纯蓝色头发和被侵蚀后带有红色挑染的头发均有训练，可以通过提示词blue hair, multicolored hair, two side up, streaked hair, red hair稳定触发，通过调整red hair的权重来调整发色比例，如(red hair:0.85)。 另附可选的剧中服装风格提示词：

dress, long sleeves, hat, witch hat, uniform, military, military uniform, witch, cloak, hat flower, skirt, jewelry, weapon, armor, brooch, corset, black headwear, jacket, wings, belt, shirt, apron, hat ribbon, shoulder armor, cape, holding sword, socks, black shirt, necklace, ribbon, white skirt, black jacket, flying, black cloak, hood, pauldrons

==============================================

This model is fine-tuned based on noobai-XL-1.0, with excellent compatibility with the noobai-XL series. However, its performance on other SDXL models is relatively poor. The dataset consists of approximately 400 high-resolution screenshots from anime scenes. Training was conducted using cosine annealing with one restart, achieving good convergence after 16 epochs (1568 steps).

The model requires the keyword chtholly to trigger. It is trained for both the initial blue hair and the red-streaked hair. These can be reliably triggered using the prompt keywords: blue hair, multicolored hair, two side up, streaked hair, red hair. The proportion of hair colors can be adjusted by fine-tuning the weights between blue hair and red hair.

Optional style prompts for in-show costumes are also available (listed above).

Trigger words

You should use chtholly to trigger the image generation.

Download model

Weights for this model are available in Safetensors format.

Download them in the Files & versions tab.