工作流下载地址:
https://civitai.com/models/1316907?modelVersionId=1486613 (含bbox检测模型)
工作流修改说明
Illustrious文生图工作流,调用ControlNet openpose及放大模型使用后,通过bbox检测脸部、眼睛、胸部、手部进行局部重绘。
PS:采样器全局输入调用会出现bug(反正我传不了参数,一直报错),并删除了lib线上无法加载的预设节点、图像选择器。
① 加载Checkpoint和LORA
原节点不修改
大模型用的是:MLiang的英雄联盟风格Illustrious模型
https://www.liblib.art/modelinfo/bfa0a416245546e28df67b1e9da01782?versionUuid=962e018f238b41e28c3c96ecb09dbbf4
lora没有使用
② 提示词
原节点:
修改为:
Positive Quality (Score)质量词
masterpiece,best quality,amazing quality, best quality, newest, absurdres, highres, detailed background, detailed hair, depth of field, detailed skin, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, absurdres, embedding:OverallDetailXL,
可固定不动
-
Character/NPC
1boy,pink_hair,blue eyes, detailed face,
根据需要设定
-
Clothing
white_shirt,black suit,eutral expression,(standing:1.1),
根据需要设定
-
Background/Location
(white background:1.25),
根据需要设定
-
提示词用的节点为comfyui-image-saver的String Literal
原预设节点为pythongosssss/ComfyUI-Custom-Scripts的Preset Text 预设文本(用不来。不纠结它,删除)
-
负向提示词
worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, sagging breasts, bad perspective, fisheye, clone, clones, chibi, child, loli,young, muscular female, multiple views,kid, deformed, big breasts, (huge breasts:1.25), large breasts, blush, blank eyes, artist name, english text, fat, plump, (bodysuit:1.15), (text on shirt:1.15), superhero costume, mask, (logo on clothing:1.4), (logo on shirt:1.4),(large breasts:1.25), nude,
---------------------
(low quality, worst quality:1.2),year2010,year2009,year2008,year2011,legacy,prestige,colored sclera,NSFW,logo,text,blurry,low quality,bad anatomy,sketches,lowres,bad proportions,cropped,watermark,signature,worstquality,grayscale,monochrome,username,out of focus,normal quality,mutated,poorly drawn face,human,
根据需要修改
③ 调用openpose
适用****Illustrious的ControlNet-OPENPOSE下载地址:(NoobAI-XL ControlNet-OPENPOSE)
https://civitai.com/models/962537?modelVersionId=1077649
PS:用其他openpose感觉未能调用到
-
NoobAI-XL ControlNet-OPENPOSE测试效果:
动作准确率95%,全身像需要开启放大修复
④ 采样设置
生成尺寸:使用sdxl的预设尺寸,或者传入openpose图的尺寸,二选一
小结:在放大前的节点布局如下
⑤ 放大及局部重绘
bbox面部重绘
bbox眼睛重绘
bbox胸部重绘
bbox手部重绘
到手部重绘后直接出图,如果需要重绘其他部分,按需添加。
-
重绘对比:
PS:测试完一个比较好使的插画模型工作流,接着搞下一个。