前言
Flux CN:XLabs发布可控组件V3,可控性里程碑
Flux ControlNet模型简介
在之前的文章中已经介绍过由XLabs-AI团队开发的最新Flux模型的最新的3款ControlNet模型(Canny硬边缘、深度Midas、线条HED),但是整体初步体验下来,整体质量还未达到可完美使用效果。但XLabs-AI团队在继续的推进和训练FLux ControlNet模型的进展,随着不久时间就由发布了这三款的ControlNet模型的V2、V3版本。今天的文章将一起体验和评测它的最新V3版本模型。笔者将长期跟踪和关注Flux模型生态的LORA、ControlNet,这也是Flux模型构建繁荣生态,以及取代SD1.5,SDXL的最后一公里的里程碑事件。
另外,本次使用ControlNet,从V2版本开始,官方训练和推荐分辨率变为1024*1024。这与之前的是有区别的(v1版本中,深度:1024x1024分辨率 ;线稿Canny,HED使用 768x768 分辨率)。
XLabs-AI Flux ControlNet演示
XLabs-AI在其主页提供了这3类flux ControlNet的演示示例如下所示:
Canny
更多信息参见之前文章:[[ComfyUI]Flux新纪元:Flux可控性新起点,第一个CN Canny硬边缘组件发布]
深度 (Midas)
HED
Flux ControlNet模型体验
当前最新版本的ComfyUI已支持Flux ControlNet模型的体验,同时还需要通过插件管理器Git安装x-flux-comfyui该插件。其中需要将模型放置到 ComfyUI/models/xlabs/controlnets目录下,如本地不存在该目录可以手工新建可以,也可以首次运行自动创建目录。更多安装详情参见文章:[[ComfyUI]Flux:新添深度和线条3款CN成员,工业级可控性即将到来,继LORA后的生态最后一公里]
本文使用的是XLabs-AI/x-flux-comfyui插件,具体安装方法请参考之前文章:[[ComfyUI]Flux:新添深度和线条3款CN成员,工业级可控性即将到来,继LORA后的生态最后一公里]
另外这个版本出图耗显存较大,24G显存机器大约每张图25步迭代需要耗时1分钟。
所有的AI设计工具,模型和插件,都已经整理好了,👇获取~
工作流界面
关于本地ComfyUI工作流体验参见之前文章:[FLUX[续篇]:12B参数23G最大开源文生图模型,Dev版直出惊艳美图欣赏
本文涉及ComfyUI工作流和模型均可在LIBLIBAI上下载或在线运行体验:
LIBLIBAI平台已支持Flux模型在线运行体验,并且笔者已增加换脸节点使用。
一:硬边缘(Canny)
以下均使用CN权重幅度为0.7体验测试。
01. 室内设计
`Chinese style, cyberpank dining room, full hd, cinematic,sunshinex`输出效果
02. 偷鸡娃
Capture an energetic scene of a chubby little boy in rural China, shirtless, chubby belly, still clinging to dirt, wearing a pair of muddy white cloth shoes on his feet, and a mischievous and happy expression on his face, running through the muddy streets of the countryside. The dust was flying, and the little boy was clutching on his arm a large, strong rooster, whose feathers were a mixture of deep Burgundy and glittering gold, in stark contrast to the boy's plain clothes. The rooster frowns slightly, but is surprisingly calm, adding a whimsical twist to the energetic chase. The background is a rural wet market. The overall tone of the film is a lively and lively atmosphere, which summarizes the free and simple youth. The sun shines through the white clouds, casting dappled shadows, and the pleasant warm picture enhances the sense of movement and the joyous chaos of the moment.
输出效果
二:深度(Midas)
01. 室内设计
Chinese style, cyberpank dining room, full hd, cinematic,sunshine
输出效果
02. 偷鸡娃
Capture an energetic scene of a chubby little boy in rural China, shirtless, chubby belly, still clinging to dirt, wearing a pair of muddy white cloth shoes on his feet, and a mischievous and happy expression on his face, running through the muddy streets of the countryside. The dust was flying, and the little boy was clutching on his arm a large, strong rooster, whose feathers were a mixture of deep Burgundy and glittering gold, in stark contrast to the boy's plain clothes. The rooster frowns slightly, but is surprisingly calm, adding a whimsical twist to the energetic chase. The background is a rural wet market. The overall tone of the film is a lively and lively atmosphere, which summarizes the free and simple youth. The sun shines through the white clouds, casting dappled shadows, and the pleasant warm picture enhances the sense of movement and the joyous chaos of the moment.
输出效果
三:线条(HED)
01. 室内设计
Chinese style, cyberpank dining room, full hd, cinematic,sunshine
输出效果
02. 偷鸡娃
Capture an energetic scene of a chubby little boy in rural China, shirtless, chubby belly, still clinging to dirt, wearing a pair of muddy white cloth shoes on his feet, and a mischievous and happy expression on his face, running through the muddy streets of the countryside. The dust was flying, and the little boy was clutching on his arm a large, strong rooster, whose feathers were a mixture of deep Burgundy and glittering gold, in stark contrast to the boy's plain clothes. The rooster frowns slightly, but is surprisingly calm, adding a whimsical twist to the energetic chase. The background is a rural wet market. The overall tone of the film is a lively and lively atmosphere, which summarizes the free and simple youth. The sun shines through the white clouds, casting dappled shadows, and the pleasant warm picture enhances the sense of movement and the joyous chaos of the moment.
输出效果
四:线稿上色(Canny)
01. 背影杀手
1girl, chinese girl, 20-old-years, simple background,Turn your back on the audience, half backless,shadow killer, light red|yellow dress, transparent silk dress, white backgroound
输出效果
最后,整个V3版本ControlNet模型体验结果,相比之前的V1、V2版本已经取到了巨大的进度,更加完善,已逐步可用于出图实践中。对于大部分线稿图质量整体已能直出不错质量,在测试中仅对中国龙神态类型质量不太好,笔者猜测可能是因为训练数据问题。具体待读者测试和评价。需要说明,官方插件结合ControlNet模型运行特别耗费显存资源,这也是Flux Dev一直以来的问题。
为了帮助大家更好地掌握 ComfyUI,我在去年花了几个月的时间,撰写并录制了一套ComfyUI的基础教程,共六篇。这套教程详细介绍了选择ComfyUI的理由、其优缺点、下载安装方法、模型与插件的安装、工作流节点和底层逻辑详解、遮罩修改重绘/Inpenting模块以及SDXL工作流手把手搭建。
由于篇幅原因,本文精选几个章节,详细版点击下方卡片免费领取
一、ComfyUI配置指南
- 报错指南
- 环境配置
- 脚本更新
- 后记
- …
二、ComfyUI基础入门
- 软件安装篇
- 插件安装篇
- …
三、 ComfyUI工作流节点/底层逻辑详解
- ComfyUI 基础概念理解
- Stable diffusion 工作原理
- 工作流底层逻辑
- 必备插件补全
- …
四、ComfyUI节点技巧进阶/多模型串联
- 节点进阶详解
- 提词技巧精通
- 多模型节点串联
- …
五、ComfyUI遮罩修改重绘/Inpenting模块详解
- 图像分辨率
- 姿势
- …
六、ComfyUI超实用SDXL工作流手把手搭建
- Refined模型
- SDXL风格化提示词
- SDXL工作流搭建
- …
由于篇幅原因,本文精选几个章节,详细版点击下方卡片免费领取