[stable-diffusion-art] 指北-5 controlnet

https://stable-diffusion-art.com/controlnet/https://stable-diffusion-art.com/controlnet/https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/mainhttps://huggingface.co/lllyasviel/ControlNet-v1-1/tree/mainControlNet V1.1 - a Hugging Face Space by hystsDiscover amazing ML apps made by the communityicon-default.png?t=N3I4https://huggingface.co/spaces/hysts/ControlNet-v1-11.安装

!wget https://huggingface.co/lllyasviel/ControlNet-v1-1/resolve/main/control_v11p_sd15_openpose.pth -O ./extensions/sd-webui-controlnet/models/control_v11p_sd15_openpose.pth

resolve is for downloading, not blob, just pull the code directly into the directory in extensions

2. use

parameter:

enable:是否启用controlnet
invert input color:交换黑色和白色,在上传涂鸦scribble时,controlnet需要黑色背景和白色涂鸦,
如果使用外部软件创建涂鸦,就需要对颜色进行更改
RGB to BGR:改变上传图像的颜色通道顺序,上传normal图等,如果使用processor,就不需要了,
说白了就是不使用annotator的话,可以直接原图上传,但是要改一下通道顺序
low vram:低于8g显存可以用这个
guess mode:配合更高的steps,例如50,图像生成不受文本提示的引导
preprocessor:边缘图,深度图等,None表示使用输入图作为控制图
model:和sd大模型一起使用

weight:提示给予控制图的影响程度,它类似于prompt中的关键字权重,但是作用于控制图
guidance strength:用于于controlnet的步骤数,类似于图像到图像中的去噪强度,如果strength为1,
则应用于controlnet所有的采样步骤,如果strength为0.7,则steps为50的话,controlnet将应用于
70%的采样步骤,即前35个步骤

resize mode:当输入图或控制与要生成的图像大小不同时,调整大小模型
envelope(outer Fit):裁剪控制图,使其和画布保持一致
scale to Fit:使整个控制图适合画布
just resize: 独立缩放控制图的宽度和高度,使其适合画布尺寸,将改变控制图纵横比

The controlnet weight is relative to the degree of compliance with the control chart of the prompt. The lower the weight, the less the controlnet has for the image to follow the control chart.

3.preprocessors

canny: edge detector, good at extracting curves and straight lines, but susceptible to noise

depth: The depth map represents the distance of the target in the image, base means near, black means far

hed: suitable for human detection edge generation contour

mlsd: Fast line detector, for interior design, architectural drawings, street views, etc., do not pass in portraits or curves

normal: generate a map with a 3d texture on the surface

openpose: key points

pidinet: detects curved and straight edges, similar to hed, but generally produces sharper lines with less detail

scribble: graffiti

seg: classify different colors for different objects

mlsd and depth are good in interior design

Guess you like

Origin blog.csdn.net/u012193416/article/details/130474211