AI Art Generation Handbook/ControlNet/Scribble
I believe some of you are old enough , may remember this how to draw an owl meme
Well , do you know that it is super easy to recreate this inside Stable Diffusion.
Assuming you got latest Xinsir Union ControlNet models, you can do it inside both Auto1111 and SD.Next.
SD.Next
[edit | edit source]We try to recreate above meme using the Scribble ControlNet.
This model is designed to work with simple sketches or scribbles. It's perfect for turning basic shapes (like two circles ala snowman) into more detailed drawings of an owl as shown in the picture below. The result should be a detailed pencil sketch of an owl, where the two circles you drew have been interpreted as the owl's body and head, with the rest of the details (feathers, eyes, branch, etc.) filled in by the AI Model.
The ControlNet is already integrated within SD.Next, therefore just go straight to the "Control" tab
Using this prompt, we can try to recreate the images
Highly detailed pencil sketch of an owl, highly detailed feathers, sitting on a branch
Control Input
[edit | edit source]We upload this scribbles into the Input Image ( Look for : Drop Image Here ) on the left side
Control Sampler
[edit | edit source]Double steps from 20 to 40 for more flushed out details
Control Elements Parameters
[edit | edit source]Pre-Processor: Line Art Realistic
ControlNet : Xinsir Scribble XL
Recommended Parameters Values
[edit | edit source]Strength:1.6
( Determines how strongly ControlNet influences the generation process. )
Values above 1.0 can exaggerate ControlNet's influence, potentially leading to over-emphasis of the control image.
Start: 0.35
( Determines at which point in the denoising process ControlNet starts to influence the generation. )
A value of 0.0 means ControlNet starts influencing from the very beginning.
End: 0.8
( Determines at which point in the denoising process ControlNet stops influencing the generation. )
End Results
[edit | edit source]
If done correctly, the AI model will try to adjust the diffusion to match as close as possible to the input images, provided that input images is closely resembled the Input image silhouette , also the input images of circles and rectangle will be not much visible