ControlNet
TODO: write about controlnet here, explain all the different kinds of controlnets available in SH.
ControlNet is a method of extracting edges and pose from an image and then using that to generate an image in Stable Diffusion, ControlNet is hardcoded to use DDIM sampler, so you don't need as many steps as other samplers; about 15 steps is enough to get good images.
there are 9 types of ControlNet in AI Horde.
since each of them are different and for different purposes, I'll use a different image for each of them to illustrate their use.
init strength is also supported for ControlNet.
this one detects edges but may sometimes pick up unwanted noise.
below is an example:
| Original image | ControlNet | Generated image |
|---|---|---|
 |
 |
 |
as you can see it changed the image but kept the original composition.
this one retains somewhat less details than canny, mainly useful for recoloring images.
| Original image | ControlNet | Generated image |
|---|---|---|
 |
 |
 |
the depth model keeps less details, but it'll keep the composition and overall shape the same, it however doesn't keep much details of the background.
this is useful if you don't want to keep the small details of an object and just want the overall shape, example:
| Original image | ControlNet | Generated image |
|---|---|---|
 |
 |
 |
can copy pose of an image, but it'll only copy the pose not the details nor the background.
generates an image from a scribble, note however it doesn't recognize colours, so your image has to be black on white.