🖼️ControlNet

Guiding image generation with another image

With ControlNet, you can guide the image generation process with another image. This is a great way to produce images with a consistent visual layout.

The standard ControlNet used at glif is controlnet-canny. This is a ControlNet variant that first tries to find edges in the reference image, and then uses those edges as guidance.

controlnet canny, via https://github.com/lllyasviel/ControlNet

Note that not all Glif image models and APIs support ControlNet - please see How to use ControlNet on how to activate it.

Examples

Baseball Card Me by @ansipedantic

Based on this reference image:

Oppenheimer Yourself by @fabian

Based on this reference image:

How to use ControlNet

Turn it on inside the advanced section of the image block:

Then put a URL to your reference image:

Make sure your URL resolves to (preferably) a .jpg or .png image. Please note that Imgur links might not resolve.

Conditioning strength

Below are a couple of images where the strength of the Conditioning parameter is increased for the prompt an illustration of a cyborg.

The first image (strength=0.0) is always the control image. If almost no control scale is applied (0.1), we get a base cyborg illustration. Moving to the right, the control image guides the image generation more and more.

SDXL k_euler_a sampler.

SDXL unipc sampler.

SD21 k_euler_a sampler.

SD21 unipc sampler.

Canny upper & Canny lower

canny upper and canny lower refer to parameters of the Canny edge-detection algorithm used inside the model.

• Canny Upper: Pixels with intensity gradients above this value are considered strong edges.

• Canny Lower: Pixels below this value are discarded. Those in between are considered weak edges and are kept only if connected to strong edges.

Here's an image to illustrate what happens:

Variation grid of the canny upper and lower thresholds.

• by moving to the right, we are increasing canny_upper, so less pixels are considered strong pixels (pixels that have a strong gradient). In the extreme case, only the edge of the ball is considered a strong gradient. (It is an abrupt change in pixel values.)

• moving down, we are increasing canny_lower and here more and more pixels with weak gradients fall off.

• in the upper left quadrant the most interesting stuff is happening: with certain mixes of lower and upper, we get a mix of strong pixels (strong gradients) and weak pixels (weak gradients) that get connected .

Tip: Move both canny upper and canny lower to a certain amount, e.g. 150, until only the most salient edges remain, then move canny lower down until you get sufficient detail.