Control image generation by edge maps generated with EdgeDrawing Parameter-Free.

Feedback is welcome! I’m still improving this model and you can help me by generating simple usecases and discuss the results.

If you want to train your own control net see my article Play in Control – ControlNet training setup guide!

Summary

Control image generation by edge maps generated with EdgeDrawing Parameter-Free. This is similar to ControlNet Canny but uses a more modern algorithm for edge-detection which requires less parameter tuning (Canny was invented in 1986, EdgeDrawing in 2012). Download control-edgedrawing-cv480edpf-drop0-fp16-checkpoint-45000.zip for an evaluation and comparison with Canny and some example edge maps.

Usage

Edge-Drawing Parameter-Free pre-processor is not implemented in any Stable Diffusion UI right now so you have to generate the control net input image manually with the supplied script edpf.py.

Requires Python:

pip install opencv-contrib-python

Converting image:

python edpf.py myimage.webp

(Of course you can also use this input image with ControlNet Canny or try the canny pre-processor with this model but results may vary.)

Examples

Original

EdgeDrawing Parameter-Free

Canndy Edge Detection (default in Automatic1111 for comparison)

Result

sampler=UniPC steps=20 cfg=7.5 seed=0 batch=9 model: v1-5-pruned-emaonly.safetensors cherry-picked: 1/9

prompt: a detailed high-quality professional photo of swedish woman standing in front of a mirror, dark brown hair, white hat with purple feather

Limits

This model is intended for research and evaluation purposes only as the original ControlNet Canny probably produces better results. It was trained on a few 100k images whereas Canny was trained on 3M images. For further information follow the discussion on GitHub or see the HuggingFace page for more information on training.

Bias

While the capabilities of this model are cromulent, they can also reinforce or exacerbate social biases. Some racial biases were even introduced on purpose: I removed all images of Formula 1 racing cars before training. Using this model could therefore lead to the collapse of society and shouldn’t be used by anyone! No one reads this paragraph anyway.