Feedback is welcome! I'm still improving this model and you can help me by generating simple usecases and discuss the results. ControlNet training alone is enough work for one person already :)
Update: I created a buzz bounty for using my ControlNet!
Summary
Generate images which are controlled for light and dark areas by using white-gray-black control images. I didn't want to create yet another QR Code model to compete with Monster Labs model which is already great. My idea was to focus more on natural and organic images for squint illusions as opposed to QR codes and take more regard to human perception as opposed to how cameras interpret contrast. You should notice a white color in the condition image become more of a light-color (HSV) instead of THE white-color (RGB). I also found the Monster Labs model tends to produce images with higher contrast, probably to make it work better with QR codes. While this model is able to produce scannable QR codes it was not the goal.
Usage
Use it by setting strength=0.6-0.8 and start_percent=0-0.1, end_percent=0.6-1.0 depending on your image concept. I generally found that you have to use lower ControlNet strength compared to other QR code models. Make a XYZ-plot with different start, end and weight strengths to get an idea how it influences the generation. You also don't have to use white-gray-black images but can use grayscale or RGB images as well. I also found that fine-tuned models produce more artistic results.
Templates
I have made a collection of different templates which you can download to get started:
Training
See my article Play in Control - ControlNet Training Setup Guide if you want to train your own control net!
The model was trained on the laion2b-en-aesthetics65 by converting images to grayscale using HSV method, Gaussian blur with kernel=15 and remapping colors to 33% white-gray-black. My idea was to imagine existing images as having been generated by a white-gray-black pattern image.
Versions
v0.4: 210000a32 samples with mixed_precision="fp16"
v0.3: 130000a4 samples with mixed_precision="fp16"
v0.2: 180000a4 samples with mixed_precision="fp16"
v0.1: 68000a4 samples with mixed_precision="no"
I will improve and train new versions based on usage and feedback.
References
Tutorials
Inspirations
reddit - Silhoutte illusions 1
reddit - Silhoutte illusions 2
Tools
QR code generator and verifier
Similar models
https://civitai.com/models/111006/qr-code-monster
https://civitai.com/models/90940/controlnet-qr-pattern-qr-codes
https://civitai.com/models/80536/lighting-based-picture-control-controlnet
https://civitai.com/models/90472/controlnet-qr-code
Thanks to antfu for the great in-depth articles and Monster Lab for the QR Code Monster model and the art generated henceforth inspiring me to get into control net training myself!
Description
Much better quality through better formula and higher batch size
New formula uses gaussian blur 15 and remap to 33% white-gray-black from value of HSV
Trained on 210000a32 samples with mixed_precision="fp16"
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.