qrtist

[2023]

Project Duration:
3 weeks

Collaborators:
Yannis Zarkadas

Personal roles:
Co-Engineering

A project developed from front to back-end intended to run on the web, QRtist made use of pre-trained machine learning models and was based on a method which was quickly gaining popularity at the time in conditioning the output of generative image models for very specific results. We used the available resources to further experiment on creating QR codes within a stylized context.

The building block behind the QRtist process is an AI image generation model, such as stable diffusion, conditioned through ControlNet (a method most used to set a person’s pose in generated images) on QR code images. This results in the possibility of creating any image in any style while maintaining the aspects that make a QR code functional and recognizable.

Part of our process revolved around striking a balance between integrated and useable results.

Several experiments took place to reach this balance, using Pytorch to iterate further on the conditioned model found in Huggingface. Besides observing the effect of the conditioning scale on the aesthetics and usability of the QR codes, we found that the three square anchors were significant in the detection of a code and to that end we implemented an additional step which strengthened their presence in the generated image. The result was a more scannable code at a lower conditioning scale, meaning that the code parts were better integrated and the functionality remained higher.

To further accomodate public use and experimentation with this new at the time concept, we set up a website which ran our iteration of the model hosted on Runpod. The site remained live on the web for a short period that we could support for free.

An initial iteration of the website that was built to host the model.

Github Repository (team account)