The author used Stable Diffusion, an open-source text-to-image model, and ControlNet, a technique that guides the output of the model, to create images with hidden text. They also used a ControlNet that can produce natural-looking images that are valid QR codes. The author faced challenges in fine-tuning the visibility of the text and found that high-contrast scenes work best. They believe this technique opens up exciting possibilities in art, cryptography, and communication.
Key takeaways:
- The article discusses a technique of embedding text within AI-generated images using Stable Diffusion and ControlNet, where the text is hidden as a low-frequency component of the image and can be revealed by zooming out or squinting.
- Stable Diffusion is an open-source text-to-image model that translates a description of a scene into a picture, while ControlNet is a technique that can guide the output of a Stable Diffusion model, allowing for precise control over the image.
- The author used Modal to offload computation to a remote A10G GPU, significantly reducing the time it took to generate these images, and has published the script used for others to try.
- The technique has potential applications in art, cryptography, and novel forms of communication, and while it presents challenges, such as fine-tuning the visibility of the text, it opens up a new frontier for exploration.