Show HN: Embedding hidden text in AI images with ControlNet and Modal

[jlowin]

A few weeks ago, a technique was shared for hiding text in the negative space of an AI-generated image and I thought it was an amazing illusion. However, the only tutorial I could find was light on details and UI-based (and seems to have disappeared?). To continuing exploring on my own, I wrote a script to generate this style of images programmatically using various Stable Diffusion checkpoints and a QR code-optimized ControlNet, running on Modal. Let me know if it works for you!

[sparrish]

Some I can't see the words until I squint a bit, then it really pops out.

They all seem to be lighter lettering on dark backgrounds. It would be more subtle if it could do dark lettering on lighter backgrounds.

[jlowin]

I bet it would work if you invert the control image!

[quickthrower2]

This is amazing! If i get time I might try to script it to run on typical rented GPU setups (Docker + SSH / Docker + script). Modal’s api is a bit different from that.

I can see this becoming the 2023 version of diorama effect, and we’ll be well bored of it by 2024 when every Medium article had one. This is a compliment by the way (like the best music is overplayed)

[lulzx]

Pretty cool application