Inspired by Jorge Luis Borges' "The Library of Babel" and Will Depue's original CanvasofBabel.com, The Canvas of Babel v2 extends this concept to real distribution from diffusion models.
The original Canvas of Babel generates random pixel noise for each coordinate. While philosophically interesting, pure noise rarely produces meaningful images. The probability of randomly generating even a simple recognizable shape is astronomically low.
By using a diffusion model (FLUX.1), we sample from the actual distribution of "possible images" rather than pure randomness. Each coordinate still maps to a unique, deterministic image through its seed, but now these images exist within the learned manifold of visual meaning.
Every position (x, y) on this infinite canvas corresponds to a unique image. The same coordinate will always show the same image, creating a stable, explorable universe of visual possibilities. You're not just viewing random images—you're navigating a deterministic space where every image has always existed at its specific location, waiting to be discovered.
Somewhere in this plane exists every possible image: every photograph that could be taken, every artwork that could be created, every moment that could be captured. The challenge is finding them in the vastness of the infinite canvas.
• Each tile position is hashed to create a deterministic seed
• Seeds are passed to FLUX.1 [schnell] with an empty prompt
• Generated images are cached to ensure consistency and reduce costs
• The canvas extends infinitely in all directions
"The universe (which others call the Canvas) is composed of an indefinite, perhaps infinite, number of hexagonal galleries..."