Interactive demos with real-time audio.
Make sure your device isn't muted.
Noise, latent space, and the imagination of AI
In 1968, Philip K. Dick asked: "Do androids dream of electric sheep?"
In 2022, a machine answered β by generating images from pure noise, guided by nothing but words.
In Part I, you saw how diffusion destroys information irreversibly. Now discover how a machine learned to run it backward β to dream shapes out of chaos, just as we see faces in clouds.
00 β A WARM-UP PUZZLE
In 1967, mathematician Vladimir Arnold discovered something astonishing. Take any image β say, a picture of a cat. Apply this formula to every pixel:
The image looks completely destroyed β indistinguishable from noise. But keep applying the same formula, and after exactly 12 steps, the original image magically reappears. Perfectly. Every pixel back in place.
Why does this matter? β Arnold's cat map is deterministic and invertible. It's a linear map on the torus β no information is ever lost. Every pixel has one and only one destination, and one and only one origin. The system is guaranteed to return (PoincarΓ© recurrence).
Now contrast this with diffusion: when you add random noise to an image, the destruction is stochastic and irreversible. You cannot simply reverse it β the randomness has erased the path back. To undo diffusion, you need a neural network to learn the reverse mapping from data. That's the miracle of Stable Diffusion.
Arnold's cat comes back on its own.
Diffusion's cat needs AI to find its way home.
01 β SEE IT HAPPEN
Below is pure random static. Press Denoise and watch as a hidden image materializes step by step β exactly what Stable Diffusion does millions of times per day.
02 β HEAR IT HAPPEN
A simple melody is buried under noise. Drag the slider to gradually denoise it. At some point your brain will suddenly "catch" the melody β the magic of denoising.
03 β READ IT HAPPEN
Drag the slider to denoise random pixels into a readable word. Just like the sound demo β at some noise level, your brain suddenly recognizes the text. That moment of recognition is the threshold where diffusion models also "see" structure.
04 β ENHANCE
A blurry photo is a partially noised version of a sharp one. Drag the slider to enhance resolution β the model fills in missing detail.
05 β THE SECRET
Everything you just experienced uses the same three-step process:
"Teach a machine how photos become static,
and it learns how to dream photos out of static."
We see faces in clouds. Machines see images in noise.
Both are finding patterns where none were placed β the essence of dreaming.
05b β THE BREAKTHROUGHS
The idea of reversing diffusion had existed since 2015, but the results were blurry and unconvincing. Then in 2020, Jonathan Ho, Ajay Jain, and Pieter Abbeel at UC Berkeley published a paper called "Denoising Diffusion Probabilistic Models" β DDPM for short β and everything changed.
Their insight was beautifully simple: don't try to predict the clean image directly. Instead, just predict the noise.
Here's what that means. Take a photo. Add a known amount of Gaussian noise to it. Now show the noisy result to a neural network and ask: "What noise was added?" If the network can answer that correctly, you can subtract the predicted noise β and recover a slightly cleaner image. Repeat this 1,000 times, starting from pure static, and a brand-new image appears from nothing.
The training objective turned out to be absurdly simple β just a mean squared error:
In plain English: measure how far off the network's noise prediction (Ρθ) is from the actual noise (Ξ΅) that was added. That's it. This one equation trained a model that produced images rivaling GANs for the first time β and it was far more stable to train.
The catch? DDPM was slow. Generating one image required running the neural network 1,000 times in sequence, each time denoising a little more. A single image could take minutes on a powerful GPU.
Stable Diffusion is the name of a specific open-source model created by a team at LMU Munich (Robin Rombach, Andreas Blattmann, and others) in collaboration with Stability AI. It solved DDPM's speed problem with one brilliant idea:
Don't diffuse in pixel space. Diffuse in a compressed space β the machine's unconscious.
A 512Γ512 image has 786,432 pixel values. Running diffusion on all of them is expensive. So Stable Diffusion first compresses the image into a tiny latent representation β just 64Γ64Γ4 = 16,384 numbers β using a pre-trained autoencoder. Then it runs the diffusion process (forward and reverse) entirely in this compressed space. Finally, it decodes the result back into a full image.
This made generation ~50 times faster. But there is something poetic about this latent space: it is a vast mathematical ocean where every concept humanity has ever photographed β every face, landscape, animal, texture β exists as a point. It is, in a very real sense, the machine's collective unconscious. But that's only half the story. Stable Diffusion also added a text encoder (CLIP) that understands language. When you type "a cat sitting on a mountain at sunset," CLIP translates those words into a numerical vector that guides the denoising process at every step β like dropping a pebble into the machine's unconscious ocean, creating ripples that steer the noise toward an image matching your words. The prompt is the trigger. The latent space is the unconscious. The denoised image is the dream.
The result: type a sentence, wait a few seconds, get a photorealistic image. And because Stability AI released the model weights openly, anyone in the world could use it, modify it, and build on it β triggering the explosion of AI-generated art you see today.
How they all fit together:
06 β TRY IT YOURSELF
Pick a prompt and watch every stage of the pipeline: your text guides the denoising, noise fills the tiny latent space, the U-Net removes noise step by step, and the decoder reconstructs the full image.
07 β THE MATHEMATICS
Now see the math behind it. Dissolve a signal into noise, then watch 300 particles converge from chaos into structure.
08 β THE REVOLUTION
09 β THE CONNECTION
| Physics | Stable Diffusion | |
|---|---|---|
| π‘οΈ | Heat spreads until uniform | Noise added until pure static |
| π | Gaussian bell curve | Noise is exactly Gaussian |
| π² | Brownian motion | Random noise injection |
| βͺ | Diffusion is irreversible | Neural net learns to reverse it |
| βοΈ | Equilibrium = no info | Pure noise = no image |
It finds faces in clouds of Gaussian noise.
It wanders a vast unconscious of compressed images.
It hallucinates form from pure chaos β guided by words.
Philip K. Dick asked whether androids dream of electric sheep.
Perhaps the answer is simpler than he imagined:
A machine dreams of whatever you ask it to β
using mathematics that Fourier, Einstein, and Anderson
wrote down long before anyone thought to ask.
The diffusion equation took 200 years to dream its first image.