Type "an astronaut riding a horse" into an AI image tool and seconds later you have a picture no camera ever took. The machine doesn't draw it, doesn't copy it from a photo library, and doesn't understand it the way you do. It does something stranger: it starts with television static and cleans it up, step by step, until a picture appears. The technique is called diffusion, and it sits behind nearly every AI image tool you have used — Midjourney, DALL-E, Stable Diffusion. This is a research-based explainer: no hands-on testing, just how the machinery works, and why it keeps tripping over hands.
Training a diffusion model begins with destruction. Take a real photo and add a little random noise — the fine speckle of an old TV with no signal. Add a bit more. Keep going, step by step, until nothing is left but static. Then teach a neural network to reverse the process: show it a noisy image and the stage it is at, and train it to predict the noise that was added. Repeat that on millions of images and the network gets very good at one skill — undoing damage. To make a new image, you run the process backwards. Start with a canvas of pure random noise, ask the network to remove a little of it, and repeat twenty or fifty times. First come broad blobs of light and colour, then shapes, then details. The modern version of this idea took shape in a June 2020 paper by Jonathan Ho, Ajay Jain and Pieter Abbeel, "Denoising Diffusion Probabilistic Models", which showed that this simple loop could produce high-quality images.
Denoising alone gives you a random image. Two more ideas turn it into the picture you asked for. The first is latent diffusion: instead of working on every pixel, the model compresses the image into a smaller "latent" representation that keeps what matters visually, does the noising and denoising there, and decodes the result back into pixels at the end. That shortcut, published by Robin Rombach and co-authors in December 2021, is what made diffusion fast enough for everyday use. The second is conditioning. A text encoder turns your prompt into a set of numbers, and cross-attention layers let those numbers nudge the network's guess at every step — pixels that should become a wheel are denoised towards wheel-like patterns. A setting called guidance strength decides how hard the model obeys the prompt: too low and it ignores you, too high and the image goes garish and overcooked.
If diffusion models are so good, why do they keep drawing hands with six fingers? Because a hand is a hard object for statistics. It has 27 bones, endless poses, and it is usually small in the frame and often half-hidden behind something. The model has no anatomy and no idea what a hand is for — it only knows what hands tend to look like in photographs. So fingers merge, extra digits appear, and a thumb lands in the wrong place. The same gap explains garbled text in AI images: the model knows what letters look like, not what they mean. Better models and more steps reduce the errors, but the underlying gap stays — the model predicts pixels, not objects, and a six-fingered hand is statistically plausible even when it is anatomically wrong.
In August 2022, Jason M. Allen of Pueblo, Colorado, entered three pieces in the Colorado State Fair's fine-art competition, in the Digital Arts/Digitally-Manipulated Photography category. One of them, "Théâtre D'opéra Spatial" — a baroque vision of robed figures in a cosmic hall — won first place and a $300 prize. Allen had made it with Midjourney, submitting it under the name "Jason M. Allen via Midjourney", after more than 80 hours of work, hundreds of prompts and revisions, and further editing in Photoshop. The two judges later said they had not known Midjourney used AI to generate images — and that they would have awarded the top prize regardless. When the news spread online, the reaction split: cheating, or a new medium? In September 2023 the US Copyright Office ruled that the AI-generated elements of the work were not eligible for copyright protection. The argument hasn't settled. The tool hasn't gone away either.
Diffusion models don't imagine, and they don't steal a single photo. They learn the statistics of millions of images and rebuild from noise, guided by words. That is why they are brilliant at plausible and terrible at precise: a hand with six fingers is statistically fine and anatomically wrong. The fair prize showed the real argument isn't about the tool — it's about what we call art when the brush is a prompt.
Related on Poptani Labs: Why AI can't count the R's in 'strawberry' — and what that tells you, Why AI sometimes makes things up — and why it sounds so sure, Why your chatbot now remembers you — AI memory, explained. More in this category: AI, Decoded.
Sources: Wikipedia, Théâtre D'opéra Spatial, Smithsonian Magazine, AI art wins at state fair, Mindox AI, How Do Diffusion Models Work?, Cloudinary, Diffusion Model: How It Works