image generation via diffusion
/ dih-FYOO-zhun IM-ij jen-er-AY-shun /
Diffusion is the technique behind most modern image generators. The idea is almost paradoxical: teach a model to create pictures by first teaching it to destroy them. During training, you take real photos and add random noise step by step until each becomes pure static, and the model learns to reverse one step — to make a slightly noisy image slightly cleaner. To generate something new, you start from a fresh field of pure noise and run that learned cleanup over and over, dozens of times, until a coherent image emerges, as if a photograph were developing out of snow.
What makes it controllable is steering that cleanup with a text prompt. A separate text-understanding model translates "a fox in a red scarf" into guidance, and at every denoising step the image is nudged toward matching those words. Most practical systems do the work in a compressed "latent" space rather than on raw pixels, which is why it is fast enough to run on consumer hardware. The result is the photorealistic and painterly image generation that became famous around 2022.
The honest framing matters because the hype is loud. A diffusion model does not understand a scene; it has learned the statistics of millions of training images and produces plausible new arrangements of what it has seen. So it can be strikingly creative and also confidently wrong — garbled hands and text, impossible physics, faces that never existed. It can echo biases and styles from its training data, raising real questions about copyright and consent, since the data was often scraped without permission. And it makes fabrication trivially easy, which is the engine behind deepfakes. Beautiful output is not the same as truthful output.
Type "an astronaut riding a horse on the moon, photorealistic." The model starts from random static and, over ~30 denoising steps each nudged by your words, resolves a convincing image. Look closely and the horse may have five legs and the astronaut's glove may blur into the reins — plausible at a glance, broken under scrutiny.
From noise to image in many small steps, steered by text — convincing at a glance, often flawed in the details.
A diffusion model recombines statistics from its training images; it does not look anything up or understand the scene. That is why it invents people who never existed, fumbles hands and text, and can reproduce copyrighted styles — and why "photorealistic" should never be read as "real" or "true."