How AI Image Generators Work: Diffusion, Explained #Shorts

3 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Jul 13, 2026

How does typing 'an astronaut riding a horse' become a photo that never existed? Image models like Midjourney and Stable Diffusion do it with a strange trick: they start with pure noise and remove it. That's diffusion. Here's the counterintuitive part: to learn to create images, the model first learns to destroy them. You take a real photo and add noise, step by step, until it's pure static. Then the model learns to run that in reverse - given a noisy image, predict the noise and subtract it. Do that starting from nothing but static, and an image appears. So generation works like this: start with a canvas of random noise; the model predicts what noise to remove to nudge it closer to the prompt; subtract a little; then look again and repeat, often dozens of times. Each pass, the static resolves a bit more, until a sharp image emerges. Your text prompt steers every step (that's called conditioning). It's sculpting a picture out of noise. The cost is speed: every denoising step is a full run of the model, which is why images take seconds, not milliseconds. It can still fumble fine details (hands and in-image text used to be classic tells, though the newest models have largely fixed both), and it only knows what it was trained on. The big speed fix is latent diffusion - denoising in a compressed space instead of on every pixel. This one idea runs the whole image-generation boom: Stable Diffusion, Midjourney, Flux, and Google's Imagen are all diffusion models. And it didn't stop at images - the same denoising trick now generates video, audio, and even 3D shapes.

Full lesson (11 free, no signup): https://systemdesign.academy/go/diffusion

#Diffusion #AIArt #StableDiffusion #Midjourney #AI #MachineLearning #GenerativeAI #SystemDesign #Shorts