What is a diffusion model? The AI tech behind every image generator

8 Просмотры

Спасибо! Поделитесь с друзьями!

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

Издатель Jul 22, 2026

From Zero · Ep 67 — Every image AI you have ever used runs on one technique. Stable Diffusion. DALL·E 3. Midjourney. Imagen. Every nudify app. All the same core idea. From zero, kept simple, no jargon.

The definition — one line. A diffusion model is a neural network trained to reverse a noise-adding process. That is it. The rest is engineering.

Forward — destroy the image:
1. Take a real image from a training set
2. Add a small amount of random Gaussian noise
3. Repeat hundreds to thousands of times
4. After enough steps, the image is pure static

Reverse — train a network to undo it. Train a network to predict which noise was added at each step. If it can predict the noise, it can subtract it, and recover the original image one step at a time.

The network doing the work — the U-Net. Encoder compresses the image, decoder rebuilds it, skip connections between matching layers preserve detail during compression. That U-shape gives the network its name.

Text conditioning — the clever part. Feed a text prompt into a text encoder like CLIP or T5 to get an embedding. Inject that embedding into the U-Net at every denoising step. Now the network denoises toward the prompt.

Sampling — how you actually generate an image:
1. Start from pure Gaussian noise
2. Feed noise + text prompt into the U-Net
3. Take one denoising step
4. Repeat 20 to 50 times — an image emerges

Classifier-free guidance. Run the model twice — once with the prompt, once without. Push the result away from the unconditioned version and toward the prompted one. Higher guidance = sharper adherence to the prompt. This is the CFG slider you see in every diffusion UI.

The products you know — different labels, same technique:
- Stable Diffusion (open source, Stability AI)
- DALL·E 3 (OpenAI, in ChatGPT)
- Midjourney (independent, Discord-first)
- Imagen (Google, in Gemini)

All four follow the same forward + reverse + text-condition structure.

Image-to-image — the nudify-app path. You do not have to start from pure noise. Start from an existing image. Add some noise but not all the way to static. Feed partially noised image + text prompt into the U-Net. It rebuilds the image guided by both. That is how every nudify app works.

Recap: two moves + one guide. Forward adds noise. Reverse removes noise. Text steers the removal. Everything else is optimization.

Companion to Ep 66 — the SF cease-and-desist against Apple + Google over 13 nudify apps.

This is From Zero. Subscribe for honest tech takes, every day.

#fromzero #tech #ai #diffusion #stablediffusion #dalle #midjourney #news #2026 #elitedevcomparisons