What Is Diffusion Model? | AI Tester Dictionary #007

6 Просмотры

Спасибо! Поделитесь с друзьями!

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

Издатель Sep 2, 2026

A model that writes text predicts the next word.

A model that makes pictures does something stranger. It starts with pure static, the kind of noise an old television made, and takes the noise away a bit at a time until a picture is sitting there. Nothing was drawn and nothing was copied. It came out of the static, one step at a time.

Welcome back to the AI Tester Dictionary — one AI term at a time, explained for people who test software.

Episode seven is the diffusion model, the other half of generative AI. #006 covered transformers, which run text generation. Most image generation runs on diffusion models instead.

A diffusion model learns to create things by learning how to undo damage. Take a real image, add a little noise, then a little more, until nothing of the original is left. Show a model that happening millions of times and it learns to run each step backwards. Nobody has to label any of it, because the answer is always known — the model is shown a noisy version and asked what the slightly cleaner version looked like.

Generating runs the same steps in the other direction. It starts from noise nobody has ever seen before, freshly made from random numbers, and your prompt goes in alongside it. Ask for a cat sitting on a laptop and every step of the cleaning is aimed at that. After enough steps there is an image that was never in the training data.

The name comes from what happens when you put a drop of ink into water. It spreads until the water looks the same everywhere and the drop is gone. Noise does the same to the information in an image, and learning to run that backwards is called denoising.

⏱️ Chapters
0:00 Static in, picture out
0:21 What the AI Tester Dictionary is
0:39 Learns to undo damage
1:01 Phase one — training
1:23 Phase two — generation
1:47 Why it is called diffusion
2:06 Any subject, same static
2:17 Two ways to generate
2:42 Not only images
2:56 Noise in, picture out