AI Trend Research

← AI glossary

AI glossary · Basic · Generative media

Diffusion model

A diffusion model is a generative AI model that starts from random noise and removes the noise little by little to create data such as images, video or sound. During training, it learns to identify the noise in noisy data.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-08
In plain words It is similar to gradually wiping a fogged-up window so that the view outside becomes clearer and clearer. In a diffusion model, though, there is no fixed view behind the window to begin with. The difference is that the wiping process itself creates a plausible new view.

How it works

During training, noise (random static) is first mixed into real images step by step until they become unrecognizable. The model is then asked to identify 'what noise was mixed into this image'. By repeating this practice a very large number of times, the model learns how to remove noise. When generating something, it starts from pure noise and repeats the process of predicting the noise and removing it bit by bit many times. At the end, a clean image comes out.

Training and generation in a diffusion modelOriginal image → Learn to guess the noise; Learn to guess the noise → Predict noise(Trained model); Random noise → Predict noise; Predict noise → Remove a little; Remove a little → Predict noise(Repeat many times); Remove a little → Finished image① Training② GenerationOriginal imageAdd noise step by stepLearn to guess thenoiseRandom noisePredict noiseRemove a littleFinished imageTrained modelRepeat many times
Training and generation in a diffusion model The top row is the training stage, where the model learns to guess the noise in images that have had noise mixed in. The bottom row is the generation stage, where the trained model repeatedly predicts the noise and removes a little of it to complete the image.

Why it matters

Diffusion models are the dominant approach in today's image generation AI, thanks to their detailed and diverse outputs and relatively stable training. When text (a prompt) such as 'a cat on a beach at sunset' is added as a condition, the model creates an image that fits that description as it removes the noise. The same principle is expanding to video, speech, music, 3D, robot motion generation and more.

Things to know

Because the noise-removal step has to be repeated many times, generation is slower and more computationally expensive than with methods that produce a result in one pass. Techniques for reducing the number of steps are therefore an active area of research, and using distillation to cut the number of steps is one of them. To save cost, removing noise in a small compressed space instead of on the original image (latent diffusion) is also widely used. Social issues such as copyright of training data and misuse for deepfakes are also being discussed.

Example

Image generation models such as Stable Diffusion are typical diffusion models. When a user describes the scene they want in text, the model starts from random noise and removes noise in line with that description to complete the picture. The same principle is now also used in text-to-video services, voice and music generation tools, and editing features that erase part of a photo and fill it in naturally.

Related terms

GAN Multimodal Autoregressive model Distillation Foundation model Prompt