---
title: Diffusion Model
description: >-
  What a diffusion model is, how iterative denoising generates photorealistic
  images, and the role of diffusion in image-based virtual try-on.
shortDefinition: >-
  A generative AI architecture that creates images by learning to reverse a
  gradual noising process.
relatedTerms:
  - image-inpainting
  - garment-transfer
heroImage: /glossary-images/diffusion-denoising-steps.webp
keywords:
  - diffusion model
  - generative AI
publishedAt: '2026-06-10'
updatedAt: '2026-06-10'
noindex: false
canonicalPath: /glossary/diffusion-model
---

## What is a diffusion model?

A **diffusion model** is a generative AI architecture that creates images by learning to reverse a noising process. During training, images are progressively destroyed with random noise and the model learns to undo each step. At generation time, it starts from pure noise and denoises step by step into a new image, guided by a text prompt, a reference image, or both.

Diffusion models underpin most state-of-the-art image generators, including the [garment transfer](/glossary/garment-transfer.md) systems used for image-based virtual try-on.

![A diffusion model denoising from random noise to a finished image](/glossary-images/diffusion-denoising-steps.webp)

## How it works

1. **Forward process (training)**: noise is added to a real image in many small steps until only static remains.
2. **Learned reverse process**: the model is trained to predict the previous, less noisy state at each step.
3. **Generation**: starting from random noise, the learned denoising is applied repeatedly (typically 20–50 steps) until a coherent image emerges.
4. **Conditioning**: at every step the model can be steered by inputs such as a garment image, a body pose, or a mask of the region to repaint.

## Role in virtual try-on

Conditioning is what turns a general image generator into a try-on engine: given a person's photo, a garment photo, and a clothing mask, the model regenerates only the masked region (see [image inpainting](/glossary/image-inpainting.md)) with fabric texture, wrinkles, and lighting consistent with the original photo.

The trade-off is speed: generation takes seconds and requires GPU compute, which is why diffusion-based try-on produces static images rather than live video overlays.

<Faq>
  <FaqItem question="Why did diffusion models replace GANs for try-on?">
    Earlier try-on research used GANs, but diffusion models proved more stable to train, more controllable through conditioning, and better at reproducing fine texture. All of these matter for clothing.
  </FaqItem>
</Faq>