---
title: "Diffusion Model"
url: "https://mahouse.world/glossary/diffusion-model"
type: "glossary term"
topic: "Generative AI"
author: "Ihor Vlasov"
reviewed: "2026-09-20"
published: "2026-08-10"
---

# Diffusion Model

A diffusion model is a generative model that learns to produce data — images above all, also audio, video and molecules — by learning to reverse a process that gradually turns data into noise: trained to remove a little noise at each step, it can start from pure noise and denoise its way to an image matching a text prompt. Stable Diffusion, DALL·E and Midjourney are diffusion models; they make the synthetic documents and faces fintech identity teams detect.

**Also known as:** Diffusion models, Denoising diffusion, Latent diffusion

## Key facts

- **Used at:** image, video and audio generation for design, media and advertising; scientific applications (protein and molecule design); and, on the other side, document forgery, deepfake identity attacks, and synthetic evidence in disputes and insurance claims
- **Binding:** no direct regulation of the method; the EU AI Act requires labelling of AI-generated content and transparency for deepfakes; identity-verification standards increasingly require liveness and injection-attack detection because of it
- **Typical mechanics:** a forward process adds Gaussian noise to training images over many steps; a neural network (usually a U-Net or a transformer) is trained to predict the noise added at each step; generation runs the network backward from random noise, guided by a text embedding; latent diffusion (Stable Diffusion, 2022) runs the process in a compressed latent space, making it fast enough for consumer hardware
- **Typical scale:** training on hundreds of millions to billions of image-text pairs; generation in seconds on a GPU; open-weight models mean anyone can run one, including fraudsters
- **Parties:** the labs and companies building the models, the creative industries using and contesting them (the copyright suits over training data), and the fraud, identity and trust-and-safety teams defending against their outputs

## How a diffusion model works in practice

Training: take an image, add noise in small increments until it is static; teach a network to predict, at each increment, what noise was added, given the noisy image and the step; repeat across a billion images with their captions, so the network also learns how text relates to what it is denoising. Generation: start with static, ask the network what noise to remove to move toward "a passport photo of a man in his forties, neutral background", remove it, repeat fifty times, and a face that never existed appears — photorealistic, consistent, and adjustable by prompt. Latent diffusion does this in a compressed representation, which is why it runs on a laptop.

The fintech consequence arrived quickly. Identity verification relies on documents and faces; a diffusion model produces both — a plausible passport with a plausible face, or a face matching a stolen document — and an injection attack feeds the generated image to the verification app in place of the camera. Vendors responded with liveness detection that tests for a real, present person, injection detection that checks the camera feed is a camera, document forensics that look for generation artefacts, and consortium data on faces seen too often. Fraud teams see the same in disputes: generated receipts, generated delivery photos, generated screenshots. The arms race is the current state.

For AI M&A the diffusion companies were the 2022–2023 rounds (Stability, Midjourney's revenue, Runway); for fintech M&A, the identity-verification vendors' deepfake defences are now a diligence item.

## Example

An identity-verification vendor sees, over 2023–2024, a rise in applications where the selfie is a generated face injected into the capture flow: from near zero to 2% of attempts at some customers. Its response: passive liveness (analysing the image for the physical signals of a real face in front of a camera), active challenges where risk is high, camera-feed integrity checks, and a model trained on generated faces to detect them. Detection of injected synthetic faces rises from 60% to 97% over a year; genuine applicants see no added friction on 94% of flows. A neobank customer of the vendor that had disabled liveness "for conversion" in 2022 re-enables it after its fraud team traces 400 mule accounts to synthetic selfies.

## Common mistakes

- Identity verification that matches a face to a document and never asks whether the face is real.
- Treating deepfakes as a future problem while the injection attacks run today.
- Disabling liveness checks for conversion.
- Assuming detection is solved; the generators improve every quarter and the detectors must too.
- Reading "AI-generated" labels as protection when the fraudster's model does not add them.

## Related terms

- [Generative AI](https://mahouse.world/glossary/generative-ai.md)
- [Generative Adversarial Network](https://mahouse.world/glossary/generative-adversarial-network.md)
- [Latent Space](https://mahouse.world/glossary/latent-space.md)
- [Multimodal Model](https://mahouse.world/glossary/multimodal-model.md)
- [Synthetic Identity](https://mahouse.world/glossary/synthetic-identity.md)
- [Training Data](https://mahouse.world/glossary/training-data.md)

— Ihor Vlasov, M&A FinTech Expert. Last reviewed 2026-09-20. https://mahouse.world/authors/ihor-vlasov