How deepfakes are made

Published June 16, 2025 Updated October 27, 2026

TLDR: Five steps turn public photos and recordings into a deepfake. AI deepfakes are audio, video, or images that a machine-learning model generates to show a real person saying or doing something they never did. A creator collects samples of the person, cleans them, trains a model, generates new media, and edits out visible flaws. Older tools trained generative adversarial networks (GANs). Many newer ones use diffusion models. Each new method changes what detection has to look for.

A note on scope

his guide explains how deepfakes are made so security, trust and safety, and fraud teams know what they’re checking for. It describes the technology, not a method to follow, and it doesn’t link to or recommend generation tools. If you’re new to the topic, start with what a deepfake is. 

A simple explanation of how deepfakes are made

Deepfake technology relies on AI systems that learn to generate convincing media of a real person by analyzing large amounts of data. Creators train a model on hundreds or thousands of images, video clips, or audio samples of one individual.

The AI system studies patterns in facial expressions, voice characteristics, speech patterns, and visual features. It then learns to generate new content that mimics the person’s appearance or voice. Older tools did this with generative adversarial networks (GANs). Many newer ones use diffusion models. In a 2026 vishing study, listeners averaged 37.5% accuracy telling AI-generated voices from human ones.

Older methods needed hundreds or thousands of samples to train a convincing model on someone’s visual and audio patterns. Newer voice and image tools need far less.

The deepfake creation process

• Step 1 (Data collection): Gathering source material, including photos, videos and audio recordings of the target person, often scraped from social media or public appearances.
  
• Step 2 (Data preprocessing): Cleaning and organizing the collected material, extracting faces from videos, isolating voice samples, and preparing data for training.
  
• Step 3 (Model training): Training a generative model, such as a GAN or a diffusion model, on the prepared data.
  
• Step 4 (Generation): Running the trained model to create new synthetic content, fine-tuning parameters to improve realism and believability.
  
• Step 5 (Post-processing): Refining the output through editing software to eliminate obvious artifacts and enhance overall quality.

Time required varies with the quality the creator wants and the computing power available. Off-the-shelf tools produce rudimentary deepfakes quickly, while high-quality results can take days to weeks of processing on more powerful systems.

Schedule a 30-minute walkthrough to see how RealScan checks uploaded audio, video, and image files for signs of manipulation.

The tools and models behind deepfakes

Deepfake software ranges from user-friendly consumer applications to sophisticated development frameworks, including mobile apps that create basic face swaps and professional-grade platforms that require technical expertise.

Generative adversarial networks pit two neural networks against each other. One generates content and the other tries to spot it, and both improve until the output is convincing.

How deepfakes use AI diffusion

Diffusion models learn by adding noise to real images until nothing recognizable is left, then learning to reverse that process. Once trained, a model starts from noise and builds a new image or video frame that matches a prompt or a reference face.

Cloud services and simpler interfaces have put these tools in the hands of people with no technical background. “Generative artificial intelligence tools make it easy for even low-skill threat actors to create deepfakes,” according to FS-ISAC’s report Deepfakes in the Financial Sector: Understanding the Threats, Managing the Risks.

The traces detection looks for

Generation methods leave traces, including visual inconsistencies, audio artifacts, and irregular movement. Detection models look for those traces rather than relying on what a person can see or hear. Reality Defender’s RealScan uses inference-based models to flag signs of manipulation or synthetic generation in uploaded audio, video, and image files.

If your team reviews uploaded or inbound media, check it before anyone acts on it. Detection is a continuous discipline, because every new generation method changes the traces a model has to find.