How AI makes images and video: from random noise to a picture

How it learned: by destroying pictures

An image model learns backwards: it practises turning static back into pictures.

1. 1 It looks at millions of captioned pictures

   Image models are trained on huge collections of pictures paired with descriptions, mostly gathered from the web. One public collection, LAION-5B, has 5.85 billion of them.

   Source: arxiv.org

2. 2 It watches each picture dissolve into static

   In training, every picture is covered with random noise, a little at a time, until nothing is left but static, like an old TV with no signal.

3. 3 It practises undoing the noise

   At each step, the model guesses what noise was added so it can be taken away. The caption is a clue: knowing it's "a cat on a sofa" makes the guess easier. After billions of rounds, it can clean up static it has never seen.

How it makes your picture, in four steps

Now it runs the clean-up for real, with your words as the guide.

1. 1 You describe the picture

   You type a prompt, such as "a red cottage on green hills at sunset", or upload a photo to change.

2. 2 Your words become numbers

   A language model turns your prompt into a long list of numbers that captures its meaning: the same idea as the embeddings inside a chatbot.

   Embedding →

3. 3 It starts from random static

   The model begins with a canvas of pure noise. A different random start, called the seed, is why the same prompt gives a different picture each time.

4. 4 It cleans up, step by step

   It removes a little noise at a time, each step steered towards your description: rough shapes first, details last. Often a few dozen steps; the fastest models need only a handful.

From pictures to video

Same idea, with one big extra job: keeping every frame consistent.

1. 0s
2. 2s
3. 4s
4. 6s

• Many pictures, cleaned up togetherA video is a stack of frames. Video models remove noise from a whole block of frames at once, so the cat in the first second is the same cat ten seconds later.
• It learns how things usually moveTrained on huge amounts of video, models pick up how water splashes and hair blows. They learn what looks right, not the laws of physics, so objects can still melt, merge or vanish.
• Sound comes tooThe leading video models now also make matching sound: voices with moving lips, music and sound effects.
• Why clips are shortKeeping everything consistent gets harder, and costs far more computing power, the longer a clip runs. Most clips last seconds, not minutes, and longer videos are stitched together.

Why it still gets things wrong

It has learned what pictures look like, not how the world works.

• Hands and fingersIt learned what hands look like, not that people have five fingers. Newer models get this right far more often, but zoom in before you trust it.
• Words in the pictureSigns and labels used to come out as gibberish. The best models now write short text well; long text and small print still go wrong.
• Counting and placingAsk for three cups to the left of a plate and you may get four cups on the wrong side. Exact numbers and positions are hard for it.
• Physics in videoPeople walk through tables, a glass refills itself, a dog grows an extra leg as it turns. Watch the edges and the background.
• StereotypesAsk for "a CEO" or "a nurse" and models lean on patterns from their training data. A 2023 Bloomberg analysis of over 5,000 Stable Diffusion images found it exaggerated race and gender stereotypes.Source: bloomberg.com
• Copying what it sawOccasionally a model reproduces a famous picture, a character or even a watermark from its training data. That is part of why image AI ended up in court.

The tools you'll hear about

Not a ranking: these change every few months. For picking one, see AI tools for designers.

OpenAI closed its Sora video app in April 2026, about six months after launch, to focus its research elsewhere. engadget.com

Whose picture is it?

• Trained on other people's workMost image models learned from pictures collected from the web, usually without asking the artists and photographers who made them. Many are suing.
• The first big ruling went to the AI companyIn November 2025, England's High Court largely rejected Getty Images' case against Stability AI, finding that the model doesn't store copies of the photos. Getty won only a narrow point about its watermark showing up in some images. Other cases, especially in the US, continue.Source: clearygottlieb.com
• You may not own what you makeThe US Copyright Office says a picture made purely by AI from a prompt can't be copyrighted; real human creative input, such as arranging or editing it, can be. Rules differ by country.Source: copyright.gov
• Businesses play it safeThat is why many companies pick tools trained on licensed images, and check the terms before using AI pictures in ads or products.

How to tell if a picture is AI-made

Labels help, but none of them is foolproof.

• Content Credentials

  A "nutrition label" saved inside the file that says which tool made it and how it was edited, based on the C2PA standard backed by Adobe, Google, Microsoft, OpenAI and others. A screenshot or re-upload can strip it off.Source: c2pa.org

• Invisible watermarks

  Google's SynthID hides a mark in the pixels themselves, which survives cropping and resizing better. Google said over 10 billion pieces of content carried it by May 2025, but it only spots content from Google's own tools.Source: blog.google

• Labelling laws

  China has required visible and hidden labels on AI content since September 2025, India has required platforms to label it since February 2026, and the EU AI Act requires AI-made content to be marked as its rules take effect.Source: hoganlovells.com

No label doesn't mean it's real. For the warning signs of fake photos, cloned voices and deepfake calls, see how to spot AI fakes and scams.

The short version is in the glossary: What is a diffusion model? Related: generative AI, multimodal AI, deepfake and how AI chatbots work.