How AI creates images and videos, explained from scratch
How does AI create images and videos from a sentence? We explain from scratch, with everyday examples, what happens inside ChatGPT, Sora and Veo.
NinaKalkazu.org
Listen to this article
Narrated by Thalita
You type on your phone: "a cat with glasses reading a newspaper on the beach". In seconds, or a minute, a photo appears that no one ever took, with fur, shadow and the reflection of the sea in the lenses. It looks real, but that cat does not exist.
The machine did not search for this photo or cut pieces from others. Here you will understand, from scratch and without formulas, how AI creates images and videos, what happens in those seconds of waiting and why it still gets wrong things that a child gets right.
Quick summary: the answer in 30 seconds. AI learns by seeing millions of images with captions. To create, most tools start from random noise, like that of a TV off the air, and clean it gradually, guided by your sentence, until an image remains. ChatGPT assembles the image piece by piece, like writing a text. Video uses the same cleaning on many frames at once.
Index
- What does it mean for AI to "create" an image?
- How does AI learn what a cat is?
- How does AI create images from noise?
- And ChatGPT, does it create images the same way?
- How does AI create videos without things disappearing in the middle?
- Why does AI still get hands, letters and details wrong?
- How to know if an image or video was made by AI?
- What this changes for you
- Frequently asked questions
What does it mean for AI to "create" an image?
It means it produces a new image, point by point, from what it learned, without copying or pasting ready-made photos. Think of an illustrator who has seen thousands of cats. They do not store the photos in their head: they store the idea of a cat and, with it, draw a cat that never existed.
This type of program is called generative AI, the kind that generates texts, images and sounds.
A term that will appear often: pixel is each colored dot that forms an image on the screen. A phone photo has millions of them. For the machine, creating an image is deciding the color of each dot.
How does AI learn what a cat is?
It learns by seeing an enormous amount of images accompanied by text that describes each one. From seeing the word "cat" next to photos of cats so many times, it begins to connect the word to pointy ears, whiskers and fur. This study period is called training and happens before you use the tool.
It is like a child with a picture book with captions: after many pages, they recognize a cat they have never seen.
A real example: in 2021, OpenAI, owner of ChatGPT, introduced CLIP, trained with 400 million pairs of image and text taken from the internet, according to the scientific paper of the project. Its task was to guess which caption matched each image.
It became a kind of translator between words and images and, as this history of the technique on TechCrunch tells, began to guide creation tools. The training of text assistants follows the same logic, as we explain in how ChatGPT, Claude and Gemini learned.
The controversial side: most of these images came from the internet without permission from those who made them, and AI sometimes repeats the prejudices of the web, as this Wired retrospective shows.
How does AI create images from noise?
Most tools start with a screen full of noise, like that of a TV off the air, and clean that noise in several steps. At each step, the mess decreases and the image becomes more like what you asked for. This technique is called a diffusion model and is behind tools like Stable Diffusion.
To learn to clean, AI trains in the opposite direction: it takes a sharp photo, throws noise on top gradually until it becomes only noise and, at each step, tries to guess the added noise. Whoever knows what went in knows what to take out for the photo to return. After millions of photos, AI becomes an expert at this.
It is like a sculptor before marble: with each blow, they remove what does not look like a statue. AI removes, step by step, what does not look like a cat with glasses on the beach. Since the initial noise is randomized with each request, the cat comes out different each time.
What guides the drawing is your sentence
At each step, the translator between words and images checks whether what is emerging matches your sentence and pushes the cleaning in the right direction. With the opening cat, first spots appear: light below, which becomes sand, and blue above, which becomes sky. Then, a shape. At the end, whiskers, newspaper letters and the reflection in the lenses.
That is why details like "late afternoon light" make a difference: they are more clues for the cleaning. The word side is in how AI understands what you ask for.
The shortcut that saves computing power
Cleaning millions of pixels, step after step, requires a powerful computer. In 2021, researchers proposed cleaning a compressed version of the image, the latent space, a summary in numbers with only the essentials. The study showed that this greatly reduces the work. It works like a .zip file, "unzipped" into pixels only at the end. Stable Diffusion was born from this idea.
And ChatGPT, does it create images the same way?
Not exactly. Since March 2025, when OpenAI put image creation inside GPT-4o, the model that powered ChatGPT at the time, it assembles the image piece by piece, as it writes a text word by word. At the end, another program translates these pieces into colored dots.
ChatGPT breaks text into tokens (little pieces of words) and chooses one after another. In the image, the little pieces are sections of the figure, like a mosaic assembled tile by tile. Those who tested it at the time, like Ars Technica, saw the image emerge from top to bottom.
Since the same "brain" understands the conversation and draws, you can adjust by talking ("now put a hat on the cat"). In OpenAI's announcement, the company says the model writes letters better and handles 10 to 20 objects in a scene, versus 5 to 8 from other systems. The price is the delay: up to a minute per image. The current version is called ChatGPT Images 2.5 (checked on 10/27/2026).
| Diffusion (Stable Diffusion and others) | Piece by piece (ChatGPT) | |
|---|---|---|
| Starting point | Random noise | A blank canvas |
| How it advances | Cleans the entire image, several times | Places one piece after another |
| Strong point | Fast, good at textures and styles | Follows long requests and writes letters better |
| Weak point | Usually gets texts in the image wrong | Slower |
How does AI create videos without things disappearing in the middle?
A video is a sequence of photos, the frames, passing quickly. AI uses the same noise cleaning, but on all frames at once, not one by one. This way it manages to keep the same cat, with the same color and the same light, from the first to the last second.
Done frame by frame, the cat would change color and the glasses would disappear.
How Sora works: the video cut into little cubes
OpenAI solved this in Sora, its video tool. Imagine the frames stacked like slices of a loaf of bread, and the loaf cut into little cubes: each has a piece of the image and a few moments of time.
According to Sora's technical report, from February 2024, a transformer, the type of program behind ChatGPT, treats these little cubes like words in a text. It works like a reader who, at each word, recalls everything that came before so as not to lose the thread. The cutting also allowed training with videos standing or lying down, of any size. For the MIT Technology Review, this recipe became the industry standard.
And the sound?
Veo 3, from Google, creates effects, background noise and even speech along with the image, according to Veo's official page, today at version 3.1 (checked on 10/27/2026). It is like filming and recording sound at the same time, instead of dubbing later: that is why the mouth moves along with the speech.
Why does AI still get hands, letters and details wrong?
Because it does not understand the world like you do. It learned how things usually look, not how they work. In photos, a hand appears in a thousand positions, with fingers hidden and crossed. AI learns "about five fingers" and sometimes gets the count wrong. With letters, it learned the drawing, not how to spell: something that looks like text comes out, but does not form a word.
In video, it becomes strange physics: a glass that goes through the table, an extra arm for a second. That is why the result is a lottery, and promotional videos are usually the best among many attempts.
There is also an invisible cost: energy. A study by Hugging Face and Carnegie Mellon University showed that creating images was the AI task that consumed the most: on average 2.9 kWh per thousand images, roughly an electric shower of 5,500 W on for half an hour. Sorting a thousand texts by type consumed only 0.002 kWh. In the least efficient model, each image consumed about half a phone charge.
How to know if an image or video was made by AI?
There is no foolproof trick, but you can suspect with method: look at the details AI usually gets wrong, search for the origin of the image and check the digital seal. The most important thing is the source: who published first and whether a serious outlet confirmed it.
- Look at the details: fingers, teeth and letters on signs in the background.
- In video, look for what changes without reason: objects that disappear, mouth out of sync with the voice.
- Find out where the image came from: send the image itself to Google Lens, which shows where it has already appeared.
- Look for the digital seal. OpenAI put C2PA in GPT-4o images, data hidden in the file that records the origin, and Veo videos carry SynthID, a watermark the eye does not see. But the label can be lost in a screenshot.
- Distrust what is too shocking and has no source.
What this changes for you
Understanding how AI creates images helps in two ways: you get better results from the tools and fall less for fake images.
| Profile | What to do with this |
|---|---|
| Curious person who wants to test | Describe light, angle and style in the sentence |
| Content creator | Use for drafts and avoid faces of real people |
| Parents and teachers | Show children that a perfect photo can be invented (see how to protect your child) |
| Those who receive news by WhatsApp | Before forwarding, check the source |
Frequently asked questions
Does AI copy images from other people?
It does not cut or paste pieces of photos. But it learned from works by other people, almost always without permission, and can imitate styles very closely.
What is a diffusion model?
It is the technique that creates images by cleaning noise gradually. In training, AI learns to remove noise from photos; when creating, it starts from random noise and cleans it, guided by your sentence.
How does Sora work?
It cuts the video into little cubes of image and time and cleans the noise from all of them together, which keeps people and objects stable.
Does AI understand what it is drawing?
Not the way you understand. It knows how a cat usually looks, but does not know what a cat is. That is why it gets hands, letters and the physics of objects wrong.
Next time a cat with glasses appears on your screen, you will know what happened: a machine that saw millions of cats, but never knew what one is, assembled the scene for you. There are more explanations like this in the artificial intelligence section.
