How AI Image Generation Works

Developer note: Series 3, Episode 2 deep dive for Paragon AI MAG. How does AI image generation work? The mechanism: diffusion models learn to reverse a noising process — start from random static, predict and remove noise over 20 to 50 passes, guided by a text embedding. CLIP (400 million image-text pairs) is the bridge between words and pictures; latent compression is what let Stable Diffusion run on 6.9 GB of home graphics memory. We also cover the memorisation finding (Carlini et al., 2023), the Getty v. Stability AI UK ruling of November 2025, and the practical side: the 5-part prompt formula and the C2PA provenance check.

How AI Image Generation Works (Series 3, Episode 2). One fake AI picture of the Pentagon moved the American stock market by 0.3% in four minutes. In the second episode of our series "How the AI Tools You Use Actually Work", we explain the true mechanics: diffusion starts from television static and removes noise step by step; CLIP linked words to pictures with 400 million pairs; Stable Diffusion reached home computers on 6.9 GB of memory. We also face the dark side: models that memorise training photos, and the Getty lawsuit the UK High Court decided in November 2025. Our verdict is YELLOW / WATCHING: a real creative tool, but a photo is no longer proof of anything. Inside: the 5-part prompt formula and the 5-point checklist for spotting a fake. Subscribe free to Paragon AI MAG — one article, one video and one podcast episode every day.

By Paragon AI MAG — Series 3, Episode 2: "How the AI Tools You Use Actually Work"

One fake photo moved the market

On the morning of 22 May 2023, a photograph spread across Twitter showing a huge black cloud of smoke beside the Pentagon. An account with a verified blue check mark, impersonating Bloomberg News, posted it. Major news outlets picked it up. Panicked investors triggered an algorithmic sell-off, and the S&P 500 fell 0.3% in just four minutes. Billions of dollars in market value vanished — and then the authorities confirmed it: no explosion had ever happened. The picture was completely fake.
That is the question this episode answers: how does a machine draw a photo of something that never happened, well enough to move the stock market?

The short answer: it starts from static

The AI does not draw. It does not copy and paste pieces from a hidden database of photos either. It starts from pure random noise — television static on an old screen — and through a process called iterative denoising, a neural network removes a tiny bit of noise, step by step. It usually takes about 20 to 50 passes to clean it up. It is, in effect, sculpting an image directly out of the static until a coherent, hyper-realistic picture emerges from absolute chaos.
Paragon verdict: YELLOW / WATCHING. Image generation is a real, revolutionary creative tool you can use right now. But a photo is no longer proof of anything. The trust layer around where an image comes from, spotting fakes, and copyright is completely unfinished.

How it learned: 400 million pairs

To understand how the AI knows what to sculpt out of all that static, look at how it learned. These models absorbed visual concepts by analysing massive datasets. Take OpenAI's CLIP model from 2021: it linked words to pictures using 400 million image-and-caption pairs. It did not memorise rigid rules. It learned statistical patterns — it worked out what the pixels of a cat or a sunset tend to look like across millions and millions of examples.

Stable Diffusion reaches everyone

What is staggering is how fast this computing power reached the public. In August 2022, Stable Diffusion launched completely for free, and it was so well optimised that it needed only 6.9 gigabytes of graphics memory to run locally. Practically anyone could fire it up on a standard home computer. Almost overnight, everyday consumer laptops became image factories.

The scale: 80 million images a day

From isolated tech demos, the technology reached global saturation in a couple of years. Tracker estimates put the daily volume at roughly 80 million synthetic images generated globally every single day — about 926 every second. (That figure is an estimate from industry trackers, extrapolated from Everypixel's 2023 measurement of 34 million a day.) It is a relentless visual assembly line operating at a scale humans cannot match, and it is blurring the line between human expression and machine output.

The dark side: memorisation

This massive scale comes with a serious complication: memorisation. If the model does not actually store photos, how did researchers manage to pull nearly exact pixel-for-pixel copies out of it? A 2023 study by Google and DeepMind exposed the flaw: if an image is heavily overrepresented in the training data — like a famous painting or a super common stock photo — the statistical model collapses and occasionally just regurgitates exact copies instead of generating something new.
The legal fallout was fast. Artists and photo agencies filed class-action lawsuits almost immediately, but the legal system is struggling to catch up with the math. Notably, in November 2025, Getty Images lost its main UK copyright claim against Stability AI. The High Court ruled that because model weights are massive files of statistical probabilities, not hard drives storing actual visual copies, they are not considered infringing material. The law is having a hard time defining what this technology even is.

The tools today

Despite those legal grey areas, commercial tools are advancing fast. Google's Nano Banana 2.1 runs at about 4 cents an image and is built for fast photo editing. Midjourney starts at around $10 a month and is known for beautiful, highly artistic aesthetics. OpenAI's ChatGPT Images 2.6 costs $20 a month and uses a massive language model backbone to give incredibly literal text adherence.

The 5-part prompt formula

The AI fills in any detail you leave out with a statistical average of everything it has ever seen. If you don't take control, the machine just guesses. The key is the five-part prompt formula: first, define your subject clearly. Second, declare the style. Third, specify the lighting. Fourth, set the mood. And fifth, add technical specs.
Don't just type "a cat". Type "a ginger cat sleeping on a windowsill, photorealistic style, golden-hour sunlight, peaceful mood, shot on a 35-millimetre lens". Anchor those five pillars and you lock the statistics exactly where you want them.

Spot the fakes: the 5-point checklist

When your feed is flooded with hundreds of new synthetic images every second, how do you spot the fakes? First, look for distorted hands or extra fingers — models still struggle with human anatomy, sometimes giving people seven fingers on one hand. Second, check for gibberish text on signs in the background. Third, look for impossibly smooth, plastic-looking, perfect skin. Fourth, scan for structural glitches in the background, like melting fences or shadows pointing the wrong way. And fifth, whenever possible, check the file for C2PA metadata credentials — hidden cryptographic tags that can verify a photo's true origin.

What comes next

The exact same diffusion mechanism — starting from noise and meticulously cleaning it up — is now being massively scaled. The next frontier is real-time video generation, where environments render around you instantly, fast 3D model creation, and full-fidelity audio and music generation. We have crossed a threshold where seeing is definitively no longer believing.
So use this as a powerful creative tool to build and imagine. But knowing what you know now — will you ever truly trust a photo on its own again?
Paragon AI MAG — one video and one podcast episode every day. Sources: NPR (22 May 2023); Associated Press; OpenAI CLIP (400 million image-text pairs, February 2021); Stable Diffusion public release (22 August 2022, 6.9 GB VRAM); Carlini and colleagues, Google/DeepMind (2023 memorisation study); Getty Images v. Stability AI (UK High Court judgment, 4 November 2025); Google Nano Banana (Gemini 2.5 Flash Image, August 2025; Nano Banana 2.1, October 2026); OpenAI Images 2.6 (4 November 2026). The scale figures (about 80 million images a day, 926 a second) are tracker estimates (AnythingCounter, extrapolated from Everypixel's 2023 measurement).

Watch the video version on YouTube: https://www.youtube.com/watch?v=b1VJCRzkslo
Read the full article on Substack: https://paragonaimag.substack.com/p/how-ai-image-generation-works