How does AI make a picture from a sentence?

The full story

The episode's narration, word for word. Numbers link to the sources below; timestamps play that moment.

The question 0:00

Last time, we saw a chatbot make things up by accident. Image generators make things up on purpose. You type a sentence, and a little later there's a photo of a place that never existed ¹. The strangest part is where many of them start: not with a sketch, but with a square of pure random static ^2,1. So how does AI make a picture from a sentence?

Six million numbers 0:26

Back in episode five, we saw that a photo is just a grid of numbers. Since Season 4, this channel's stills have been made at nineteen hundred and twenty by ten eighty-eight pixels, with three colour numbers each ³. That's over six million numbers to choose ³. Learning to get every one right in a single leap is very hard. The trick that works breaks the job into many small steps, and it begins by doing the opposite: wrecking a picture ^4,2.

Wrecking is easy 0:58

Here's a tiny picture of a tree, sixteen pixels square, drawn for this episode. Now mix in random noise, a little at a time. A quarter noise, then half, then three-quarters, until nothing of the tree is left ^5,6. Each step is just arithmetic: some of the picture, plus some random numbers ^2,5. No intelligence needed. Destroying a picture is easy. The question is whether a machine can learn to run it backwards.

An idea from physics 1:29

In 2015, Jascha Sohl-Dickstein and colleagues at Stanford and Berkeley borrowed that idea from physics ⁴. Think of ink spreading through water. Their plan was to slowly destroy the structure in pictures through a process like that spreading, called diffusion, and then train a network to reverse it ⁴. The analogy breaks in one important place. Real ink never un-mixes. The network doesn't reverse physics; it learns what a slightly less noisy picture usually looks like ^4,2.

A thousand small steps 2:06

In 2020, Jonathan Ho, Ajay Jain and Pieter Abbeel at Berkeley showed diffusion making convincing pictures ². They used a thousand steps, each adding a little noise ².

Guess the noise 2:22

Here's what the network actually learns ². Take a training picture, mix in a random amount of noise, and ask the network one question: what noise was added ²? Say one pixel was nought point eight, the noise was minus one point two, and we mixed them half and half, giving minus nought point two. If the network guesses minus nought point nine, it's out by nought point three, a squared error of nought point nought nine. Then gradient descent, from episode four, nudges its weights to do better ². Repeat that across an enormous number of pictures, pixels and noise levels ².

Running it backwards 3:04

Why is guessing the noise enough? If you know the noise, you can subtract it and work out the picture ². So to make a new picture, start with pure static, guess the noise, take some of it away, and repeat ². Here's that on our toy, with eight little training pictures and four steps ⁶. At first, every picture is equally likely, so the best guess is a blur of all eight at once. One step later, it's torn between a tree and a house. By halfway, it's ninety-seven per cent sure: a tree. Large shapes first, details last, just as Ho's team saw ².

Where the toy cheats 3:45

Our toy cheats. It uses the perfect denoiser for just eight pictures, so all it can ever do is give one of them back ⁷. Real networks are trained on far more. In a 2024 study, a network trained on one or ten face photos memorised them, but trained on a hundred thousand, it made new faces ⁷. Big models still copy now and then: in 2023, researchers got Stable Diffusion and Google's Imagen to give back more than a hundred near-copies of training pictures, mostly ones repeated many times in the training data ⁸.

Words and pictures meet 4:23

So far there's no sentence at all. In January 2021, OpenAI described CLIP, trained on four hundred million pictures with their captions, collected from the internet ^9,10. Its whole job was to predict which caption goes with which picture ⁹. It turns each picture, and each piece of text, into a list of numbers, so the two can be compared ⁹. Matching pairs score high, like nought point nine seven here, and mismatches score low.

The sentence steers 4:55

Now give those numbers to the denoiser at every step ^11,6. Its noise guess depends on the words, so the words steer where the static goes ^11,12. Here's our toy again, from exactly the same static. With no words, it lands on a tree. Add "sun", "moon" or "house", and the same noise becomes a sun, a moon or a house. In 2022, Google's Imagen team found that a bigger language model reading the prompt mattered more than a bigger picture network ¹³.

Turning the words up 5:29

Many models add one more trick, called guidance ¹⁴. At each step, make two guesses, one with the prompt and one without, then push further in the direction the words point ¹⁴. With made-up numbers: without words, nought point two; with words, nought point five; pushed three times further, one point four ¹⁴. Turn it up, and pictures follow the prompt more closely, though too much makes them look worse ^14,13.

Working small 5:58

Six million numbers per step is slow, so in late 2021, Robin Rombach and colleagues in Munich and Heidelberg compressed the picture first, then ran diffusion on the small version ¹¹. That became Stable Diffusion ¹⁵. Three of those authors later co-founded Black Forest Labs, which made the model behind this channel's stills ¹⁶. For our pictures, it squeezes six million numbers into about half a million, twelve times fewer, in eight thousand one hundred and sixty patches ^6,3.

Our own pictures 6:33

So here's this channel's own kit since Season 4, read straight from the files ³. The model is FLUX.1 schnell, released by Black Forest Labs in August 2024, with twelve billion parameters ^1,17. Its licence allows commercial use, which is why we use it, not a recommendation ^1,3. It runs on a laptop in a squeezed-down copy, using just four steps, where its sister model FLUX.1 dev uses fifty and Ho's 2020 model used a thousand, because schnell was trained to take big strides ^2,18,1,6,3. It skips guidance altogether ^6,3. Across a hundred and eighty-three of our stills, all made in October 2026, making a typical one took about two and three-quarter minutes ³.

The prompt that asked for no heads 7:21

For our season on movie physics, we asked for a cluttered nineteen-eighties special-effects bench: latex sheets, tiny motors, cables, sculpting tools ³. And to be safe, we added: no faces, no masks, no heads ³. Here's what came back. A pale sculpted head, a skull-like face, and little heads along the bench ³.

Same noise, three sentences 7:47

So we added more: no skulls, no skeletons, no dolls, no figures, and we swapped latex for rubber ³. It came back with a whole row of skulls ³. Every version started from the very same static, the same seed, so only the words changed ³. On the third try, we deleted every phrase starting with no, and swapped sculpting tools for screwdrivers and pliers ³. Clean bench, no skulls. Prompt words pull the picture towards what they name, and a 2025 study found that models like the ones that read prompts often handle "not" no better than a random guess ¹⁹. This model, as we run it, has no separate way to say "avoid this" ^6,3.

Words on things 8:35

Writing went wrong too ¹². We asked for a children's book with a blank cover and no text, and got a made-up title, something like "The Bortims of the Uisyer" ³. One reason is the tokens from episode six. OpenAI's 2022 DALL-E 2 paper suggested that its tokens, which hide the spelling of words from the model, were likely part of the problem ¹². In 2022, a Google team showed that letting the model see individual letters helped, beating rival models by more than thirty points on rare words ²⁰. Newer models spell much better: Stability AI reported better lettering in March 2024, and OpenAI says its March 2025 generator reliably puts text into pictures ^5,21.

Why hands went wrong 9:25

And then hands ²². In 2023, researchers at the University of Sydney catalogued extra fingers and strange shapes in hands made by Stable Diffusion and Midjourney ²². They point to two reasons ²². A hand has sixteen joints and twenty-seven ways to move, so there's very little room for error ²². And a hand is three-dimensional: in a flat photo, fingers hide each other, so the training pictures show different numbers of fingers ²².

Not the only way 9:57

As of October 2026, there are many of these tools: Midjourney, Stable Diffusion, FLUX, Google's Gemini image models, Adobe Firefly, and the image maker inside ChatGPT ^{1,15,21,23,24,25}. Not all of them say how they work ^23,24,25. And not all are diffusion: OpenAI says the generator it added to ChatGPT in March 2025 is autoregressive, building a picture piece by piece, more like text ^21,26.

Where the pictures came from 10:28

One fair line on where all this learning came from. Many of these models learned from billions of captioned pictures collected from the web ^27,15. One public list from 2022, called LAION-5B, links to five point eight five billion ^27,15. Whether that's fair, or legal, is disputed ^28,29,30. Artists and photo libraries have gone to court, the companies are contesting it, and a UK ruling in November 2025 settled only a narrow part ^28,29,30. It's a story for next season.

What a sentence does 11:03

So that's a picture from a sentence: start with static, guess the noise, take a little away, and let the words steer every step ^2,11,14. The model doesn't paint, and it doesn't copy and paste, though it can memorise ^7,8. And it struggles with the word no ¹⁹. Everything so far, chatbots included, waits for us to type something. So what happens when a model can use tools and take steps by itself? What is an AI agent, and what can it really do on its own?

Sources

Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.

1. Black Forest Labs, FLUX.1 [schnell] model card — huggingface.co
2. Ho, Jain and Abbeel, Denoising Diffusion Probabilistic Models (arXiv 2006.11239, 2020) — arxiv.org
3. Aussie Science Clips production files: pipeline/localstills.py and the mflux metadata… — youtube.com
4. Sohl-Dickstein, Weiss, Maheswaranathan and Ganguli, Deep Unsupervised Learning using… — arxiv.org
5. Esser et al., Scaling Rectified Flow Transformers for High-Resolution Image Synthesis… — arxiv.org
6. Black Forest Labs, flux reference code: src/flux/sampling.py, util.py, cli.py (main,… — github.com
7. Kadkhodaie, Guth, Simoncelli and Mallat, Generalization in diffusion models arises… — arxiv.org
8. Carlini et al., Extracting Training Data from Diffusion Models (arXiv 2301.13188, 2023) — arxiv.org
9. Radford et al., Learning Transferable Visual Models From Natural Language Supervision… — arxiv.org
10. OpenAI, CLIP: Connecting Text and Images (blog, 5 January 2021; read via the Wayback… — web.archive.org
11. Rombach, Blattmann, Lorenz, Esser and Ommer, High-Resolution Image Synthesis with… — arxiv.org
12. Ramesh et al., Hierarchical Text-Conditional Image Generation with CLIP Latents… — arxiv.org
13. Saharia et al., Photorealistic Text-to-Image Diffusion Models with Deep Language… — arxiv.org
14. Ho and Salimans, Classifier-Free Diffusion Guidance (arXiv 2207.12598, 2022) — arxiv.org
15. CompVis, Stable Diffusion repository README — github.com
16. Wikipedia, Flux (text-to-image model) (read 2026-10-15) — en.wikipedia.org
17. Black Forest Labs, Announcing Black Forest Labs (1 August 2024) — bfl.ai
18. Sauer et al., Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion… — arxiv.org
19. Alhamoud et al., Vision-Language Models Do Not Understand Negation (CVPR 2025) — arxiv.org
20. Liu et al., Character-Aware Models Improve Visual Text Rendering (arXiv 2212.10562, 2022) — arxiv.org
21. OpenAI, Addendum to GPT-4o System Card: 4o image generation (25 March 2025) — openai.com
22. Lu, Xu and Zhang, HandRefiner: Refining Malformed Hands in Generated Images by… — arxiv.org
23. Midjourney — midjourney.com
24. Adobe Firefly — adobe.com
25. Google DeepMind, Gemini Image – Nano Banana (deepmind.google/models/imagen/ now… — deepmind.google
26. OpenAI, Addendum to GPT-4o System Card: Native image generation (PDF, 25 March 2025),… — cdn.openai.com
27. Schuhmann et al., LAION-5B: An open large-scale dataset for training next generation… — arxiv.org
28. Cleary Gottlieb, UK High Court issues landmark ruling in Getty Images v Stability AI… — clearygottlieb.com
29. CourtListener, Andersen v. Stability AI Ltd., 3:23-cv-00201 (N.D. Cal., filed 13… — courtlistener.com
30. CourtListener, Getty Images (US), Inc. v. Stability AI, Ltd., 3:25-cv-06891 (N.D.… — courtlistener.com

Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.