How Does AI Generate Images?

The model never draws. When you type a sentence into an image generator, the machine begins with a rectangle of pure random numbers, static, the grey snow an old television showed when the channel went dead. Then, in twenty to fifty passes, it removes what it believes is noise. What survives the subtraction is your picture. Nothing was drawn. Something was uncovered, out of a field that had nothing in it.

That is the literal training objective, and it goes back to a 2015 paper by Sohl-Dickstein, Weiss, Maheswaranathan and Ganguli. The idea came out of physics: slowly destroy the structure in a dataset, then learn the reverse process that puts it back. Ho, Jain and Abbeel, in the 2020 paper that made the approach practical, ran that ladder a thousand rungs deep. Then you train a network on one question, asked a few hundred million times: which part of this is the noise? Generation is that ladder climbed the other way, and Song, Meng and Ermon showed you can do it in twenty to a hundred steps instead of a thousand. Stable Diffusion's default pipeline runs fifty.

Doing that to pixels is punishing, so in December 2021 Rombach, Blattmann, Lorenz, Esser and Ommer stopped denoising pixels. A separate network compresses the image first: at the factor the public release used, a 512 by 512 image collapses to a 64 by 64 grid, one sixty-fourth of the area. That paper became Stable Diffusion, released publicly as version 1.4 on 22 August 2022. Your sentence gets in two ways. CLIP, trained on 400 million image and caption pairs, puts a picture and the words describing it in the same place in a shared mathematical space. Classifier-free guidance runs the denoiser twice at every step, with your prompt and without, then overshoots the difference. In their own words, it is a method to increase sample quality while decreasing sample diversity. That dial is the house style of AI art. It fell out of a sampling trick.

Then the captions. Stable Diffusion was trained on subsets of LAION-5B: 5.85 billion image and text pairs filtered by CLIP, 2.32 billion of them English. And LAION-5B contains no images at all. It is a list of web addresses and the alt text beside each one, typed by whoever posted the picture, back before any of this existed.

What stays open is whether the model holds a composition or only local statistics. Ask for a red cube on top of a blue sphere and you may get a blue cube beside a red sphere. Conwell and Ullman tested fifteen basic relations on DALL-E 2 in 2022 and found raters matched only about 22 percent of images to the prompt, in work still sitting as a preprint. And the catch: Somepalli and colleagues, at CVPR 2023, found 1.88 percent of random Stable Diffusion samples matched a training image closely enough to count as copied content, the copied ones appearing an average of 34 times in the training data against about 12 for an ordinary image. Memorising and generalising are the same machine. The law has mostly declined to answer: in the London judgment of 4 November 2025 Getty had abandoned its central training claim before closing submissions, and as of September 2026 no American court has decided whether training an image generator on copyrighted pictures is fair use.

Sources (arXiv ids): Sohl-Dickstein et al. 1503.03585, ICML 2015; Ho, Jain and Abbeel 2006.11239, NeurIPS 2020; Song, Meng and Ermon 2010.02502; Rombach et al. 2112.10752, CVPR 2022; Stability AI, Stable Diffusion Public Release, 22 August 2022; Radford et al. 2103.00020, ICML 2021; Ho and Salimans 2207.12598; Schuhmann et al. (LAION-5B) 2210.08402, NeurIPS 2022; Conwell and Ullman 2208.00005 (preprint); Liu, Garrette et al. 2212.10562, ACL 2023; Esser et al. 2403.03206, ICML 2024; OpenAI, Introducing 4o Image Generation, 25 March 2025; Somepalli et al. 2212.03860, CVPR 2023; Carlini et al. 2301.13188, USENIX Security 2023; Getty Images v Stability AI, 2025 EWHC 2863 (Ch); Andersen v. Stability AI, No. 3:23-cv-00201-WHO (N.D. Cal.).

Image credits, via Wikimedia Commons: Seine valley fog by Raimond Spekking (CC BY-SA 4.0); darkroom by Kecko (CC BY 2.0); ice crystals by Jojoo64 (CC BY-SA 4.0); Bremerhaven container terminal by Carsten Steger (CC BY-SA 4.0); starling murmuration by Skander zarrad (CC BY-SA 4.0); Roman sea-life mosaic by Tyler Bell (CC BY 2.0); Royal Courts of Justice by The wub (CC BY-SA 4.0); ink in water by Schbence (CC BY-SA 4.0); silicon wafer by Rob Bulmahn (CC BY 2.0); GPU by Joydeep (CC BY-SA); CRT static by Boreal11 (CC0); archive storage by Samuel Zeller (CC0); federal courtroom by Carol M. Highsmith, Library of Congress (public domain). Sand ripples by Todd Vogel and piano keys by Gabriel Barletta are CC0 via Unsplash. Other visuals are AI-generated conceptual illustrations.

Subscribe: https://youtube.com/@howaithinks Playlist: https://www.youtube.com/playlist?list=PLzCZMjbB9LCpQXSeehGk68uZDVMEA9cXF

#AI #AIArt #StableDiffusion #DiffusionModels #MachineLearning