TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · EVOLVING
Images out of noise
The image has no source, only a process.
The idea
A diffusion model starts with a square of pure noise, like an untuned television. Step by step it removes a little noise. At each step the prompt steers what the cleaner image should look like. After tens of steps, a picture stands where the static was. The picture was not pulled from a library, and no photograph was copied in. It formed, guided by patterns the model learned about how pictures tend to look. That is why a generated image can look right everywhere and still be made up in every part.
Why it matters
"Where did it get this image?" has no answer. The image has no source, only a process. Knowing that changes how you can use it and what you can claim about it.
See it in the studio
A render shows a staircase that starts in the courtyard and leads into a wall. Nobody drew that stair. At each denoising step, "stair-like" and "wall-like" were both likely, and the process settled on both. That is not a bug. That is the process, with nobody checking it.
Watch for this
Asking why the model "chose" a detail. It did not choose. Every detail is where a noisy path ended, not a decision.
Try it
Find a tool or demo that shows the intermediate steps (the Diffusion Explainer does this in the browser). Watch one image form. Note the step where you first recognise a building, and the step where the first wrong detail appears.
Prove it
Talk through the denoising steps in plain words using a row of images from noise to finished picture, then explain why every detail can look believable while none of it is actually correct.
How it works
Training adds noise to real images and teaches a network to predict that noise. Generation runs the prediction backwards, starting from pure noise. Text steers through a text encoder — often a CLIP-style one (SHARED SPACE) — and, in many pipelines, a setting called guidance (GUIDANCE on this strand). EVOLVING: diffusion is the dominant engine for images and video today, but some recent image systems generate differently — token by token, as language models do — and the field moves fast. The lesson holds for both: no source, only a process.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.