TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · DURABLE
Latent space, the field of possibilities
Diffusion runs inside a compressed grid, and pixels are decoded only at the end.
The idea
Many image generators do not work in pixels. An autoencoder first compresses an image into a much smaller grid of numbers — its latent — and can expand a latent back into a picture. The diffusion process runs on that small grid: it starts from noise there, is denoised there, and the result is decoded to pixels at the end. This is the third number-space on this strand. Its job is compression, so that generation is fast enough to use. It is not a map you walk about in, and a prompt is not a pin. The prompt steers each denoising step.
Why it matters
When every result looks the same, the model's learned habits are steering every run the same way. Change what steers it — a different anchor, a reference, a control input — rather than the number of words.
See it in the studio
Ten runs of "a contemporary house in Goa" all give white render, laterite accent, pool, palms. That is a well-worn habit of the model. Change one anchor — "monsoon verandah, no pool, pitched Mangalore tile" — and the steering changes.
Watch for this
Writing a longer prompt when you want a different result. Length does not change the steering. A different anchor does.
Try it
Run one prompt eight times and lay the results out. Circle what every image shares. Then change one word and run eight more. What moved, and what stayed, shows you what your prompt decided and what the model's habits decided.
Prove it
Explain what a latent is, why a diffusion model works in one instead of in pixels, and why two prompts close in wording give similar-looking images — without using a map picture.
How it works
Latent diffusion (Rombach et al.) compresses images with an autoencoder and runs the denoising inside that compressed space, then decodes back to pixels. Compression is what makes generation fast enough to use. The space is learned, so it is uneven (DENSE, SPARSE on this strand), and it has no measuring tape in it (NO RULER on Foundations). Some newer generators use other latents, or none; the lesson holds for all of them — generation is a learned transformation of noise, not a walk across a map.
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.