TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 18—20 · FACTUAL · DURABLE
Upscaling and the resolution question
Higher resolution raises confidence without adding any truth.
The idea
Models make images at a working resolution, then add detail in upscaling passes. An upscaler cannot know what was lost at the small size. It infers detail consistent with the small image, and where several answers fit, it picks a plausible one: grain in the stone, joints in the brick, letters on a signboard you never asked for. So a 4K render can look sharper and more certain than the model ever was. Some of the sharpness is honest inference. Some is new invention over the old invention, at a finer grain, and nothing in the image tells you which.
Why it matters
Detail reads as evidence. A contractor, a client or a juror will trust a crisp joint line more than a soft one, though neither was designed. Higher resolution raises confidence without adding any truth.
See it in the studio
The upscaled street view shows a shop sign in something like Kannada, a tile pattern on the plinth, a meter box by the door. Zoom in. The letters are not letters, the tiles change size halfway up the wall, and the meter box floats. Each of these was invented in the pass that made the image "print-ready".
Watch for this
Sending an upscaled image for material approval. The texture the client approves does not exist as a specification.
Try it
Take one generated image. Upscale it. Put the two at the same screen size and zoom into the same 200 mm of wall on each. List what the upscaler added. Then ask: would anyone downstream read that as information?
Prove it
Explain why an upscaled render can gain a lot of detail that means nothing, and show, on one blown-up image, where inference consistent with the source ends and invented detail begins.
How it works
Modern upscalers are themselves generative networks, trained on pairs of degraded and clean images to infer a likely clean image from a degraded one. Where the small image genuinely constrains the answer, the inference is faithful; where it does not — the fine texture, the lettering — the network supplies a plausible answer, not the true one. Diffusion-based upscaling runs another denoising pass at the larger size, with all the freedom that brings.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
- diffusion-how-ai-paints-from-noise
- ch03
- practical-applications
- Wang et al., Real-ESRGAN
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.