TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · DURABLE
Text-to-image vs image-to-image
Choosing between the two modes is really choosing how much of your source to keep.
The idea
In a diffusion or flow pipeline, text-to-image starts the denoising loop from pure noise, so layout, massing and viewpoint are invented fresh every run. Image-to-image starts from a noised-up version of a picture you supply. The less noise you add, the more of your picture's structure survives. The more noise, the freer the model gets. The setting that controls this is usually called denoising strength. Choosing between the two modes is really choosing how much of your source to keep. Keep the massing and change the mood, or keep almost nothing and take a fresh look.
Why it matters
Most "it ignored my sketch" complaints come from a strength setting that is too high. Most "it just traced my sketch" complaints come from the same setting, too low.
See it in the studio
You have a massing screenshot and want it at dusk, in brick, with people. Text-to-image gives you a different building. Image-to-image at moderate strength keeps your masses and repaints the rest.
Watch for this
Using image-to-image on a weak drawing and expecting it to be redesigned. It keeps what you gave it. A wrong plan, lit beautifully, is still a wrong plan.
Try it
Take one sketch. Run image-to-image at low, middle and high strength. Lay the three beside the sketch. Mark, on each, what survived and what was invented.
Prove it
Explain why image-to-image keeps the layout that text-to-image makes up fresh each time, and describe what you give up as you turn denoising strength up.
How it works
Image-to-image adds noise to your picture part of the way along the same path training used, then denoises from there. The idea was set out in a method called SDEdit. It is the simplest form of grounding. CONTROL INPUT on this strand is the stronger form, where edges or depth are fed alongside the prompt at every step rather than only as a starting point. Other generator families — autoregressive image models, for one — have no denoising loop and do "start from my picture" differently; this card describes the diffusion and flow family.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.