TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 18—20 · FACTUAL · DURABLE

Guidance, how hard it steers

Run three guidance values on one seed and pick the zone between blur and burn.

When to use

Whenever a diffusion or flow tool exposes a guidance setting (often called "CFG", for classifier-free guidance) and the result is either a soft blur that ignores your prompt or a stiff, over-saturated image that looks shouted. If the tool hides the setting, this card does not apply to it.

The method

Keep the seed and the prompt fixed. Run a low guidance value, a middle one and a high one. The tool's documentation gives the usual range. Lay the three out. Low gives a soft, loose image that drifts from the words. High gives hard edges, burnt colour and repeated details. Pick the zone between, where the image follows the prompt and still looks natural. Note the value. Only then change the prompt or the seed. Guidance controls how closely the image follows the prompt. The seed controls the starting point. Change one at a time.

Watch for this

Rewriting the prompt when the fault is guidance. A burnt image is not over-described. It is over-steered. A vague image is not always under-described. It may be under-steered.

Try it

Three images, one seed, one prompt, three guidance values. Label each with its value and write one word under it. Keep the sheet. It is your calibration for that tool.

Prove it

Shown an image that is either blurred or burnt, trace the fault to the guidance setting rather than the wording, and say clearly what guidance controls and what the seed controls.

How it works

At each denoising step the model makes two predictions, one with your prompt and one with no prompt. Guidance pushes the result further along the difference between them. More push means the image follows the prompt more closely, until the push overshoots and produces artefacts. This is a mechanism of particular diffusion and flow pipelines, not a law of image generation: some newer, distilled models build guidance in and hide the setting, and "too high and it burns" is a habit of today's pipelines rather than a durable principle. Where the setting is hidden you cannot calibrate it, and this method does not apply.

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.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON