Key takeaways
- One variable per generation turns every render into a readable experiment.
- Diagnose by section: light, palette, framing — then touch only that section.
- Keep a “control” render and compare against it, never against memory.
- If three single-variable tries fail, re-word the phrase instead of repeating it.
The iteration method
Start from a structured prompt — the generator hands you one with labeled sections, which is exactly why structured output matters here. Generate a baseline. Decide the single biggest miss. Edit only that section. Generate again. You now know what that phrase was doing. Repeat. Five disciplined iterations teach more than fifty scattergun ones.
Section-level editing
Because each section controls a different axis, you can target precisely:
- Image too flat? Touch lighting only — add direction and shadow behavior.
- Wrong mood? Swap the two palette anchors, or the single mood word.
- Subject lost in clutter? Edit composition (“subject on the right third”) and add one negative (“cluttered background”).
- Looks synthetic? Add one texture phrase and one flaw — grain, a blown highlight. See writing prompts for the vocabulary.
Baseline unchanged except Lighting: replace “even ambient light” with “soft key from camera left, gentle falloff into shadow on the far side.”
One edit. Whatever changes in the render is the lighting phrase’s actual effect.
A/B habits that stick
Name your renders by what changed (“v3_warm-palette”) so the history stays legible. Keep the control image open beside each new one. Judge at two distances — squint for grade and framing, zoom for texture — the same comparison discipline used in recreating images. And log surprises: when a phrase does nothing, that model can’t hear it — re-word, don’t repeat. Understanding why runs vary is covered in why results differ.
The three-try rule
Three consecutive single-variable attempts with no movement means the vocabulary is wrong for that model. Translate the idea into different words before touching anything else.
FAQ
How many iterations is normal?
Three to seven for a known look; more when exploring. If you’re past ten, the prompt structure itself needs rework.
Should I change seeds between iterations?
Keep seeds varied early to test robustness; once the prompt converges, run a few seeds of the final version to confirm stability.
Does this work for every model?
The method does; the vocabulary differs per family. That’s why per-model output formats exist in the generator.