Iteration loops: change with intent

1. Define success before generating.
List three to five criteria in priority order. For a course thumbnail: recognizable subject at small size, open corner for a badge, palette aligned with the course, no generated text, and no unsafe equipment.

2. Generate a small batch.
Several candidates reveal the model’s interpretation better than one. Keep prompt and settings constant within the batch. You are sampling variation, not testing a revision yet.

3. Diagnose specific gaps.
Replace reactions such as “boring” with observations: the subject occupies only 15 percent of the frame; background contrast competes with the face; the reserved corner contains a lamp; the illustration uses glossy 3D rendering instead of flat shapes.

4. Revise one variable family.
Change composition while holding style and subject steady, or change palette while holding composition steady. Generate another batch and compare it against the same criteria.

This cycle can end in selection, targeted editing, compositing, or a decision that generation is the wrong method.

A minimal log turns trial and error into reusable knowledge. Record:

• Prompt version and date.
• Tool and model version if visible.
• Aspect ratio and relevant settings.
• Reference inputs and their permissions.
• Seed or variation identifier when supported.
• Intended hypothesis: “A medium close-up will make the hands readable.”
• Outcome: what improved, regressed, or remained unstable.
• Decision: keep, branch, revise, or reject.

Do not assume a seed reproduces an image forever. Providers can update models, safety systems, or preprocessing. Save approved outputs and prompts according to your project’s retention rules. For important work, preserve provenance and final editing history too.

When a candidate has a strong composition but the wrong palette, branch from it. Do not throw away the successful frame. Image-to-image editing, regional editing, masks, or variation controls can preserve more structure than generating from scratch.

Think in branches:

• Branch A: preserve composition; explore warm and cool palettes.
• Branch B: preserve style; test closer and wider framing.
• Branch C: preserve subject pose; simplify background.

Name versions by meaningful changes rather than “final-final-3.” A filename such as v06-close-shot-muted-palette helps collaborators understand lineage.

Use masking for local problems when the tool supports it. If only a mug is misshapen, regenerate that region rather than risking the entire scene. Check boundaries, shadows, reflections, and texture after a local edit; repairs can create new inconsistencies.

Different failures call for different interventions:

• Wrong concept: clarify subject and action.
• Weak layout: revise framing, hierarchy, or viewpoint.
• Wrong feel: change palette, medium, lighting, or texture.
• One defective region: inpaint or composite.
• Need more canvas: outpaint, then inspect added content.
• Unstable identity or object: use an authorized reference or a consistency workflow.
• Exact geometry or text: create those elements with deterministic design tools.

Prompt length is not the solution to every failure. If the model cannot reliably draw an exact interface, diagram, logo, or sentence, use generation for supporting visuals and build precise elements separately.

Selection by first impression tends to reward polished images that may miss the brief. Score candidates on the declared criteria. A simple 0–2 scale works:

• 0: fails or contradicts the criterion.
• 1: partially satisfies it or needs repair.
• 2: satisfies it clearly.

Keep hard constraints as gates rather than averages. An image that scores beautifully on mood but exposes private information should be rejected, not rescued by its total score.

Invite the right reviewers. A designer can assess hierarchy; a subject-matter expert can catch incorrect equipment; a community representative may identify harmful portrayal; a legal or brand reviewer can evaluate higher-risk uses. “Looks good to me” is not enough when the image makes factual or reputational claims.

Iteration can consume unlimited time and credits. Stop when:

• One candidate meets every hard requirement and enough preferences.
• Remaining defects are cheaper and safer to edit manually.
• New rounds no longer improve the rubric.
• The model repeatedly fails an exact requirement.
• Risk or uncertainty makes another production method preferable.

Record why you stopped. A good workflow includes declining an output method when it does not fit the job.