Image prompting: build a brief you can actually revise

Describe the job the image must do

Start with the intended use, not a list of aesthetic adjectives. A website header needs space for type and must survive a narrow crop. A product detail image needs accurate visible features. A concept illustration can tolerate more interpretation. Write those requirements before describing the scene.

Choose the subject and the main relationship in the frame. Specify what should dominate, where the viewer should look and which area needs to stay clear. If everything is equally important, the generator and the reviewer have no common basis for deciding whether the result works.

Build the prompt in layers

Use a simple order: subject, setting, composition, light and treatment. Example: 'A small silver desk lamp on an uncluttered dark table, placed in the right third of a wide frame. Soft light from the left, subtle shadows, restrained editorial photograph. Leave the left half clear for a headline.' This is a starting brief you can adapt.

Keep exact product features in a separate acceptance checklist. A prompt can request a feature without reliably preserving it. If the image must represent a real object, compare the output against approved reference material before using it in public-facing work.

Use references with a defined purpose

Decide whether a reference supplies the subject, the composition or the visual treatment. Those purposes can conflict. A photograph with the right palette may contain a layout you do not want. Explain which qualities should carry across rather than asking for a vague resemblance.

Midjourney and BFL both document reference-related controls, but their interfaces and model capabilities differ. Read the instructions for the selected tool instead of copying parameters between services. Keep the reference file and the prompt together so your next revision starts from the same information.

Review systematically

Inspect the image at thumbnail size to judge composition, then at delivery size to inspect details. Check edges, repeated objects, text and unusual anatomy where relevant. Review a crop preview for every destination. A technically large image can still have the wrong visual hierarchy for a small screen.

Change one layer of the prompt per round. If the composition works but the lighting does not, keep the composition instructions. Record a small set of accepted and rejected examples with reasons. That record becomes a practical style guide for the project.

Finish exact elements outside the generator

Set final headlines, legal text and logos in a design tool when they must be exact. This makes typo fixes and localization straightforward. Preserve an editable master rather than baking every required element into a single generated raster image.

Before exporting, verify the final crop, contrast and file size. Write alt text that describes the image's role in its page. Save the successful prompt, references and editing steps. A useful prompt is part of a reproducible production process, not a secret phrase you hope to remember.

Questions & answers

Sources & further reading

Product details should be checked against these primary sources. Editorial recommendations are our interpretation; they are not hands-on benchmark results.

• Midjourney: Style Reference ↗
• BFL: FLUX.2 documentation ↗