Define the image's job
Our invented brief is a hero image for a ceramic travel cup. The cup must remain recognizable, the background should leave usable space for a heading, and no text should appear inside the generated image. This is an editorial exercise, not a completed advertising campaign. An attractive image that crops the handle or fills the heading area may fail the brief. Write the acceptance criteria before adding cinematic language. That makes the review about the asset's intended use instead of a vague preference for something more impressive.
Establish a simple baseline
Start with one cup, a matte ceramic surface, a warm neutral tabletop and soft side lighting. Specify where the cup belongs in the frame and where the empty area should be. Some generators accept an aspect ratio as a parameter rather than through prose; use the product's documented control when available. Avoid stacking contradictory instructions such as centered composition and far-left placement. The baseline is a prompt you can interpret, not a guarantee that any model will obey it exactly.
Change composition before adding effects
Create variant B by moving the cup to the left third and reserving the right half for a heading. Keep subject, materials and lighting unchanged. Compare the generated images for cup completeness and usable empty area. Then create variant C by changing only the lighting direction. This order helps distinguish the effect of a compositional instruction from the effect of a dramatic style term. If you change camera angle, background, material, color and lighting simultaneously, a better result will not tell you which change helped.
Use the authored comparison worksheet
The example below contains prompt variants and a blank review plan. It does not contain measured scores or claimed generation results. Record whether the cup is complete, whether there is enough heading space, whether unintended text appears, and whether the material resembles matte ceramic. Use yes, no or needs review for each criterion. Preserve the model version and available generation settings. Repeat a comparison when randomness makes a single pair misleading, and do not infer universal performance from one visually pleasing output.
Diagnose the failure you can observe
If text appears, simplify the scene and clarify that any label area is blank. If the cup is distorted, reduce competing scene elements and review whether the desired camera angle hides important geometry. If the image looks glossy, distinguish material from lighting: a matte surface can still show a bright highlight under some conditions. These are hypotheses to test, not promises that one phrase fixes the problem. Generated product imagery may misrepresent a real product, so compare proportions, colors and features against an approved reference before publication.
Keep rights and delivery separate from style
Check that you have rights to uploaded reference images and that the generator's terms permit your intended use. Do not use a person's likeness or a trademarked asset without appropriate permission. After selecting a candidate, inspect at final delivery size, verify the crop and add website alt text that describes the visible asset. Text overlaid by your own layout should stay readable on mobile. Style commands are starting points for iteration; the value of this exercise is the controlled review workflow, not claiming that borrowed adjectives are secret technology.
Worked example
BASELINE
One matte ceramic travel cup on a warm neutral tabletop. Soft side light. Complete cup visible. Blank surfaces; no text.
VARIANT B — ONLY COMPOSITION CHANGES
Same subject, surface and lighting. Cup in the left third. Reserve the right half as uncluttered space for a heading.
REVIEW WORKSHEET
Model/version: ___ Settings: ___
Complete cup: ___ Heading space: ___
Unwanted text: ___ Material match: ___
Observed difference: ___ Next single change: ___How to check the result
Record observations from your actual images. Leave the worksheet blank until a run occurs. A visual preview on this site illustrates direction and is not evidence of a prompt benchmark.