A more useful way to iterate on AI images
Turn a vague idea into a repeatable image workflow, from first draft to a deliverable you can use.

Begin with the image's job
A blog cover, product mockup and social post have different constraints. Before writing a prompt, decide where the image will appear, which aspect ratio it needs and whether text will be added later. If the image must leave room for a headline, describe that composition rather than hoping a perfect empty area appears.
Before the prompt, decide what the image needs to do: what readers should notice first, whether it appears as a banner or a small card, where text needs room, and what mood fits the article. One or two clear sentences make results easier to evaluate than a pile of style adjectives. Avoid relying on generated tiny text when spelling must be exact.
Change one decision at a time
Describe subject, setting, lighting and framing in ordinary language. Generate a small first set, pick the strongest direction, and adjust one or two elements: perhaps the crop and the light, not the entire scene. Keep notes on the model, prompt and settings so a good result is easier to reproduce or adapt.
Generate a few candidates with the same brief and choose the strongest subject, composition and color direction. Change only one or two factors per iteration, and save the prompt, model and important settings. If a result drifts, identify whether the subject, light or crop caused the problem. Random bulk retries can add cost without explaining what worked.
Composition, light and color can be refined one decision at a time.Review at the size people will see
A dramatic image can lose its subject when cropped into a small card. Check the result on mobile and desktop, look at edges and hands, and inspect whether the visual communicates the article rather than merely decorating it. Add meaningful alt text when the image carries information; decorative covers can use an empty alt attribute.
Place candidates in the actual page and check mobile cards, desktop lists and social-share crops. Look for obscured subjects, sufficient text contrast and obvious artifacts near the edges. Informative images need accurate alternative text; decorative images need not repeat the article title to a screen reader. Check the rights and permitted use of portraits, brands and third-party references.
Handle the finished asset deliberately
Use a suitable compressed format, preserve an original, and check that the uploaded image loads from your own approved storage. Confirm the rights and permitted use of any reference or input image. Models, sizes and output formats vary, so check the selected model's documentation and current price before making a large batch.
Keep the original and an appropriately compressed version, choosing dimensions and format for each use. After upload to your approved storage, verify that the public asset loads and provide a fallback for failure. Track which file belongs to which article so a replacement does not break links. Balance visual quality with loading performance rather than minimizing bytes at any cost.
Save the creative brief, final prompt, selected file and necessary usage permissions. A teammate can then distinguish firm requirements from an experiment that happened to work once. If the article title or card ratio changes, first check whether the existing image still works before generating another version without a clear purpose.
Start with the purpose of the image
Decide where the image will appear before writing the prompt: a wide blog cover, a vertical poster or a small product card. The placement affects aspect ratio, subject size and negative space. If you will add text in a design tool later, leave a clean area rather than relying on a model to render every letter correctly.
Describe the subject, action, setting and camera viewpoint first. A close-up cup on a desk and a cup held on a busy street ask for different images. Use a few clear requirements instead of a stack of contradictory style words. Before supplying reference images, check both the model's supported inputs and your rights to use the material.
Change one important decision at a time
After the first result, identify the main problem: subject, composition, lighting or style. Change one or two decisions while keeping the rest stable, so you can tell what helped. Save the prompt, model and important settings with each candidate.
Place candidates in the actual page as well as viewing them full size. A mobile crop can hide the subject; text overlays may lack contrast; small details can disappear in a thumbnail. Check hands, edges and repeated objects. Use human review and post-production for exact text, logos or product details, and keep the source, export and usage conditions together.
Refine a specific scene
For a remote-work article, try a clear starting brief: “An open laptop, coffee and notebook on a wooden desk in soft morning light; wide composition with quiet space on the right.” It describes the subject, setting, light and purpose without assuming that every model interprets the details identically.
If the desk is crowded, remove objects first. If the image is cropped on a narrow screen, move the subject toward the safe area. Add concise alternative text for informative images, and check rights and privacy before publishing. Record the model and date if your team may need to produce related assets later.
A prompt you can refine
For a hypothetical article cover, start with a subject and a use: “A small camera on a filmmaker’s desk, soft window light, horizontal composition, room for a title on the left.” If the first result feels crowded, keep the subject and change only the framing. Compare the full image with a card-sized crop before deciding whether it works.
Describe the subject, action or relationship, setting, framing, light and constraints. For example, establish “a camera on a desk” before adding window light, a horizontal crop and room for a title. For editing, say what should change and what must remain. Models differ in reference inputs, sizes and formats; check the actual model instead of copying another model’s options blindly.