One-shot prompting is tempting: describe a beautiful scene, press generate, and hope the first result is usable. That approach can produce a great image, but it is difficult to repeat. The moment you need the same character, a cleaner product label, more space for a headline, or a different aspect ratio, the composition starts to drift.

The solution is to treat image generation like a small production pipeline. Each stage has one job, and every edit has a clear success criterion.

Good prompts describe the image. Reliable workflows also describe what must not change.

A two-pass workflow that scales

Separate exploration from finalization. Mixing them causes you to spend time polishing a direction that may not survive the next edit.

• Flare for exploration: test composition, color language, camera angle, and visual direction quickly.

• Sunburst for finalization: refine the selected direction, preserve important details, and prepare the asset for delivery.

During exploration, make each candidate structurally different. Four nearly identical images do not give you four useful options. Change the camera angle, subject placement, background geometry, or lighting direction—one major decision per candidate.

Step 1: write an image brief before the prompt

A useful prompt starts with an output contract. Before adding style words, answer these questions:

• Placement: Is this a hero image, social thumbnail, product visual, or editorial illustration?

• Subject: What must be present, and what must be recognizable?

• Composition: Where should the subject and negative space sit?

• Camera: What framing, angle, and depth of field are required?

• Lighting and palette: Which colors dominate, and which should be avoided?

• Delivery: What aspect ratio, background, and file format are needed?

Here is a reusable prompt structure:

Create [asset type] for [placement].

Subject: [what must appear]
Composition: [framing, layout, negative space]
Visual direction: [medium, style, era]
Lighting and palette: [specific direction and colors]
Must preserve: [identity, proportions, product details]
Avoid: [artifacts, unwanted objects, forbidden changes]
Output: [aspect ratio, background, intended crop]

“Premium, cinematic, beautiful” is subjective. “One centered bottle, label facing the camera, empty space on the left, no extra objects” is testable.

Step 2: explore structure, not polish

Use the first pass to answer large visual questions. Does the silhouette read clearly? Is the focal point obvious at thumbnail size? Is the empty space actually usable for interface copy?

Choose the candidate with the strongest structure, not necessarily the most detail. Fine texture and dramatic lighting are easy to add later; a weak composition is expensive to repair.

Step 3: lock the invariants

Most failed edits are not ugly. They simply change something that was already correct. Before requesting an edit, list the invariants explicitly:

• preserve the subject’s identity and facial features;

• keep the camera position, crop, and pose unchanged;

• retain product geometry, logo placement, and label text;

• maintain the background layout and shadow direction;

• do not introduce new objects.

Then state the one change you want:

Keep the composition, product proportions, label, and camera angle unchanged. Replace only the gray background with a warm off-white studio backdrop. Preserve the existing shadow shape.

This gives the model a stable boundary. It also gives you a simple review question: did the requested delta improve the image without breaking the invariants?

Step 4: change one class of variables at a time

A one-delta edit changes a single class of variables: composition, lighting, material, environment, typography, or cleanup. If you change all six at once, you cannot tell which instruction helped or hurt.

Save each accepted version as a checkpoint. A simple naming pattern is enough: concept-A-v01, concept-A-lighting-v02, concept-A-bg-v03, and concept-A-final. If an edit drifts, return to the last good checkpoint instead of trying to repair the damaged version.

Step 5: design for the final delivery

The production target should influence the prompt from the start.

• PNG is useful for transparency, UI assets, diagrams, and crisp edges.

• JPEG is efficient for photographic images that do not need transparency.

• WebP is a strong web-delivery option when file size matters.

• Generate close to the final aspect ratio so cropping does not destroy the composition.

• For transparent assets, request a transparent background and choose a format that supports alpha.

Text inside the image deserves its own mini-brief. Put the exact copy on separate lines, specify hierarchy and placement, and forbid any additional words. If typography is business-critical, generate the visual without text and add the final type in a design tool.

A complete example: a developer-tool hero image

Suppose the final asset is a 16:9 hero image with space for a headline on the left.

Create a 16:9 landing-page hero image for an AI developer tool.

Subject: one translucent glass interface panel floating above a dark desk
Composition: panel on the right third; generous clean negative space on the left
Camera: eye-level, medium-wide view, subtle depth of field
Lighting: cool blue rim light with one restrained amber accent
Style: realistic product visualization, precise and minimal
Avoid: people, readable UI text, logos, watermarks, extra screens, visual clutter

Generate several distinct layouts with Flare. After choosing the strongest composition, move to Sunburst and request a controlled lighting edit:

Preserve the camera angle, interface-panel geometry, desk layout, and left-side negative space. Soften the blue rim light and reduce the amber accent. Do not add text, icons, logos, or new objects.

To run the same workflow in the browser, open GPT Image 2.5. Begin with the image brief, use Flare to explore, and move to Sunburst only after the composition is worth preserving.

A 60-second review before export

• Are all required objects present, with no unintended additions?

• Does the subject match the reference and remain consistent?

• Does the crop work at the final aspect ratio?

• Are hands, faces, reflections, edges, and shadows coherent?

• Is text correct, or should it be added outside the image?

• Is the background and file format appropriate for delivery?

Closing thought

The most dependable image workflow is not the one with the longest prompt. It is the one that makes decisions in the right order. Explore while changes are cheap, lock the invariants once a direction works, and finish with small, testable edits. That turns image generation from a guessing game into a repeatable creative system.

Further reading

• Introducing ChatGPT Images

• Images API: Generate an image

• OpenAI Academy: Image generation