GPT Image 2.6 Prompt Guide: A Practical Framework for Flare and Sunburst

A useful GPT Image 2.6 prompt states the production goal first, then describes the subject, action, setting, composition, visual treatment and constraints in that order. Use Flare to compare clearly different directions; use Sunburst to refine an accepted image while naming everything that must remain unchanged. More adjectives rarely produce more control. A better prompt makes the decision visible.

OpenAI’s November 5, 2026 announcement introduces Flare as the faster model and Sunburst as the higher-quality model with stronger editing coherence. The announcement also mentions image labeling and SVG export. At verification time, however, OpenAI’s public API documentation still listed GPT Image 2.5 model identifiers rather than 2.6. The framework below follows the official image-prompting principles that are documented today, but it does not invent 2.6 parameter names, prices or limits.

Begin with the job the image must perform

Before describing style, write one sentence that explains what the image is for. “Create a cinematic image” is vague because it does not define the viewer, placement or communication task. “Create a 16:9 opening storyboard frame that establishes a lone courier entering a flooded train station” gives the model a narrative and compositional job.

A production goal should answer four questions:

• Where will the image appear?
• What must a viewer understand immediately?
• Which element has visual priority?
• What space must remain available for later design work?

For a product page, recognizability may be the priority. For a thumbnail, the silhouette and focal point must survive at small size. For a storyboard, spatial relationships and action matter more than surface polish. This first sentence acts as the acceptance criterion for every later detail.

Build prompts in six controllable layers

Use a consistent information order so that omissions are easy to spot:

1. Purpose: the asset type and its communication job.
2. Subject: identity, objects, wardrobe and defining features.
3. Action: what is happening at the exact captured moment.
4. Environment: location, time, weather and relevant background objects.
5. Composition: shot size, camera position, lens feel, aspect ratio and negative space.
6. Treatment and constraints: lighting, medium, palette, exclusions and elements to preserve.

For example:

Create a 16:9 storyboard keyframe for a science-fiction short. An exhausted adult botanist in an orange utility suit kneels beside a cracked greenhouse window, sealing it with transparent repair film while a dust storm approaches outside. Medium-wide eye-level view, botanist on the right third, damaged window and storm readable on the left, cool interior work lights against warm exterior dust, restrained realistic concept art. Keep the face visible, show both hands clearly, no text, no logos, no extra people.

This prompt is specific without prescribing every pixel. It identifies what the frame must communicate and what would make it unusable.

Prompt Flare for deliberate variation

Flare’s announced speed is most valuable when each result tests a different hypothesis. Do not send one crowded prompt and hope repeated generations discover the right answer. Create variation lanes.

For a campaign key visual, the lanes might be:

• human-first, with the user’s expression carrying the message;
• product-first, with a clean hierarchy and controlled negative space;
• environment-first, showing the product in a credible context;
• graphic-first, using bold shapes and a simplified palette.

Keep the product, audience and format constant while changing one strategic variable. Name the lane at the beginning of the prompt. A review team can then discuss “environment-first” rather than comparing anonymous files.

Flare should answer broad questions: Which composition works? Which setting communicates the idea? Which silhouette reads best? Once the answer becomes “make this accepted image more precise,” the task has moved from exploration to refinement.

Prompt Sunburst with a change list and an invariant list

Editing prompts need two separate instructions: what changes and what stays fixed. A request such as “make the coat blue” leaves the model free to reinterpret the face, crop, lighting and background. A controlled request is explicit:

Change only the adult subject’s raincoat from yellow vinyl to matte navy waterproof fabric. Preserve the subject’s facial identity, age, expression, hairstyle, body proportions, pose and hands. Preserve the camera position, crop, street geometry, reflections, lighting direction and all background objects. Do not add text or logos.

The invariant list is not filler. It defines edit coherence in terms a reviewer can verify. When the result breaks an invariant, reject that revision rather than adding corrective instructions onto a damaged branch.

Make one semantic change per turn. Changing clothing, weather, pose, lens and background together prevents you from knowing which instruction caused a failure. It also makes version history difficult to audit.

Describe camera and layout with observable language

Prompting improves when camera terms correspond to a visible decision. “Cinematic” can mean almost anything. “Low-angle medium-wide shot, subject centered below a tall arch, shallow haze separating foreground and background” is reviewable.

Useful composition instructions include:

• extreme wide, wide, medium, close-up or macro framing;
• eye-level, low-angle, high-angle or overhead viewpoint;
• centered, symmetrical, rule-of-thirds or edge-weighted placement;
• foreground, middle-ground and background relationships;
• negative space on a named side;
• visual hierarchy at thumbnail size;
• portrait, square or landscape delivery format.

Avoid stacking contradictory terms such as “minimalist, highly detailed, crowded marketplace with empty negative space.” Decide which requirement is dominant and explain where detail belongs.

Control style through production attributes

Instead of naming a living artist or relying on a broad aesthetic label, describe the properties that matter: medium, shape language, surface texture, contrast, palette, lighting and degree of realism.

“Editorial vector illustration with flat geometric forms, two-tone navy and acid green palette, hard edges, minimal texture and generous white space” is more transferable than a reference name. It also gives the designer criteria for judging whether subsequent images belong to the same system.

For a sequence, create a small style block and reuse it unchanged. Keep subject and scene instructions separate from this block. If the style drifts, you can identify whether the reusable block or the changing scene description caused the problem.

Put text outside the image when accuracy matters

Image models can render text, but brand, legal and pricing copy should still be treated as structured design content. Generate the scene with intentional blank space, then add verified copy in a design or editing tool. This keeps spelling, hierarchy, accessibility and localization under human control.

If text inside the image is necessary, specify the exact string, location and priority. Review every character. Do not treat a plausible label as an accurate one, especially for packaging, medical diagrams, maps, interfaces or safety instructions.

The same rule applies to factual symbols. A convincing interface is not necessarily functional, and a realistic technical diagram is not evidence of correctness.

Use negative constraints sparingly

A short exclusion list can prevent predictable defects: no logos, no watermarks, no additional people, no unreadable interface text. A long list of negatives often competes with the main brief and makes review harder.

Phrase the desired state positively whenever possible. Instead of “no busy background,” request “an uncluttered pale wall with one soft shadow.” Instead of “do not crop the hands,” request “both hands fully visible inside the frame.” Positive constraints describe the target, not only the failure.

Keep safety and rights checks outside the prompt as well. Confirm that reference images are authorized, that real-person likeness use is appropriate, and that generated claims are not misleading. A prompt cannot replace editorial review.

Create a reusable prompt card for characters and worlds

Long-form visual stories need stable references. Store a character card with identity features that should persist: apparent age range, face shape, hair, body proportions, signature wardrobe and prohibited changes. Store a world card with architecture, materials, technology level, color logic and recurring landmarks.

In Elser AI, an approved image can become part of a browser-based story, storyboard or image-to-video workflow. In ElserStudio, a production can organize scripts, World assets, shots, takes and the final composition. The safe linkage is an asset handoff: export the approved image and prompt record, then confirm the currently available import and provider options. Do not assume a native GPT Image 2.6 integration unless the live product documents it.

The prompt card should travel with the image. Record the model shown by the source tool, date, aspect ratio, prompt, edit history and approval notes. This metadata matters when a later shot must be recreated.

Review prompts with a failure matrix

Do not rewrite the whole prompt after every weak result. Classify the failure first:

| Failure | Likely prompt problem | Better revision | | --- | --- | --- | | Wrong focal point | Purpose or hierarchy is vague | State the primary subject and intended viewer takeaway first | | Character drift | Identity features are mixed with scene details | Use a fixed character card and explicit invariants | | Busy composition | Too many equal-priority objects | Remove nonessential objects and define negative space | | Inconsistent sequence | Style and world rules change between prompts | Reuse locked style and world blocks | | Edit damages other areas | Only the requested change is named | Add an invariant list and change one layer at a time | | Unusable marketing copy | Text was delegated to the image | Reserve space and typeset verified copy separately |

This matrix creates information gain. A failed generation becomes evidence about the brief instead of a reason to roll the dice again.

A complete workflow from prompt to production

Start with a written acceptance criterion. Use Flare to produce a small, intentionally varied set. Score the candidates for communication, composition, subject integrity and downstream usability. Select one direction, then use Sunburst for constrained refinement. Export the approved still with its prompt and version record. Finally, move it into the relevant Elser workflow for story development, shot planning or animation.

The model choice is only one part of the system. Naming files, preserving sources, tracking revisions and approving claims are what make generated images usable by a team.

Frequently asked questions

How long should a GPT Image 2.6 prompt be?

Long enough to define the job, subject, action, environment, composition and constraints. A shorter structured prompt is usually easier to debug than a paragraph of adjectives. Add detail only when it changes a visible decision.

Should I use the same prompt for Flare and Sunburst?

Reuse the stable creative brief, but change the task. Ask Flare to explore distinct directions. Ask Sunburst to refine an accepted image and explicitly preserve approved elements.

Can GPT Image 2.6 generate accurate text?

The supplied announcement does not establish a text-accuracy guarantee. Even when text looks correct, verify it character by character. Typeset legal, pricing and brand copy separately when accuracy matters.

What if the model keeps changing the character?

Use one approved base image, a fixed character card, an invariant list and one change per edit. Reject a revision when identity breaks rather than continuing to edit the damaged result.

Is GPT Image 2.6 available through the API?

The OpenAI announcement identifies the models, but the public API references checked on November 5, 2026 had not yet listed 2.6 model identifiers. Verify the live model catalog before implementing it.

Conclusion

The best GPT Image 2.6 prompts behave like production briefs. Flare benefits from clearly separated variation lanes; Sunburst benefits from explicit changes and invariants. Define the decision, preserve the prompt record and hand off only approved assets. That workflow creates better images and makes them easier to use in Elser AI or organize across a multi-shot ElserStudio production.