Image workflow
How to Use GPT Image 2: A Practical Creation and Editing Workflow
Turn one visual brief into a reviewable generation or edit by defining the deliverable, assigning reference roles, choosing the frame, and checking every protected detail.
Editorial workflow illustration created for ZMS AI. It is not a GPT Image 2 output or a current ZMS AI live-generation test.
Define the deliverable before opening the generator
Start on ZMS AI by writing the image’s real job: where it will appear, who should respond to it, what the viewer must notice, and what would make the result unusable. A product hero, social portrait, editorial illustration, and wide landing-page banner need different compositions even when they share the same subject.
Record the final aspect ratio, minimum working resolution, safe space for copy, approved visual facts, and the person who will review the output. These decisions turn a loose idea into a specification. They also prevent an attractive image from being approved before anyone checks whether it fits the destination.
The current GPT Image 2 workspace prepares tasks through a registered ZMS driver, but signed-in task completion, returned media, and live credit charging have not yet been verified. Treat creative readiness and service readiness as separate questions from the beginning.
- Name the channel, audience, and central message.
- Choose the intended frame before writing composition details.
- List protected facts and obvious rejection conditions.
- Separate task preparation from proof of live service delivery.
Choose text generation or reference-led editing
Use text alone when the subject, layout, and art direction can be interpreted from a written brief. Adding any reference image changes the current ZMS request to an edit task. There is no separate generation-versus-editing switch, so the input set itself determines the path.
Open the GPT Image 2 workspace and attach only the visual facts that matter. The interface accepts one to ten references, but the maximum is not a target. One authoritative product image and one palette reference are often easier to reconcile than a large group with overlapping or contradictory roles.
Upload only material you are authorized to use. If a reference contains a person, confirm the required consent and permitted use. If it contains a trademark, packaging, artwork, or licensed photography, decide whether the requested transformation and publication context are allowed before generation.
Give every reference one explicit role
Label the authority of each input in the prompt. One image may control subject identity, another product geometry, another palette, and another composition. Say which visual facts can change and which must survive the edit. Without that hierarchy, two references can silently compete for the same decision.
For a controlled product edit, write that the first image owns shape, proportions, cap details, and material; the second contributes only the pale-blue and sage-green palette; and the prompt defines the new room and lighting. This is more precise than asking the system to “combine the references.”
The present GPT Image 2 interface does not expose an input-fidelity control. The upstream model is documented as processing image inputs at high fidelity, but ZMS does not provide a user-facing fidelity setting on this route. Review every supposedly protected region rather than assuming preservation.
Choose ratio and resolution from the final placement
The ZMS workspace lists ten ratios: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. Pick the destination frame first, then describe subject scale, negative space, horizon, crop, and text-safe areas for that canvas.
Use 1K for an exploratory pass and 2K when the selected concept needs more working detail. Resolution does not repair an unclear hierarchy, incorrect text, malformed geometry, or weak reference preservation. Approve composition and meaning before treating extra pixels as production readiness.
Quality is not a visible choice on this page. The current adapter submits medium quality internally. The workspace also lacks output-format, compression, transparent-background, mask, batch-count, negative-prompt, and seed controls. Do not promise a deliverable that depends on an absent option.
Write the prompt as a compact art-direction brief
Put the deliverable and main subject first. Follow with composition, action or state, environment, camera perspective, lighting, materials, palette, and constraints. Each phrase should describe something a reviewer can see. Replace broad words such as “premium” with specific surface, light, spacing, and hierarchy decisions.
For an edit, split the prompt into change and protection. For example: move the same white lamp into a pale-sage studio, add one clear glass vase on the right, preserve lamp geometry and finish, keep the camera angle fixed, and leave clean space above the subject for a headline. The instruction explains both the new result and its continuity contract.
Avoid embedding several unrelated deliverables in one request. A square catalog tile and a 21:9 homepage hero may share references, but they should be separate tasks with framing written for each destination. That keeps failures attributable to one brief instead of a compromise between incompatible layouts.
Verify the task before spending credits
Before Generate, confirm that the selected ratio matches the destination, 1K or 2K is intentional, every reference opens correctly, and the prompt assigns each input a role. Save the exact prompt, filenames, and settings outside the transient form so a later result can be reproduced or diagnosed.
The current client estimates six credits for one GPT Image 2 task. That value is configured in the interface, but a real account debit has not been verified. Check the displayed balance and current product terms, then use a low-risk validation task before planning a larger batch.
After submission, record whether authentication succeeds, the provider accepts the task, status polling progresses, credits change, and a playable or downloadable image URL returns. A missing result is an operational failure, not evidence that the creative prompt was poor.
Review the full image, not only the focal point
Inspect the returned image at full size. Check faces, hands, text, symbols, product geometry, reflections, repeated patterns, edge transitions, contact shadows, and background objects. A convincing thumbnail can hide errors that become obvious in a campaign crop or product page.
Compare every protected fact with the authoritative reference. Mark the largest failure category: instruction following, identity, geometry, composition, material, color, text, or delivery. Do not blend all feedback into a longer prompt before deciding which visible problem matters most.
Use the broader AI image workspace when another model or control set better matches the brief. Prompt inflation cannot replace a required mask, transparent-background option, exact output format, or other control that the current route does not expose.
Revise one failure category at a time
Keep the strongest result as a baseline. If composition failed, revise subject scale and negative space without replacing the references or palette. If geometry drifted, strengthen the protected shape language while keeping the frame and environment stable. If the style is wrong, change treatment after confirming that the subject and layout already work.
Changing prompt, references, ratio, resolution, and purpose in the same pass creates an unrelated image rather than a useful comparison. Controlled revisions reveal which instruction helped, reduce accidental regressions, and give teammates a repeatable production recipe.
Name every pass with the project, deliverable, version, and review state. Save the task identifier, prompt, reference names, settings, output, actual credit behavior, and one sentence describing the decision. This small log separates creative iteration from service debugging.
Hand off the approved image with its evidence
An approved asset should travel with the brief that created it. Preserve the accepted prompt, authorized inputs, aspect ratio, resolution, actual output dimensions, task history, rights notes, and review decision. Record any conventional edits made after generation so the published file is not mistaken for an untouched model result.
Return to the GPT Image model family when comparing GPT Image 2 with the previous GPT Image 1.5 route. OpenAI documents a broader and changing image API surface, while ZMS exposes a narrower provider contract. Recheck current controls and lifecycle notes instead of treating this guide as a permanent capability list.
The durable workflow is simple: define one deliverable, use the minimum authorized reference set, assign every input a role, choose the frame before composition, verify the live task, review protected facts at full size, and preserve the evidence behind the approved result.