Skip to main content

Browse guides

Open the full guide library for quick switching.

Image Generation9 min readUpdated August 25, 2026

Image Generation: Prompts, References, and Presets

Review a fictional recorded image workflow with a concrete prompt, optional helper inputs, and a disclosed sample result.

This guide uses a fictional analytics product and a recorded sample. It is not customer work or proof that another run will reproduce the result. The walkthrough covers freeform prompt, platform preset, sketch, references, and preview review.

What you’ll learn

  • How to keep the workflow in General instead of switching to a template
  • How to structure a prompt for the dashboard visual shown at the end of the guide
  • Where Sketch and References live, even when you leave them empty
  • How to explain preview actions without hand-waving over the decision point

Before you start

  • Open Generate
  • Switch the widget to Image mode
  • Keep LinkedIn Post (4:5) selected in Platform preset so the setup matches the final output
  • Leave brand guidelines off for this specific run so the prompt alone drives the image
01Choose the setup

Stay in General and keep the LinkedIn Post (4:5) preset selected

Start in Image mode with General selected. Keep the walkthrough focused on the prompt rather than mixing in too many controls at once.

  • Keep Platform preset on LinkedIn Post (4:5).
  • Leave the Try a prompt chips unused for this run.
  • Leave Sketch and References closed until the next step.
  • Do not switch to a content-type template.
  • Keep the model on the default image option shown in the UI.

Expected result

You have a clean prompt-driven image workflow with the same 4:5 publishing preset used throughout the rest of the guide.

02Describe the visual

Use a prompt that reads like a design brief

Paste this exact prompt into the main textarea:

Create a polished product marketing visual for an analytics dashboard. Dark navy background, glowing glass panels, clean SaaS layout, subtle cyan accents, and a premium editorial lighting style.
  • Name the subject first: the analytics dashboard visual.
  • Name the visual treatment second: dark, glowing, premium, editorial.
  • Keep Platform preset on LinkedIn Post (4:5).
  • Keep the brief tight enough that another human could art-direct from it.
  • Click Generate.

Expected result

The first draft is specific enough to judge immediately instead of leaving the operator to guess what the model should do.

03Show the helper inputs

Open Sketch and References so the team sees where manual steering lives

Open the two helper controls that sit below the main image prompt. You do not need to upload anything in this specific run; the goal is to teach where the controls are and what each one is for.

  • Click Sketch to reveal the rough layout canvas.
  • Click References to reveal the Add slot for reference images.
  • Once both panels are open, the pills change to Hide sketch and Hide refs.
  • Leave both panels empty for this prompt-only example.
  • Use Sketch when placement matters most, and use References when mood, angle, or lighting matters most.

Expected result

The operator understands where to add layout guidance or reference imagery when text alone is not enough.

04Approve or refine

Use preview to decide whether to keep, revise, or brand the result

When the dashboard visual appears, narrate the decision using the button labels exactly as they appear in the footer.

  • Use Generate again when the direction is right but the execution is weak.
  • Use Edit prompt when the brief itself needs better instructions.
  • Use Save if the raw draft is worth keeping for review.
  • Use Brand it when the image is approved and now needs an overlay or logo.

Expected result

A strong output can be saved immediately or moved into branding, while a weak one can be corrected without restarting the whole explanation.

Sample generated image

Sample generated output
The same accepted dashboard visual shown in the preview above

Recorded fictional sample from the documented run; it is not a repeatability, quality, or customer-performance claim.

Keep the baseline result

The walkthrough above is the baseline ImageLayer image flow. Model availability and exact credit quotes come from the live selector and server-provided catalog, so check both before comparing results.

Use GPT Image 2 as a second pass when a team wants to compare the current OpenAI image path with the default Gemini 3.1 Flash Image path.

When to choose GPT Image 2

Consider GPT Image 2 when you want to compare its supported contract against the other admitted image models for a constrained draft:

  • the output includes short, exact text such as headlines, labels, badges, or UI copy
  • a reference image supplies product, identity, layout, or lighting guidance that will be reviewed closely
  • the asset is a product photo, packaging mockup, infographic, diagram, UI screen, poster, or ad creative
  • the team has reviewed the current quote and wants a controlled model comparison

For rough exploration, a lower-credit image model can be the right first pass. For constrained assets, compare the available models with the same prompt and references; outputs vary and still require review.

How to compare Gemini 3.1 Flash Image and GPT Image 2

Run the same asset idea twice and document both runs:

  1. Baseline run: keep Gemini 3.1 Flash Image, the current default image model, use the same platform preset, and save the result as the default path.
  2. GPT Image 2 run: switch only the model to GPT Image 2, keep the same preset, and tighten the prompt around exact text, product labels, UI copy, or preservation constraints.

In customer-facing docs, always state:

  • model used
  • content type or General
  • platform preset
  • whether Apply brand guidelines was on or off
  • the exact prompt or exact field values
  • the actual generated image

That lets a reviewer compare recorded settings, quotes, and outputs without claiming that one model is consistently faster, higher quality, or more reliable.

GPT Image 2 prompt structure

When the model is set to GPT Image 2, teach customers to write prompts in labeled blocks instead of one long paragraph:

Scene:
Where the image exists, including background, time of day, and environment.

Subject:
The main product, person, screen, or object.

Important details:
Materials, texture, camera angle, lighting, composition, exact text, and mood.

Use case:
Product mockup, social ad, UI screen, infographic, poster, or editorial photo.

Constraints:
What must not appear, and what must stay unchanged.

For revisions against a reference image, switch to edit language:

Change:
The one thing that should change.

Preserve:
The product geometry, label text, face, pose, layout, lighting, camera angle, or background that must stay locked.

Constraints:
No extra objects, no logo drift, no additional text, no watermark.

The preserve list gives the edit model explicit guidance, but it cannot guarantee that every listed detail remains unchanged.

Review practices to share with operators

  • Write prompts like short art directions, not keyword piles
  • Quote exact text and specify placement, type style, and size when the image includes readable copy
  • Use Change / Preserve / Constraints for reference-image edits
  • Open Sketch and References only when the base prompt is no longer enough
  • Use preview before branding so teams do not waste finishing steps on a weak base image
  • Keep the first demo simple, then add helper inputs in the second pass

Previous guide

Getting Started with the Widget

Walk through a fictional recorded path from a blank image prompt to a reviewed launch-graphic draft.

Next guide

Content Type Templates: Structured Inputs for Graphics

Use structured fields instead of an open-ended prompt, then review a fictional recorded stat-card sample.