Create images with AI Studio

An image request usually starts with a deceptively small sentence: “Make this feel more premium.” The useful part is everything around it. Which product must stay recognizable? Is the result a rough direction or the final campaign visual? Does it need readable text? Is speed more important than preserving the reference exactly?

The short version

Choose Image, leave Model, Style, and Ratio on Auto, and keep Count at 1 for the first draft. Add references only when the result must preserve something specific, then describe the subject, composition, audience, and details that must not change. Review the first result before increasing quality or generating more versions.

Choose Image in AI Studio when the result should be a visual object you can place, compare, comment on, and share from the canvas. AI Studio can generate a fresh image from text or use selected and attached images as references, but the model you choose changes what it can preserve, how long it usually takes, and the relative cost.

Use another output when the job is not primarily visual:

For the broader creation flow, read Create content with AI Studio. This article stays focused on Image.

Start with the least expensive useful draft

1. Open the canvas where the image belongs.
2. Select Image in AI Studio.
3. Leave Model, Style, and Ratio on Auto for the first attempt.
4. Keep Count at 1 until the prompt and references produce the right direction.
5. Add selected canvas objects or image attachments only when the result must preserve something specific.
6. Describe the subject, composition, audience, visual treatment, and details that must not change.
7. Submit, then review the image on the canvas before generating more versions.

This workflow keeps the first decision cheap. A weak prompt multiplied by Count 10 is still a weak prompt, only ten times.

The picker puts the important tradeoffs in one place:

Choose a model by the job, not the brand

Auto is the right default for most requests. It considers the prompt, references, expected time, and cost, then chooses a compatible route. If a selected canvas image or attached image is present, text-only models are unavailable and Auto stays on a reference-capable path.

Choose a specific model when you know which tradeoff matters. The Cost value below is the relative value shown in the picker.

The labels matter:

• Reference/edit models accept selected or attached images and can preserve or change them.
• Text-only models create from the written prompt. Flux 2 Pro and every Imagen 4 option become unavailable when an image reference is present.
• A text-only model is not a worse model. It is the wrong model only when the request depends on an existing image.

The Nano Banana 2 and Pro catalog can handle larger reference sets and higher-resolution work than the base Nano Banana route. GPT Image 2 is the stronger GPT choice for precise edits and larger output. GPT Image mini is for cheaper drafts, not for upscaling. Imagen 4 is strongest when you want a new photorealistic or illustrative image and do not need to preserve a source.

Set Count, Style, and Ratio deliberately

Count

Count accepts 1-10 images. One is the default because every additional result starts another image job and can consume more credits.

• Use 1 while testing the prompt, references, and model.
• Increase Count when the direction is already right and you need variants.
• Use a smaller batch with an expensive model, then keep the best result instead of repeating a large batch.

Each result gets its own placeholder. In a multi-image request, deleting one processing placeholder cancels that output slot. Deleting every image placeholder cancels the remaining task.

Style

Style is direction, not a substitute for a concrete prompt. “Infographic” still needs the message, audience, hierarchy, and facts that belong in the visual.

Ratio

All current picker models support these five explicit ratios. Auto chooses among them from the request. If a saved draft contains a choice the selected model cannot use, AI Studio normalizes the request to a supported route rather than sending an invalid combination.

Give the model references it can actually use

A good image prompt answers five questions:

1. What is the subject?
2. What should the composition show?
3. Who will see it?
4. What visual treatment should it use?
5. What must stay unchanged?

For example:

Create a 16:9 launch visual for a client presentation. Keep the bottle shape and label from reference #1. Use the warm amber palette from reference #2. Place the product on the right with open space for a title on the left. Do not add text inside the image.

That prompt separates identity, style, layout, and constraints. “Make this premium” leaves all four for the model to guess.

Attachments are numbered before selected canvas objects, so references such as #1 and #2 stay clear. Use those numbers when several images play different roles. Keep only references that affect the result. A mood image, product image, old layout, and unrelated screenshot in one request create four competing instructions.

The safest selection rule is simple:

• Use Nano Banana 2 🍌 when references, multilingual text, or a moderately dense layout matter.
• Use Nano Banana Pro 🍌 when the reference set or layout is genuinely difficult and the result is headed for external review.
• Use GPT Image 2 for a precise edit or explicit text and layout instruction.
• Use GPT Image mini · Low or GPT Image mini to test an edit cheaply.
• Use Flux 2 Pro or Imagen 4 only when you want a fresh image from text.
• Use Auto when none of those needs dominates.

This walkthrough shows a reference image and canvas context becoming a finished image beside the source material:

Spend credits on decisions, not retries

Use this sequence for expensive work:

1. Write the prompt with Count 1, Model Auto, Style Auto, and Ratio Auto.
2. Generate a first composition.
3. Fix the prompt before increasing quality. Add the missing subject, layout, reference number, or constraint.
4. Use a low-cost tier to test a precise edit.
5. Move to Nano Banana Pro, GPT Image 2 · High, or Imagen 4 Ultra only when the direction is stable.
6. Increase Count only when you need useful alternatives, not reassurance.

Manual edits can be cheaper than asking the model to rediscover a nearly finished result. If one label, crop, or small canvas detail is wrong, fix it directly when possible.

Deleting a finished image does not refund credits. A successful result can consume credits even when you dislike it, remove it, or generate another version. For the workspace balance and ledger, read Understand AI credits.

Watch progress and cancel from the canvas

After submission, AI Studio places image placeholders on the canvas. The server updates each placeholder as its image finishes, so a batch can reveal usable results before every slot is done.

While a placeholder is processing, selection and deletion remain available. Editing, preview, download, duplicate, copy, lock, and thumbnail actions remain unavailable until the result finishes.

To cancel:

• Delete one processing image placeholder to cancel that slot.
• Delete every image placeholder from the request to cancel the whole image task.
• Deleting the canvas also stops generation still running there.

Do not resubmit because a slower model still shows progress. GPT Image 2 and GPT Image mini show a 60s-2min picker hint, and the relative speed hints are expectations rather than guarantees.

When image generation does not behave as expected

If you contact support, send the canvas link, approximate request time, selected Model, Count, Style, Ratio, a prompt summary, whether image references were selected or attached, and whether credits were blocked or consumed. Do not send private customer images or source files unless support specifically needs them and you are allowed to share them.