How to Use Reference Images on Image 2 (Up to 16)

Reference images let GPT Image 2 and GPT Image 2.5 follow a face, product, pose, or style instead of inventing everything from text. On Image 2 you can attach up to 16 references per run—from upload, history, or a previous output.

Open the generator to try the flow below.

When references help

Add references when you need:

• Identity — the same person, mascot, or packaging across styles
• Composition — keep a pose, crop, or layout while changing the scene
• Style lock — match lighting, materials, or brand look from a photo
• Edit loops — use a generated frame as the next reference and refine

Skip references for open-ended brainstorms where you want the model free to invent.

How many references to use

More is not always better.

• 1–3 — identity or a single hero product (clearest signal)
• 4–8 — style boards or multi-angle product packs
• 9–16 — only when each image adds a distinct cue; noisy sets confuse the model

Prefer sharp, well-lit crops. Blurry screenshots and busy collages dilute the brief.

A clean Image 2 workflow

1. Write the prompt first: subject, setting, lighting, then style.
2. Attach the strongest reference(s)—identity photo before mood boards.
3. Pick GPT Image 2, Flare, or Sunburst.
4. Generate, review, then either download or use as reference for the next turn.
5. Clear old references when you switch projects so leftover faces do not leak into a new brief.

You can multi-select from generation history when adding references, up to the remaining slot count.

Prompt + reference pairing tips

• Say what should stay: “keep the same person / bottle / logo placement.”
• Say what should change: “new background, evening light, poster layout.”
• Quote on-image text in quotation marks.
• On Sunburst, change one local element per turn after the first frame lands.

Common mistakes

• Dumping 16 unrelated images with a one-line prompt
• Mixing two different identities in one batch
• Forgetting to remove refs when starting a new concept
• Expecting references to fix a vague prompt—text still leads

Next steps

Generate with references on /generate, browse proven prompts in the gallery, or read Flare vs Sunburst if you are choosing a GPT Image 2.5 route for edit-heavy work.