Gemini Image Editing Guide: Edit Photos by Chatting

Editing in Gemini (gemini.google.com) works the way you would brief a retoucher. You upload a picture, say what you want changed, and get a new version back. There are no layers, masks or brushes, so the result depends on how clearly you describe the change and how carefully you protect what should stay the same. This guide covers the editing workflow, keeping a character or product consistent across a series, and borrowing a style from a reference image.

When you attach an image and ask for a change, Gemini uses a native image model (the one nicknamed "Nano Banana" in 2025) rather than generating from scratch. Its strength is conversational editing: it reads the picture, applies your instruction and tries to keep identity, pose and composition intact unless you ask otherwise. That is why the same face, outfit or product can survive several rounds of edits in one chat.

You can upload one image or several. With more than one, Gemini can combine them: a product from one photo placed into a scene from another, or a character’s outfit swapped for one you show it. Outputs carry a SynthID watermark, and upload and edit limits vary by plan and region.

Step by step: a clean edit

1. Start a new chat for the job. Edits build on the conversation, so keeping one subject per chat avoids the model mixing earlier instructions into later ones.
2. Upload the best source you have. Sharp, well-lit, with the subject clear of the frame edges. The model cannot restore detail that was never there.
3. Describe one change. "Change the jacket to dark green" rather than "make it look better". Vague requests produce vague edits.
4. Say what must not change. "Keep the face, pose, lighting and background exactly the same." This sentence is the most useful thing you can type in Gemini.
5. Review at full size. Zoom in on hands, edges and any text. Then ask for the next change in the same thread.
6. Record the chain. Save the final image and copy the sequence of instructions into a note.

Keeping a character or product consistent

Because edits start from your upload rather than a text description, the subject’s identity has a fixed anchor. To hold it across a series:

• Always edit from the original or the last approved version. Each generation adds a little drift. If round four looks off, go back to round two’s image and re-attach it instead of editing forward.
• Name the subject the same way every time. If you call her "the woman in the yellow raincoat" in message one, keep that exact phrase. Changing descriptors invites the model to reinterpret.
• Change pose and scene, not identity. "Same woman, now sitting at a cafe table outside, same raincoat" works. Asking for a different age or hairstyle is a new character.
• For products, keep the hero angle. Place the product in new scenes with "same bottle, same label, same angle, new background: a marble kitchen counter in morning light." Labels survive best when they are not asked to rotate.
• Use multiple uploads as a reference sheet. Attach two or three views of the same character and say "this is the same person; keep her consistent".

The Midjourney consistent characters guide shows the same ideas in a parameter-driven interface.

Style references and transfers

Gemini can take the look of one image and apply it to another. Upload the reference and the target, then be explicit about which is which: "Apply the colour palette and brushwork of the first image to the second image, keeping the second image’s subject and composition." Without that sentence the model may blend subjects instead of styles.

Describe the style in words as well. "Muted teal and rust palette, grainy film look, soft contrast" gives the model something to hold onto if the reference is ambiguous. Describe qualities rather than naming living artists. For a brand, upload one approved image as the style anchor at the start of every chat so each new asset matches the set.

Settings at a glance

Common problems and fixes

• The face changes slightly each round. Re-attach the original and ask for all edits relative to it, or attach the original plus the latest version and say "match the face in the first image".
• The edit spread beyond the area you named. Add a stronger preserve clause and name the region: "only the sky changes; the building and street stay pixel-identical".
• Text on packaging is garbled. Ask Gemini to keep the label unchanged, or regenerate without text and add it back in a design tool.
• Style transfer changed the subject. State which image is the style source and which is the content source, and say "do not add elements from the first image".
• Requests involving real people are declined. Edits of identifiable people may be limited. Work with fictional characters or products.

Try it

[Attach a product photo]
Keep this exact bottle, label, angle and lighting. Replace the background with a pale terrazzo countertop in soft morning light, add a faint natural shadow under the bottle, and keep the overall colours neutral. Do not change anything on the bottle itself.

Once that works, follow up with "Same bottle, same angle, now on dark slate with a single warm spotlight" to build a matching set. If the label drifts, re-attach the original photo and ask again.