Editing Real Photos With Gemini's Image Model (Not Just Generating New Ones)

You have a real photo. Maybe it's a product shot with a cluttered background, a headshot with bad overhead lighting, or a listing photo with a stray trash can in the corner. You don't want a new image in that style. You want that exact photo, fixed. That's a different job from typing a description and getting an original image back, and Gemini's image model (known as Nano Banana, with Nano Banana 2 as the current everyday version and Nano Banana Pro available to paid Google AI subscribers) handles it well once you stop treating it like a text-to-image generator and start treating it like an editor you're giving instructions to.

If you haven't used Gemini's image tools at all yet, the Complete Beginner's Guide to Gemini covers how generation works first. This article is specifically about the upload-a-photo-and-fix-it workflow, which behaves differently enough that it deserves its own treatment.

What "editing" means here

When you upload an existing photo to Gemini and ask for a change, the model isn't redrawing the scene from a text description. It's looking at the actual pixels you gave it, understanding what's in them, and modifying specific parts while leaving the rest recognizably intact. That's why it can keep a person's face, a product's exact shape, or a room's layout the same while changing one thing about the picture.

The practical difference shows up immediately in how you write the prompt. A generation prompt describes a whole scene from nothing. An editing prompt describes one change against a scene that already exists:

The shape of a good edit

The edits that actually work well tend to follow the same arc: a clear before, one specific instruction, and a result you check before moving to the next fix.

Edits that come up constantly

Swapping a background. This is the most common request, and the one people over-explain the most. You don't need to describe your ideal studio setup in three sentences.

That last sentence matters more than it looks. Without it, the model sometimes "helpfully" smooths out a texture or color on the product along with the background, since it's treating the whole image as fair game unless told otherwise.

Fixing lighting. A photo shot under mixed indoor light usually reads as yellow, flat, or unevenly shadowed. Naming the specific problem gets a better result than "make it look professional."

Removing an object. Gemini is genuinely good at this when the object is a clear, separable thing, less good at it when the object overlaps something important like a face or hands.

Cleaning up a headshot. People use this for LinkedIn photos and team pages more than anything else: fixing a distracting background, evening out lighting, or removing a stray object from the frame, without turning the person into an obviously different-looking version of themselves.

A bad, better, excellent edit, worked through

Say you're listing a bedroom on a rental site, and the photo is technically fine but won't make anyone stop scrolling: the light is patchy, and there's a pile of laundry on the chair in the corner that shouldn't be there.

Bad prompt: "Make this bedroom look nicer."

That produces something like: the room reads a little brighter overall, but the laundry pile is still sitting on the chair, since nothing told the model it was the problem. The window light and the overhead light get nudged up by roughly the same uniform amount, so the space doesn't look meaningfully more staged, just slightly washed out. Nothing structural about the actual mess changed, because "nicer" isn't a target, it's a mood, and the model can only guess at what would produce one.

Better prompt:

That produces something like: a warmer, more evenly lit version of the same room, genuinely closer to listing-ready. But the laundry pile is still there, because "staged" never named it. The photo reads better in a quick scroll, but the one detail that would actually change a renter's mind, an uncluttered room, never got addressed, since the prompt described a feeling instead of a fix.

Excellent prompt:

Notice what changed between the three prompts isn't tone, it's specificity on two separate axes: what exactly should change, and what exactly should not. "Bad" answers neither question. "Better" answers the first one loosely (a category, not an object) and skips the second entirely. "Excellent" answers both: the laundry pile is the named target, and the furniture, wall color, and floor are explicitly protected. That second half matters as much as the first. Without it, a model asked to fix lighting will sometimes also smooth a texture or nudge a color it decided looked better, because nothing told it that part of the photo was off-limits.

Where first attempts go wrong

The single most common mistake is asking for too much in one instruction: "fix the lighting, swap the background, remove the guy in the back, and make it look more professional." Each of those is a distinct edit, and stacking them into one prompt gives the model too many competing priorities at once. It often does a mediocre job on all four instead of a clean job on one. Work through changes one at a time, checking the result before moving to the next.

The second mistake is expecting pixel-perfect fidelity on the hardest details: hands, small text in the background, reflections, and exact brand colors. These are the parts of a photo that are genuinely difficult for any image model to edit without introducing small distortions. If a result is close but a hand looks slightly off, it's usually faster to ask for a targeted fix ("just redo the left hand, keep everything else") than to regenerate the whole image and hope.

When this isn't the right tool

If you need pixel-exact color matching to a brand's hex value, or retouching precise enough for print production, a dedicated photo editor with layer-level control is still the more reliable choice. Gemini's edits are very good for everyday marketing, listings, and profile photos, but it's an approximation engine, not a color-calibrated one.

It's also worth pausing before editing photos of real, identifiable people in ways that change what actually happened, altering their appearance, placing them somewhere they weren't, or changing the context of an event. That's a different use case from cosmetic touch-ups, and it carries real consent and honesty questions worth thinking through before you hit generate.

Where you'll actually use this

Nano Banana editing works the same way whether you're in the main Gemini app or the Chrome side panel, which can edit an image right on the page you're looking at without switching tabs (this is available on desktop platforms; check your account's current Gemini plan for exact limits). If you're doing a batch of product photos for a storefront, that in-browser version is often the faster path since you never leave the listing page you're editing.

Official sources

• Generate & edit images with Gemini Apps
• Edit images from the web in Gemini in Chrome

Checked on September 21, 2026. Features, plans and names change often, so the vendor's own pages are the final word.