AI Image Editor
Upload a picture and describe the change you want. This is editing, not generating: the photograph you started with is still the photograph you end up with, minus the thing you asked to remove or plus the thing you asked for.
How to edit an image with AI
Upload the image
A photograph, a screenshot, a product shot, a piece of artwork, or a still you generated earlier. The point of this tool is that the picture already exists and you want to keep most of it.
Say what should change
Name the thing and name the change. "Remove the car parked behind the shop", "make the sky overcast", "put the bottle on a dark stone surface". An instruction that identifies a specific element gets a specific edit; "improve this" gets an opinion.
Check what stayed the same
The first thing to look at in an edit is not the part you changed. It is everything else. Faces, logos, hands, text and reflections are where an edit gives itself away, so inspect those before you accept the result.
Iterate, then take it forward
Make one change per pass and stack them, rather than describing five at once. When the picture is right it goes to your library, ready to be upscaled, animated into video, or edited into a sequence with voiceover and captions.
What people edit
Backgrounds that were never going to work
A good shot of the right subject in the wrong place. Replacing the setting is far quicker than reshooting it, and unlike generating a fresh picture, the subject stays the exact object you photographed.
Removing the thing that ruins the frame
A bin, a cable, a passing stranger, a reflection of the photographer. These are the edits people used to open Photoshop for, and describing them in a sentence is a shorter route to the same place.
Product shots in new settings
One photograph of the product, placed on marble, on wood, on a desk, outdoors. The packaging keeps its real shape and its real label, which is precisely what a generated lookalike would not do.
Preparing a still before animating it
An image-to-video model inherits every problem in the frame you give it. Cleaning up the background and fixing the light first is the cheapest fix available, because stills come back faster than clips.
Reworking a shot for a different shape
A landscape photograph and a vertical post want different framing. The aspect ratio sits next to the instruction and defaults to matching what you uploaded, so you can ask for the change and the shape you need in one pass.
Fixing what the generator got nearly right
A generated image where the composition is excellent and one detail is wrong is not a reason to start again. Edit the detail and keep the composition you liked.
Consistency across a set
When a batch of images has to feel like one series, editing them towards a common palette and light is more reliable than hoping the next generation matches the last one.
What is AI image editing?
AI image editing means changing an existing picture by describing the change in words, rather than selecting, masking and painting it by hand. You supply the image and an instruction: remove this, replace that, relight the scene. The model returns the same picture with that change made.
The important word is "existing". A generator starts from nothing and invents a picture that matches your description. An editor starts from your picture and is meant to preserve it. That difference decides which tool you should be using, and it is the mistake most worth avoiding: if you ask a generator for a version of your product shot, you get a picture of something in the same category, not a picture of your product.
What separates a good edit from a bad one is what the model leaves alone. Anyone can change a background; the skill is in the untouched 90% of the frame coming back identical. The same face, the same label, the same grain, the same shadow directions. Judge results on the parts you did not ask about.
Resolution is part of the same argument. Nano Banana 2 handles editing as well as generation, up to 4K, so an edit does not have to mean handing back a smaller or softer version of what you uploaded.
Editing an image vs generating a new one
These two tools look adjacent and are used for opposite things. Generation answers "make me a picture of this". Editing answers "keep this picture, change that part of it". The test is whether a specific real thing has to appear in the result: your product, your premises, a person, a logo, a photograph with sentimental or evidential weight. If it does, you edit. If nothing in the frame has to be any particular object, you generate.
The failure mode is quiet, which is why it catches people out. Ask a generator for "my kettle on a marble counter" and you will get a kettle on a marble counter. It will be a plausible kettle. It will not be yours: the spout will be a different shape, the badge will say something almost-but-not-quite right, and the version of it you put in an advert will not match the box the customer opens. An edit starting from a photograph of the actual kettle cannot make that mistake, because the kettle is not being invented.
Editing is also usually the faster route when you are close. A generation that is 90% correct is tempting to fix by regenerating, and regenerating throws away the 90%. Every fresh generation is a new roll: you may lose the composition, the light and the mood you liked in exchange for fixing one object in the corner. Editing changes the corner and leaves everything else where it was.
Generation earns its place when the subject is unconstrained or does not exist yet. Concept work, backgrounds, atmospheric shots, illustrations for a script, first drafts of an idea nobody has photographed. These are all faster to describe than to source, and five variations to choose between is an afternoon where five photo shoots is a month.
In practice the two are used in sequence rather than as alternatives. Generate a frame, edit the one thing that came back wrong, upscale it, then animate it. Each step is doing what it is good at, and each one is reversible, which is more than can be said for regenerating from scratch and hoping the good version comes back.
It is also worth naming the traditional option honestly. A masked selection in a conventional editor is still more precise when you need pixel-exact control over a boundary, and manual retouching remains the right answer for work that must be defensible. Prompt-based editing wins on speed and on the edits that would otherwise take real skill: relighting a scene, replacing a background convincingly, rebuilding a cluttered corner. It does not win on every task by default.
Getting better results
Name the object, not the feeling
"Remove the blue recycling bin on the left" is actionable. "Clean this up" invites the model to decide what offends it, and it will often decide something you wanted to keep. Point at one element and say what should happen to it.
One change per pass
A single instruction that changes the sky, removes two objects and warms the colour will get you a muddled attempt at all three. Stack edits instead. Each pass gives you something to accept or reject, and you always know which instruction caused which result.
Say what must not change
Where an element matters, protect it in the wording: "replace the background, keep the product and its label exactly as they are". It costs a clause and it meaningfully reduces the chance of the model helpfully redesigning something you never mentioned.
Give the light somewhere to come from
The most common tell in a background replacement is a subject lit from a direction the new scene cannot explain. Describe the new light to match the original: "overcast light from the left, soft shadows". Otherwise the composite will read as one.
Start from the best version you have
An edit inherits the source. A screenshot of a compressed thumbnail cannot become a clean image because you asked politely; the detail is gone. Upload the original file, at the largest size you have, before any resizing or re-encoding.
Inspect the parts you did not ask about
Check hands, text, logos, teeth, patterned fabric, straight architectural lines and anything reflective. These are where edits drift, and they are exactly the details a viewer notices without being able to say why.
Edit before you animate, not after
If the still is heading for image-to-video, fix it now. Once a clip is generated, an unwanted object is in every frame of it, and cleaning that up is a much larger job than cleaning one picture.
Keep the original
Edits stack, and stacked edits accumulate their own softness and drift. Keeping the untouched file means you can restart from a clean source when the fourth pass goes somewhere you did not intend.
Models available
- Nano Banana 2 — Google's best stills. Generation and editing, to 4K.
- Kling 3 — Cinematic motion that holds a character across a cut. For animating an edited still.
- Seedance 2.0 — ByteDance, up to 24 seconds. The long shots.
- Grok Video — xAI. Fast image-to-video, and text-to-video.
Questions
- How is this different from the image generator?
- The generator invents a new picture from your description. The editor starts from a picture you already have and changes part of it, leaving the rest as it was. If a specific real object or person has to appear in the result, you want the editor.
- Will the rest of the picture stay the same?
- That is what the tool is for, and it is what you should check first. Look at faces, hands, text, logos and reflections in the areas you did not mention. Those are where an edit shows itself. If something drifted, restate the instruction and name what must be preserved.
- Does editing reduce the resolution?
- Nano Banana 2 covers editing as well as generation, up to 4K, so an edit does not have to mean stepping down in size. Start from the largest original you have. The edit can only work with the detail you upload.
- Can the edit change the shape of the image?
- There is an aspect ratio control beside the instruction, and it defaults to matching the image you uploaded. Leave it alone when the framing is already right; change it when the picture has to work in a slot it was never shot for, and say in the instruction what should stay in view.
- What is this not good at?
- Pixel-exact boundaries and work that has to be defensible. A masked selection in a conventional editor is still more precise when a specific edge matters, and manual retouching remains the right answer where the edit has to be accounted for. Describing the change wins on speed and on the edits that would otherwise take real craft.
- Can I make several edits to the same image?
- Yes, and stacking single changes works better than describing all of them at once. Keep the untouched original so you can restart cleanly if a later pass goes somewhere you did not want.
- Can I edit an image I generated here?
- Yes. It is a common sequence: generate a frame, fix the one detail that came back wrong, then upscale or animate it. Regenerating instead would throw away the composition you already liked.
- What happens to the image after I edit it?
- It goes to your library, where it can be upscaled, turned into a video clip, or edited into a sequence with a script, voiceover and captions. Finished videos can be published to YouTube, TikTok and Instagram on a schedule.
- Can I use the edited images commercially?
- What you make is yours. Check the terms for the specifics, and note that the rights in whatever you upload are still your responsibility.
AI Image Editor
Upload a picture and describe the change you want. This is editing, not generating: the photograph you started with is still the photograph you end up with, minus the thing you asked to remove or plus the thing you asked for.
Start creating