AI Image Editor — Edit Any Photo With One Sentence
Describe the change in plain English. Get the photo back, edited. No layers, no masks, no account.
1 — Your photo
2 — What should change
2 free edits a day, no signup.
Nano Banana (Gemini image) · 1024 × 1024 · about 15 s per edit
What an instruction-based AI image editor is
Most editors ask you to operate the image: select an area, build a mask, clone a patch, adjust a curve. This one asks you to describe an outcome. You supply a photo and one sentence — make the sky overcast, remove the bin behind the car — and a generative model returns a new image that matches the description.
That is a different trade-off from Photoshop, and it is worth being precise about it. A layer-based editor gives you exact control and expects you to know where every edge is. An instruction-based editor gives you speed, handles regions you could never mask by hand, and has no idea what "exactly" means. It is the right tool when the edit is expressed in words and the result can tolerate a little interpretation.
What you can change with one sentence
- Remove something. "Remove the stranger in the background and fill the area with the same paving."
- Change a colour or material. "Make the jacket navy blue, keep the face and pose unchanged."
- Replace the background. Plain walls, studio backdrops, gradients — the strongest category for this kind of model.
- Clean up a product photo. Dust, scratches and small blemishes go; label text often survives if you ask for it explicitly.
- Restyle an existing photo. Film tone, flat illustration, overcast light — the composition stays, the treatment changes.
- Fix an obvious distraction. Power lines, a stray cable, a mark on a wall.
Each of these is a single pass. There is no undo stack, so keep the original file — the fastest way to compare two instructions is to run both and look.
How to edit a photo with AI in three steps
1. Upload the original
Drop a JPEG, PNG or WebP up to 6 MB. Use the largest version you have: the model works on a 1024 × 1024 canvas, so a small source loses detail before it starts.
2. Write the instruction like a brief, not a prompt
The sentence that works best names three things: what to change, what it should become, and what must not move. Compare "make it better" with "remove the bin on the left and fill the area with the brick wall, keep the car unchanged". The second one gives the model a target and a boundary.
3. Compare, then re-run
Results vary between runs. If an edit comes back soft or misaligned, change one clause rather than rewriting the whole sentence — that is how you find out which part the model was ignoring.
Where instruction-based editing still fails
Being honest about the limits saves more time than a longer feature list:
- Text. Long words, small labels and logos are frequently garbled. Ask for text to stay sharp and inspect it at 100%.
- Hands, faces and fine structure. Small faces and fingers degrade fastest. If a person is the subject, check them first.
- Precise geometry. "Move this 20 pixels left" is not an instruction this kind of model understands. Use a normal editor for measurement.
- Anything that must stay pixel-identical. Every pass re-renders the whole frame. If the rest of the image has to be untouched, mask it in a layer-based tool instead.
- Real people in misleading contexts. The model refuses these, and so do the terms of use.
Free edits and what Pro will be
The free tier is two edits per day, no account, with a small watermark in the corner of the output. Capacity is shared across the whole site and resets at 00:00 UTC; when the day's budget is spent the tool closes until the reset rather than quietly degrading.
Pro is not on sale yet. When it opens it will add a higher daily limit and no watermark. We are not taking money for a tool that is still this young — the pricing page has the current state.
What happens to your photo
This site keeps no copy. The image is read into memory, forwarded to the model provider to run the edit, and returned to your browser. Nothing is written to disk, an object store or a queue, and nothing survives the request.
One thing we will not overstate: the model provider processes the image on its own infrastructure and, like every API provider, may retain request data for a limited period under its abuse-monitoring policy. We cannot promise on their behalf, so we are not going to claim "never stored" end to end. What we can promise is what we control — we keep nothing. Full details are in the privacy notice.
Frequently asked questions
Is this AI image editor free?
Yes. Two edits per day, no account required. The free output is a WebP file with a small watermark in the corner.
Do I need to sign up?
No. There is no account system in this version. Free edits are counted per network address and reset at 00:00 UTC.
Are my photos stored?
Not by this site. The image is held in memory only long enough to send it to the model provider and return the result. The provider may retain request data briefly under its own abuse-monitoring policy; see the privacy notice.
How is the two-a-day limit counted without an account?
A random ID in a first-party cookie plus a one-way hash of your network address. Both counters are checked on every edit, so clearing cookies does not reset the allowance — the other counter still applies.
What formats and sizes are supported?
JPEG, PNG and WebP up to 6 MB in, and a 1024 × 1024 WebP out.
How do I remove an object from a photo?
Name the object and say what should replace it: "remove the person on the right and fill the area with the same paving". The model fills from surrounding pixels, so busy textures and straight architectural edges are where it struggles most.
Why was my edit rejected?
Edits that put a real person in a misleading situation, involve minors, are sexually explicit, or reproduce protected characters are refused by the model's safety filter. Rewording usually does not change that outcome.
Can I use the output commercially?
You keep responsibility for the image you upload and for how you use the result. Check the terms before publishing anything commercially.
Why is the result only 1024 × 1024?
That is the resolution this model renders at. Upscaling afterwards works, but it will not recover detail that the edit did not produce.