AI Background Remover
Remove the background from a photo and keep a transparent PNG, or set the cut-out on a solid color as a JPEG. It finds the subject itself, or you tap or circle the one you want.
The photo
Photo
Drop a photo here
PNG, JPEG or WebP · one picture at a time
—
—
Brush
Kept objects
Cut out
Add a photo and the subject comes away from the rest. You choose transparent or a flat color before you save.
Cutting one out
- Drop the photo into the Photo box, or press Take photo on a phone and shoot the thing where it already sits.
- Press Remove background. The cut lands on a checkerboard in the same frame, and Original | Cut out flips between the photo and the cut. If it kept the wrong thing, press Pick the subject and tap or circle what you want to keep: each one is kept as its own object, and a second tap drops it. With a keyboard, tab to the photo, steer the crosshair with the arrows and press Enter.
- Choose what goes behind. Transparent keeps the alpha for compositing; Solid color fills it with a color you type, pick from the swatch, or lift from the screen with the eyedropper where the browser has one.
- The first cut-out fetches about 18 MB, the model plus the runtime that runs it, and the browser keeps both, so the next photo starts at once. Picking fetches its own model the first time you pick.
- Tidy the edge with Add back and Erase, then save it with Download PNG, or Download JPEG when the background is a solid color.
About this tool
Removing the background from an image is usually one of two jobs. Either you want a transparent PNG to drop into a design, or you want the subject sitting on a flat color for a store listing. Both start the same way. Add the photo, press Remove background, and a small segmentation model named U2Netp works out where the subject ends and the rest begins. It traces that outline on a 320-pixel square copy of the picture, then a second pass fits the outline back onto the edges the full photo actually has, so the cut follows the fur or the collar rather than a blocky curve.
When the photo holds two things, or the model keeps the wrong one, tap the one you want on the photo or draw a loop around it; once a cut is made, press Pick the subject first. A second model reads the picture once, and from then on every tap or loop keeps one object and lists it under Kept objects. Once the cut is made, Add back and Erase brushes sit behind Undo and Redo, so a missing ear or a leftover speck of wall gets fixed on the spot. The badge above the photo says what each model costs to fetch on your device before you press anything.
A mug on a cluttered counter
Say you photographed a ceramic mug on a kitchen counter with a kettle behind it. Remove background will very likely keep the kettle as well, because it is the loudest object in the frame. So tap the mug before pressing anything and give the picker a few seconds to read the picture: the mug comes back outlined and listed under Kept objects, and Remove background now cuts only that. If the tap catches the saucer too, drop it with a second tap and draw a loop around the mug alone.
Now choose Solid color under Background and leave the color on #ffffff. What downloads is a JPEG of the mug on white, which is the plain look most marketplace listings ask for. If the inside of the handle kept a sliver of counter, press Add back, switch the brush to Erase, drag the Brush slider down and take it off before you save.
Good to know
- Fine detail is where a small fast model shows its size. Hair strands, fur tips, veils and motion blur come out fully in or fully out. The outline is traced on a 320-pixel square, or a 256-pixel one when you pick, and then fitted to the photo, so a bigger photo is not a sharper cut. What helps is contrast between the subject and what is behind it.
- A transparent cut saves as PNG, the format that carries an alpha channel. A cut on a solid color has nothing left to be transparent, so it saves as a JPEG instead, which is a fraction of the PNG's size.
Frequently asked questions
When should I press Pick the subject?
Any time the photo holds more than one thing worth keeping, or when Remove background keeps choosing the wrong one. On its own, the cut-out model looks for the single most salient object. Pick the subject asks you instead, and answers each tap or loop with an outline from a promptable model. The trade is about 38 MB fetched the first time and a few seconds while it reads the picture, after which every tap is near-instant.
It cut out the shirt instead of the whole person. What now?
A tap is genuinely ambiguous. Point at someone's chest and both the shirt and the person are honest readings, and the model shows the one it rates highest. Draw a loop around the whole figure instead: the loop tells the model the person is inside it and the grass around it is outside, and the answer is kept to the loop. Failing that, run the removal anyway and paint the head and legs back with Add back.
How do I get a plain white background for a product photo?
Choose Solid color under Background and leave the color at its default white. The subject is composited onto that fill and the download is a finished flat JPEG, no transparency involved. Any other color works the same way, so the same cut gives you a colored backdrop behind a headshot or a banner.
Why is the first run slower than the ones after it?
Because the model and the ONNX runtime have to arrive before anything can be cut. That is about 18 MB for the cut-out and about 38 MB more the first time you pick. Your browser caches both, so the second photo starts almost at once. After that, speed depends on your machine and how many pixels the photo has.
Which image formats can I feed it?
Whatever your browser can decode, which in practice covers JPEG, PNG, WebP, AVIF and GIF. The Photo box accepts any image file. Very large photos still work, they just take longer, and the result is written back out at the original pixel dimensions.
Missing something in AI Background Remover? Suggest a feature →
Tell someone who needs this.