How to Remove a Background from a Photo (and Get Clean Edges)

How AI background removal actually works

Automated background removal has gotten very good, but it isn't magic. A segmentation model looks at the whole image and predicts, pixel by pixel, how likely each one is to belong to the main subject. The output is a mask: a grayscale map where white means "keep", black means "remove", and the grays in between mean "partly see-through". That mask becomes the image's alpha (transparency) channel.

The partial values are what make a cutout look natural. A strand of hair or the soft edge of an out-of-focus shoulder is genuinely a blend of subject and background, so a good mask keeps it semi-transparent rather than forcing a hard yes/no line. When people say a cutout looks "pasted on", it is usually because those in-between edge pixels were handled badly.

The NextG Design Background Remover runs this model directly in your browser using WebAssembly. The first time you use it, the browser downloads the model; after that your photo is processed on your own device and never uploaded anywhere. That makes it suitable for personal photos, ID documents and client work you can't send to a third-party server.

Start with a photo that gives the model a fair chance

Models do best when there is real contrast between the subject and the background — different colors, different brightness, and ideally a little separation in focus. A subject wearing a color close to the background, or standing right up against a cluttered, similarly lit backdrop, is the single biggest cause of a rough cutout. If you're taking the photo yourself:

• Leave space behind the subject. A metre or two of distance from the wall reduces shadows on the background and adds natural blur.
• Pick a contrasting backdrop. Dark clothing in front of a light wall, or a light product on a dark table, gives the model a strong edge to follow.
• Use soft, even light. Hard shadows cast onto the background can be mistaken for part of the subject.
• Keep the subject sharp. Motion blur smears the boundary across several pixels and makes the true edge ambiguous.
• Don't crop the subject too tightly. A little margin around the edges helps the model understand what the object is.

Resolution matters less than people expect. A sharp 1,500-pixel photo with good contrast will cut out better than a blurry 6,000-pixel one.

Know where AI cutouts usually struggle

• Fine hair and fur. Individual strands are thinner than the model's effective resolution, so hair edges are the most common place for slight haloing or over-smoothing.
• Semi-transparent or reflective material. Glass, sheer fabric, veils and smoke have no single clean edge, so the model has to pick one — and it won't always pick the one you'd want.
• Low-contrast boundaries. A dark jacket against a dark background, or pale skin against a light wall, gives the model the least information to work with.
• Small gaps and holes. The space between an arm and the body, or inside the handle of a mug, can be filled in or left transparent inconsistently.
• Multiple subjects. In a group shot the model has to decide what "the subject" is; people at the edges or partly hidden are sometimes dropped.
• Objects the subject is holding or touching. A chair someone is sitting on, or a bag on their shoulder, may be treated as background.

Step by step: removing a background in your browser

1. Open the Background Remover and drop in a PNG, JPG or WebP photo.
2. Click Remove Background. The first run takes a little longer while the model loads; later images are faster.
3. Zoom in on the edges — hair, fingers, the gap under an arm — and look for leftovers or missing pieces.
4. Use Erase to remove background that leaked into the subject and Restore to bring back parts that were cut away. Undo is available if a stroke goes wrong.
5. Click Review & Export, check the cutout on both light and dark previews, optionally upscale it, and download the transparent PNG.

Use the touch-up brushes instead of starting over

Rather than accepting an imperfect result or trying a dozen different source photos, correct the specific areas the model got wrong, directly on top of its first pass:

• Erase removes background pixels that leaked into your subject — useful for stray bits near hair or busy edges.
• Restore brings back subject pixels the model mistakenly cut away — thin details like glasses arms, wisps of hair, or a clipped finger.
• Zoom in before brushing. Edge mistakes are usually only a few pixels wide; it's far easier to be precise at high zoom than to guess at full size.
• Shrink the brush near edges. Use a large brush to clear big areas and a small one close to the boundary.
• Work in short strokes. It's easier to stop exactly where you meant to, and a single undo then removes only a small mistake.

Two minutes of touch-up on the trouble spots is almost always faster than hunting for a "better" photo.

Check your export, not just the preview

Before downloading, view the result against both light and dark backgrounds. A cutout that looks fine on white can reveal a faint pale halo once it's placed on a dark or colored background — and a cutout checked only on black can hide dark fringes that appear on white. The halo is leftover color from the original background mixed into the semi-transparent edge pixels. A quick Erase pass around the affected edge usually clears it.

Also think about where the image is going next:

• Layering in a design, slide or website? Keep it as a transparent PNG. Our PNG vs JPG guide explains why JPG will lose the transparency.
• Need a white background for a marketplace listing? Convert the PNG with the PNG to JPG converter, which fills transparent areas with a solid color. See our product photo guide for marketplace rules.
• Small source image? Use the upscale option in Review & Export, or the standalone Image Upscaler.
• File too large? The Image Compressor can export WebP, which keeps transparency at a much smaller size.

Frequently asked questions