How to remove the background from a photo
Cutting a subject out used to mean a lasso tool and twenty minutes in Photoshop. A model can now do most of that work in a couple of seconds, drop a photo in and you get the subject back on a transparent background. Here's how it actually works, and how to get a result you're happy with.
What background removal actually does
The tool looks at a photo, decides which pixels belong to the main subject and which belong to everything behind it, then makes every background pixel transparent. What's left is the subject alone, on a checkered transparent canvas until you place it against something else.
That's the whole job, but it's a harder job than it sounds. A person's outline is straightforward for a model to find. The edge of that outline, hair against a bright sky, a sleeve the same colour as the wall behind it, is where most of the real difficulty lives, and it's the part worth understanding before you judge a result as good or bad.
How the automatic cutout works
This tool runs a segmentation model, a neural network trained specifically to separate a subject from its background, entirely inside your browser tab using WebAssembly. Drop a file in and the model outputs a mask: for every pixel, a value saying how much it belongs to the subject rather than the background.
That mask isn't just on or off. Edge pixels, hair, fur, a blurred motion edge, get a value somewhere in between, which is what lets the final cutout have a soft, natural edge instead of a hard jagged line stamped straight through a strand of hair.
The mask gets scaled back up to your photo's full resolution and applied as the image's alpha channel. Everything the model was confident is background becomes fully transparent, everything it was confident is subject stays fully visible, and the in-between pixels along the edge get partial transparency, which is what makes a good cutout look clean rather than cut-out.
Doing it in your browser
Open the background removal tool, drop a photo in, and wait a few seconds. The first time you use it, the browser downloads the segmentation model itself, roughly 80MB, which is a one-time cost. After that it's cached, and later photos process without downloading anything again.
There's no setting to fiddle with beyond the output format. This isn't a tool with sliders for edge feathering or subject detection sensitivity, it runs the model and gives you the result. If the cutout isn't clean, the fix usually isn't in a setting, it's in the photo, more on that below.
Batch works the same way it does for every other tool on this site: drop more than one file and each gets processed in turn, downloaded together as a single ZIP when they're all done.
Why the output is PNG or WebP, never JPG
Transparency needs an alpha channel, a place in the file format to store 'how see-through is this pixel', and JPG has never had one. Every pixel in a JPG is fully opaque by definition, there's no way to represent a transparent background in the format at all.
So this tool only ever offers PNG or WebP as output. It's not a missing feature, it's the format doing the one thing a background-removal tool actually needs, and offering JPG here would mean silently handing back a photo with a solid colour where the transparency was supposed to be, which defeats the entire point of running the tool in the first place.
Why nothing gets uploaded to do this
The usual way a background-removal tool works is to send your photo to a server, run the model there, and send the result back. That's a reasonable way to build one, and it's also a copy of your photo leaving your device and landing on hardware you don't control.
This one runs the model on your own machine instead. The 80MB you download once is the model itself, not a connection to a server, and once it's loaded, removing a background works with no network request at all. You can check that directly: load the page, disconnect from the internet, and run a photo through it anyway. It still works.
That matters more for this specific tool than for most. A background-removal tool is often used on photos of people, staff headshots, family photos, ID-style shots for a form, exactly the kind of image where a silent upload step is the last thing you want.
What makes a clean cutout, and what doesn't
The model does best with a subject that stands out clearly from its background, a person against a plain wall, a product on a table, an object with a defined edge. Even lighting across the subject and reasonable contrast between subject and background give the model the clearest signal to work from.
It struggles more with a subject that blends into its background in colour or brightness, a person in a grey jumper against a grey wall, a dark subject on a dark background. It also struggles with motion blur, since the model is looking for an edge and a blurred edge is, by definition, not a clean line.
None of this means the tool fails outright in those cases, it usually still produces something usable, but the edge is where the difference between a great cutout and an adequate one actually shows up.
Hair, fur and other hard edges
Fine detail along an edge, hair, fur, the fuzzy edge of a sweater, is the single hardest thing for any background-removal system, automatic or manual. A strand of hair is a handful of pixels wide, semi-transparent where light passes through it, and different from the strand next to it, which is a lot to ask any model to get pixel-perfect.
The mask this tool produces handles that with partial transparency rather than an all-or-nothing cut, so a soft, layered edge on hair generally comes through recognisably rather than as a hard silhouette. It won't be flawless on every strand, no automatic tool's is, but it's built to fail gracefully at the edge rather than produce a harsh outline.
If the hair or fur detail in a specific result matters a lot, a bright, evenly lit background behind the subject when the photo was taken makes a bigger difference to the final edge than anything the tool itself can do after the fact.
Fixing a bad cutout without starting over
If a result comes back with a background element you didn't expect, a chunk of the wall next to a shoulder, a shadow left in, or the reverse, part of the subject too faint and made partially transparent, the fastest fix is usually a different source photo rather than trying to coax a better result from the same one.
A plainer, more evenly lit background, moving the subject a step or two away from anything the same colour as them, or retaking the photo with the subject better separated from what's behind them, tends to fix more than re-running the same image ever will. The model isn't guessing at random, it's reading exactly what's in the photo, so a cleaner photo in is what produces a cleaner cutout out.
Common uses for a transparent-background photo
A product photo cut out onto transparency drops cleanly into a listing template or a catalogue page without a background colour clashing with the page it lands on. A staff photo or headshot with the background removed goes onto a team page, a slide, or a name badge without needing to be shot against the exact colour that page happens to use.
A subject cut out from one photo can be composited into a different scene entirely, a product against a lifestyle background, a person into a group graphic, without a visible seam if the lighting is reasonably close between the two. And for print or merchandise, a transparent PNG is usually the exact input format a print shop or a t-shirt printer's upload tool is asking for.
Batch removing backgrounds
The batch limit here is the same 200 files per run every other tool on this site uses, and it works the same way, drop a folder in, everything processes one after another so an ordinary laptop doesn't run out of memory partway through, and you download the lot as one ZIP when it's finished.
Because the model itself is a fixed 80MB download that only happens once per browser session, running a large batch doesn't cost you that download again for every file, it's paid once and then every subsequent photo, in this batch or a later one in the same session, reuses it.
Transparent background, not a coloured one
The output here is transparency, a checkerboard until you place it somewhere, not a white or coloured background painted in behind the subject. That's a deliberate scope decision rather than a missing option: transparency is the more useful, more general output, since you can drop a transparent PNG onto any background colour afterwards, but you can't easily undo a colour that's already been baked in if it turns out to be the wrong one.
If what you actually need is the subject on plain white, most places that ask for that, marketplace listing photos are the common example, will accept a transparent PNG directly and treat it exactly like a white background themselves. Where a tool insists on a flattened image with no transparency at all, opening the transparent PNG in any editor and placing a white rectangle behind it takes a few seconds, one direction is easy, the other isn't.
How photo resolution affects processing
Very large photos, a full-resolution shot straight off a modern phone or camera, get resized down before the segmentation model actually looks at them. Running the model at full resolution on a large image would ask more of your browser's memory than a WebAssembly model reasonably should, and a segmentation model doesn't need every one of twenty-plus megapixels to find an edge, it needs enough resolution to see where the subject stops and the background starts.
The mask the model produces is then scaled back up to your photo's actual resolution before it's applied, so the final cutout keeps your original pixel dimensions and detail everywhere except right at that edge, which is where the model's own precision was the limiting factor anyway, not the resizing step.
PNG vs WebP: which output to pick
PNG is the safer default and what most software expects when it asks for 'an image with transparency'. It's lossless, so nothing about the subject itself is degraded by the format, and it's supported everywhere, every editor, every upload form, every version of every operating system.
WebP with transparency does the same job at a meaningfully smaller file size, and every current browser reads it without issue. Pick it if you're publishing the cutout to a website you control and the smaller file matters. Pick PNG if you're not sure what the destination expects, or if it's headed somewhere older or more particular about file types, a print shop, an older editor, a form that only lists specific formats.
How this compares to other background removers
Most background-removal tools you'll find, remove.bg among the best known, work by uploading your photo to their server and sending back a result, often with a daily limit on the free tier, a resolution cap unless you pay, or both. That's a fine trade for a lot of people, but it does mean your photo leaves your device to get processed.
This tool trades a one-time 80MB model download for never sending a photo anywhere, no daily cap, no watermark, and no account. The cutout quality it produces is comparable for an ordinary photo of a person or product, the actual difference between tools in this category is almost entirely in how they handle your data and what they charge, not in wildly different segmentation quality on an easy subject.
What this tool isn't built for
This is a single-image tool, not a video background remover, removing a background from a video needs the mask recalculated for every frame, a genuinely different and much heavier job that a browser tab isn't the right place to do.
It's also not built for extremely fine, deliberate manual masking, the kind of frame-by-frame hand correction a professional retoucher does on a hero image for a magazine cover. For that level of control you want a proper editor with a manual selection tool. For an ordinary product photo, headshot, or a subject you want cut out for a slide or a listing, the automatic result is usually exactly what you need without any of that manual work.
Quick questions
Is background removal really free?
Yes, no account, no daily limit, no watermark. The only cost is a one-time ~80MB model download the first time you use it in a browser session.
Are my photos uploaded to a server?
No. The model runs entirely in your browser. Load the page, disconnect from the internet, and it still works.
Why can't I export as JPG?
JPG has no alpha channel, so it can't store transparency at all. The tool only offers PNG or WebP, both of which support it.
Why is the edge around the hair not perfect?
Fine hair and fur are the hardest thing for any segmentation model, automatic or manual, to separate pixel-perfectly. The tool blends the edge with partial transparency rather than a hard cut, which usually looks better than it sounds but won't be flawless on every strand.
Can I remove backgrounds from a whole folder at once?
Yes, up to 200 photos per batch, downloaded together as one ZIP.
Will it work on a photo where the subject is a similar colour to the background?
It will usually still produce a result, but the edge quality drops when subject and background have similar colour or brightness. A more evenly lit, higher-contrast photo gives a cleaner cutout.
Does it work on product photos, not just people?
Yes. The model isn't limited to people, it works on any distinct subject against a background, products, objects, animals.
How is this different from remove.bg?
The main difference is where the processing happens. This tool never uploads your photo; most alternatives, remove.bg included, process it on their own servers.