Alpha is a channel, not a hole
The colours are still in the file. What is added is an opacity value per pixel, which is why a transparent PNG is not smaller than an opaque one and why a viewer that ignores alpha shows white.
Transparent background
A PNG or WebP with a real alpha channel, cut by a model that runs on this device. The checkerboard in the preview is what see-through looks like.
The short answer
To make an image background transparent, add the photo and press the button. A matting model runs on your device, works out how much each pixel belongs to the subject, and writes that as the alpha channel of a PNG or a WebP. Every part that was background becomes see-through rather than white, so the file sits correctly on any colour you place it on later. The photo is not uploaded and the export keeps its full resolution.
Drop a photo, click to browse, or paste one from the clipboard. The first run downloads the model and counts the megabytes as they arrive; after that it is cached and the cut starts immediately. This tool only writes formats that can hold transparency, so the result cannot come back opaque by accident.
Presets
Cut-out quality
Save as
Both of these carry a real alpha channel. PNG is lossless and larger; WebP is smaller and every current browser reads it.
AI model
Checking what this device can run.
Checking what this device can do. Nothing is uploaded.
Add a photo and run the tool. The download appears here.
Try another tool on the same photo
The short version
The phrase “transparent background” hides a technical fact worth understanding, because almost every complaint about these files comes from not knowing it. Transparency is a channel, not an absence. The file still stores a colour for every pixel that used to be background; what changes is that each of those pixels now also carries an opacity of zero, and any program that draws the picture is expected to honour it.
That expectation is where things go wrong. A viewer that ignores alpha shows white. A printer that ignores alpha prints white. A GIF export flattens the whole channel to one bit and the edges go jagged. None of those are the cut-out being wrong, and all of them are avoidable once you know that the format, and the thing opening it, both have to agree that alpha exists.
The colours are still in the file. What is added is an opacity value per pixel, which is why a transparent PNG is not smaller than an opaque one and why a viewer that ignores alpha shows white.
PNG and WebP hold real alpha. GIF holds a one-bit version, which is why its cut-outs look stepped. JPG holds none, so a cut-out saved as JPG is a cut-out on a colour.
The model returns partial values along the outline, and those partial pixels are what make hair read as hair. They are kept, never rounded, which is the difference between a cut-out and a sticker.
Most operating system viewers draw transparency as white. Before deciding the file is wrong, drop it on a dark web page or into a design tool and look again.
It is cached by the browser and reused by every background tool on this site, so the second file you cut out starts immediately and works with the connection off.
Clear glass is partly transparent and partly refracting what is behind it. One alpha channel cannot say both things, so glassware needs hand work whatever tool produced the mask.
Step by step
Four steps, all inside this page, with the photo read from disk into this tab and nowhere else.
Drag a file onto the panel, click it to browse, or paste an image from the clipboard. The tool reads the file, applies any camera rotation, and shows its real dimensions.
The panel counts the megabytes as the model arrives, then spends a few seconds setting it up on your graphics card. That step is silent and it is not stuck. It happens once per device.
Leave the softness at one pixel for a photograph, or take it to zero for a logo or a screenshot. Press the button and the model finds the subject and writes the mask into the alpha channel.
The checks show through wherever the file is see-through. Look at the hair, the thin parts and the gaps that should be open, adjust and run again if needed, then download the PNG or WebP.
PNG if something downstream is old or fussy, WebP if the size matters. PNG is lossless, which means it is exactly the pixels the tool produced, and it is the format every piece of software on earth has read for twenty years. The cost is weight: a photograph as a PNG is routinely several times the JPG it came from.
WebP with alpha is the same cut-out at a fraction of the bytes, and browser support has been universal for years. The reason to hesitate is not browsers, it is everything else: some print workflows, some older design tools and some upload forms still refuse it. When you do not know where the file is going, PNG is the answer that never needs a second attempt.
If you are converting between them later, the converter keeps alpha intact in both directions and tells you when a target format cannot.
Because a cut-out drawn on a white card is indistinguishable from a photograph of a white wall, and you would have no way to tell whether the tool did anything. The checkerboard is the convention every image editor uses for “nothing is here”, and it is drawn by this page rather than being part of the file.
It is also the fastest quality check available. Zoom into the hair and look at whether the checks come through in the gaps between strands. If they do, the mask found real detail. If the hair is a solid silhouette with checks only outside it, the model treated it as a block, and the detailed quality setting is worth trying.
Hair against a busy background of a similar tone. This is the genuine limit of the technology rather than of this implementation. The detailed model is much better at it than the fast one and neither is perfect.
Anything you can see through. Glass, water, smoke and thin fabric are partly transparent and also bend what is behind them. A single alpha channel can express the first half of that and not the second, so the result looks flat.
Fine repeating gaps. Chain-link, wire mesh, lace and bicycle spokes lose their small holes, because the mask resolution is finite and the holes are smaller than it. Zooming out, the result usually reads fine; zoomed in, the mesh is filled.
Motion blur. A blurred hand or a moving ball has no crisp boundary to find, so the model picks one. Adding edge softness helps the result sit more naturally on its new background.
| Property | Detail |
|---|---|
| Input formats | JPG, PNG, WebP, AVIF, GIF, BMP. HEIC where the browser can decode it. |
| Output formats | PNG lossless, or WebP with alpha. Both carry a genuine alpha channel. |
| Not offered | JPG and GIF. One has no alpha, the other has one bit of it. |
| Model, detailed | BiRefNet lite, fp16, about 115 MB, MIT licensed. |
| Model, fast | MODNet, fp16, about 13 MB, Apache 2.0 licensed. |
| Download | Once per device, cached by the browser and shared with the other background tools. |
| Backends | WebGPU where available, single-threaded WebAssembly everywhere else. |
| Edge control | 0 to 8 pixels of feather over a continuous mask, never a threshold. |
| Resolution | Full. The mask is upsampled and applied to the untouched original. |
| Where it runs | In this browser tab, on your device, in a background thread. |
Slide the table sideways to see every column.
Worth avoiding
JPG cannot store alpha. The transparent area becomes a colour, and this tool does not offer JPG at all for exactly that reason. Use the background remover if a flat colour is what you want.
Most previews draw transparency as white. Open it over a dark background in a browser tab before concluding the cut-out failed.
GIF transparency is one bit, so every edge pixel is either fully opaque or fully gone. That is what produces the stair-stepped outlines people associate with cheap cut-outs.
Feather blends an edge; it does not find detail the model missed. If the hair came out as a silhouette, change the quality setting and run it again rather than blurring the problem.
Work on the largest version you have. A mask computed on a small copy and then enlarged loses precisely the edge detail that makes the cut-out worth doing.
Plenty of forms flatten PNGs onto white on the way in. If the destination is not a design tool, check what the uploaded version looks like before building anything on top of it.
Questions
Every answer is in the page source, so a reader and an answer engine get the same words.
A fourth channel. A normal photo stores red, green and blue for every pixel; a transparent one adds alpha, which says how opaque that pixel is. Nothing is deleted, so the file is not smaller for having a transparent background, and the transparency only appears when something else is drawn behind it. The checkerboard you see in the preview is this page drawing that something, not part of the picture.
PNG and WebP, which are the two this tool writes. AVIF also supports alpha but very few browsers can encode one from a canvas, so it is not offered here. GIF has a one-bit version of transparency, meaning a pixel is either fully there or fully gone, which is why GIF cut-outs have jagged edges. JPG has no alpha at all: saving a cut-out as JPG replaces the transparency with a colour.
Because PNG is lossless and JPG is not. A photograph re-encoded as PNG is routinely three to five times the size of the JPG it came from, and the alpha channel adds a little on top. If the size matters, choose WebP with alpha instead: it is usually 25 to 35 percent of the PNG and every current browser reads it.
Almost certainly not. Most image viewers, and most operating system previews, draw transparency as white or as a checkerboard, and some print white. Drop the file onto a coloured slide, a dark web page or a design tool and the transparency will show. If you need to be certain, open it in a browser tab over a dark background.
It widens the band of partly transparent pixels around the subject. Zero cuts on the model's own edge, which is the right answer for a logo or a screenshot. One or two pixels blends the cut-out into whatever it eventually lands on, which is what hair, fur and fabric need. It is never a threshold: the mask keeps its intermediate values either way, which is what stops the result looking like a sticker.
Your device sets the ceiling rather than a plan does. The model looks at a bounded copy of the picture because it resizes internally regardless, so a very large photo does not need a very large amount of memory for the inference itself. The composite happens at full size, and browsers cap a drawing surface at 16384 pixels an edge, which is well past anything a camera produces.
No. The first run downloads the AI model from a content delivery network, and that is the only network traffic the tool ever causes. The photo is read from disk into this tab, the model runs in a background thread in the same tab, and the result is written back out as a download. After the model is cached the tool works with the connection off.
The same model and the same cut. This page will only write formats that carry an alpha channel, so a file cannot come out opaque by accident. The background remover also offers a solid colour, because most people arriving there are heading for a marketplace that wants plain white.
Put the subject on a colour or a gradient, blur the room behind it instead of removing it, or resize and compress the finished file for wherever it is going. All of it runs on your device.