How to remove the background from an image free (and without uploading it)
Updated on · Reading time: 7 minutes · By Gabriel Maglia
Removing the background from a photo is probably the most repeated edit in
the world: for an online store, a WhatsApp sticker, a profile photo or an
app icon. Most popular tools solve it by uploading your image to a server,
processing it there and sending back the result — which means your photo
travels, is stored who-knows-how-long and is usually limited by a paywall.
There's another way: process it right in your browser,
without the file leaving your device.
BeforeThe original image, with its background.AfterNo background, as a transparent PNG.
The three methods (and when to use each)
1. AI segmentation: the fast path
A neural network trained to separate "subject" and "background" analyzes
the image and generates a mask automatically. It's the best starting
point for photos with people, pets or products on reasonably
distinguishable backgrounds. In PhotoCut Studio
this is the "AI Cutout" button (shortcut I):
the model downloads once and runs in your browser: on a computer with a
GPU the browser can use, it fetches BiRefNet-lite (~120 MB), which
cuts out finer detail; on phones and tablets, or without a GPU, it
fetches u2netp (~6 MB) and runs it with WebAssembly.
Once you've used it, it works even offline.
2. Magic wand: one click on the background
When the AI gets confused and the background is plain or a uniform color,
the Magic wand selects by color: you click the
background and that whole contiguous area of the same color is removed.
Shift-click adds other areas to remove, and the tolerance
sets how much color variation is accepted. It's deterministic and
predictable: if the result isn't right, you undo and adjust the
tolerance. If you only need to narrow down the area, "Crop area" makes a
rectangular crop (it doesn't separate the background).
3. Refine brushes: fine control
No automatic method is perfect at hard edges. That's why the right
workflow always ends with two brushes: keep (marks areas
the cutout removed by mistake) and remove (erases
leftover background). With zoom on the canvas you can work at pixel level.
The recommended workflow, step by step
Open the photo (or paste it with Ctrl/Cmd+V from the clipboard).
Try the AI Cutout first. On most photos it's already good.
If it fails, or the background is plain, use the Magic wand: one click on the background removes it (shift-click adds areas).
Turn on the before/after compare to spot leftovers.
Refine the edges with the keep/remove brushes (zoom with the wheel).
Adjust the edge feather (1–3 px is usually natural).
Export as a transparent PNG — or apply a color background, sticker or shadow.
The AI Cutout applied. In the left rail are the brushes and the compare tool mentioned in the steps.
Tricks for hair and fine edges
Resolution rules: cut out on the largest image you have and downscale later; the other way around loses detail.
Moderate feather: 1–3 px of feather makes the edge natural; more than 5 px looks "cut with blunt scissors".
Small brush on the hair: several passes with an 8–12 px brush beat one with a 40 px brush.
Check on light and dark backgrounds: light halos only show on a dark background, and vice versa. Use the white/black preview before exporting.
Which format to export?
Format
When
PNG
The standard with transparency: stickers, logos, icons, print.
WebP
Transparency with much smaller files — ideal for the web.
JPG
No transparency: only if you apply a color or image background.
Common mistakes when removing a background
Confusing "transparent" with "white". If you export as JPG (which doesn't support transparency) the gap is filled with solid white or black. To keep transparency you need PNG or WebP.
Not checking at real zoom. At preview size, background leftovers and halos don't show; zoom to 100–200% before calling the cutout done.
Cutting out on an already-reduced photo. If the image comes compressed or low-res, the algorithm has less contrast information to decide the edge. Always start from the largest version available.
Overusing the "remove" brush. Running it too close to the subject eats parts that do belong to the figure (fingers, strands of hair). Several soft passes beat one aggressive one.
On your phone (Android and iPhone)
Most photos we want to cut out live on the phone, and you don't need an
app for it: PhotoCut Studio runs in the phone's browser (Chrome, Safari,
Samsung Internet…), takes no storage space and processes the photo on
the device itself.
Tap to pick a photo from the gallery or take one with the camera on the spot.
Apply the AI Cutout: it separates the subject from the background right there.
Touch up with your finger anything left over or missing, zoomed in on fine edges (hair, fingers).
Keep the background transparent or add a color, and download the PNG to your gallery.
Three things that change on a small screen:
Turn the phone landscape for more room when touching up edges, and always use zoom: your finger covers exactly the detail you want to see.
Small brush on the outlines. A big brush, on a finger, erases more than you want.
Huge photos and little memory. Phones shoot many-megapixel photos; on a phone with little memory a very large one can take a few extra seconds. If it stutters, try a slightly smaller version: for social, profile or stickers you don't need the maximum resolution.
The cutout doesn't go anywhere else: you stay on the same phone to give
it a new background,
keep it transparent for a
sticker or
export a round avatar.
Since processing is local, mobile data use is limited to loading the site
(and the AI model) the first time.
Genuinely free: no watermark, no sign-up, no limit
"Free" usually comes with fine print because every cutout on a server
costs the provider compute and bandwidth, and that cost resurfaces as a
watermark you only remove by paying, a daily quota of cutouts, reduced
resolution on the free plan, or a mandatory account to sell you a
subscription later. When the cutout runs in your browser there's no
per-image cost, so none of that applies here:
No watermark: PNG, WebP and JPG come out clean, with no logo on top.
No sign-up: no email or personal data; you open the page, add the photo and download.
No usage or resolution limit: the same flow (AI Cutout, brushes, export) repeats photo after photo, handy for a product catalog or a sticker pack.
Commercial use: the images you cut out and export are yours; use them in your store, social media or any project with no license restrictions on the result.
To spot a "free" tool that isn't, try it first without giving any data
and check whether the full-resolution, watermark-free export sits behind
a plan.
Why does it matter that the image isn't uploaded?
ID documents, photos of minors, unreleased products, client material under
NDA… a good part of what gets cut out daily is sensitive information. If
the tool processes the image in your browser, that question disappears: the
file never leaves your device, there's no server to hack or retention
policy to read. That's the underlying technical difference between PhotoCut
Studio and cloud cutout services.
Frequently asked questions
Is it really free? Does it add a watermark or limit usage?
Yes, it's genuinely free: the cutout, the brush refinement and PNG/WebP/JPG export have no cost, no watermark and no usage limit. It also works the same on Android and iPhone, because it runs in the browser.
We tested it: the photo doesn't leave the browser, and the model takes as long as it takes
I opened /editor/ in a tab with an empty cache and loaded the photo with “Open photo”.
I pressed “AI cutout” and timed it from the click to the “AI cutout applied” notice.
I downloaded the transparent PNG with “Download”.
I ran the AI cutout again in the same tab and after reloading the page, with the model already in the browser cache.
I logged every network request of the session.
What I measured:
AI cutout, first time (with the model downloaded from my local server): 22.6 s. In the same tab, second run: 3.9 s. After reloading, with the model in the cache: 16.1 s.
This computer has a GPU, so the AI downloaded BiRefNet-lite: 114.8 MB of model (in 2 parts) + 24.0 MB of WebAssembly/WebGPU engine, uncompressed. The model travels as is; the engine with Brotli (level 11, computed on those same files from dist/) drops to 3.5 MB. In total 118.4 MB, the “~120 MB” the guide mentions for a device with a GPU.
Exported PNG: 720 × 900 px, 532 KB, 47.7% of pixels fully transparent and 3.3% semi-transparent (the smoothed edge).
Network: 93 requests in the whole session and none with a body (0 bytes sent). The 26 requests to other domains (pagead2.googlesyndication.com, googleads.g.doubleclick.net, ep1.adtrafficquality.google, ep2.adtrafficquality.google, www.google.com) are Google advertising and don't carry the photo.
What went wrong or what limit I saw: With this portrait the AI left no fringe I could see on a magenta background, so I can't describe a hard case. I measured on a Mac with a GPU, so BiRefNet did the cutout: the first time took 22.6 s because, besides downloading the model (from my machine; over the real internet that's about 120 MB), the browser prepares the GPU, and after reloading it took 16.1 s again for that preparation even though the model was already cached. Without a GPU u2netp does the cutout and downloads about 6 MB: with it I previously measured 5.0 s the first time and 2.9 s the second. Each timing is from a single run. In another run of the same test I saw four Google ad-measurement POSTs (up to 435 bytes, none with the photo): the ads make requests, but the image doesn't travel.
Tested by Gabriel Maglia on with PhotoCut (web) in Chrome 155.0.8059.80 · macOS 14.8.8 · PhotoCut version: commit 7ec5265. AI cutout with BiRefNet-lite on the GPU (WebGPU).