Make a clean transparent PNG from any photo with AI — or instantly knock out a solid color. Then keep it transparent or add a new background. Everything runs in your browser.
Drop an image here or browse
JPG · PNG · WebP — your photo stays on your device
Click the original image to pick the exact background color to remove.
Find people, cars, pets and more, then keep only what you tap. The model downloads once.
🔒 The AI runs in your browser — your photos are never uploaded to a server.
AI background removal is powered by @imgly/background-removal, licensed under AGPL-3.0 — source & licenses.
Cut out profile photos and headshots for avatars, IDs and team pages.
Get clean transparent product shots for shops, listings and catalogs.
Drop in a white, black or brand-color background and download a ready-to-use PNG.
The three tabs are genuinely different approaches, and picking the right one matters more than any setting inside it.
| Mode | How it decides what to keep | Best for | Weak on |
|---|---|---|---|
| AI | A neural segmentation model that has learned what a foreground subject looks like | Almost anything — people, products, animals, objects on a messy background | Fine hair, fur, glass and anything semi-transparent |
| By object | DeepLab v3, which labels every pixel as one of 21 known categories | Scenes with several recognised subjects, where you want to choose which to keep | Anything outside its 21 classes — shoes, food, jewellery, tools |
| By colour | Straight arithmetic: any pixel within a set distance of the sampled colour becomes transparent | Real green screens and genuinely uniform studio backdrops | Any background with shading, gradient or texture — and subjects containing the key colour |
Object mode's categories are worth knowing before you rely on it: background, aeroplane, bicycle, bird, boat, bottle, bus, car, cat, chair, cow, dining table, dog, horse, motorbike, person, potted plant, sheep, sofa, train and TV monitor. Anything it detects covering more than half a percent of the image appears as a chip, with the largest selected automatically. If the thing you want is not on that list, the model cannot see it and AI mode is the answer.
Colour mode measures distance in RGB space from the colour you sampled — it starts by averaging the four corner pixels, and clicking anywhere on the image re-samples. Because it is pure arithmetic with no notion of what a subject is, a red jumper will be cut full of holes if you key out a red background.
Hair and fur are the hard case, universally. A cutout is fundamentally a decision about each pixel, and a strand of hair thinner than a pixel is genuinely part background and part hair. Models approximate this and the approximation shows: expect a slightly chopped outline on flyaway strands, and worse against a busy background. The same applies to motion blur, glass, veils, smoke and anything else that is partly see-through — none of these has a correct hard edge to find.
Contrast drives quality more than the subject does. A dark jacket against a dark wall gives the model very little to work with. If you can influence the photograph, that is where to spend the effort — a plain, evenly lit backdrop that contrasts with the subject makes every mode work better and can make colour mode sufficient on its own.
The by-colour and by-object modes preview at reduced size. They work on a copy capped at 1200 pixels wide so that adjusting tolerance stays instantaneous, then scale the finished cutout back up to your original resolution on export. The exported file is full size — no resolution is held back, and there is no watermark — but edges from those two modes are slightly softer than AI mode's, which runs on your original file throughout.
AI mode needs a download and a moment. The model is fetched the first time you use it and cached afterwards, so the first run needs a connection and takes noticeably longer than later ones. Object mode fetches its own model separately. If either fails to load, the other modes still work — colour mode needs nothing beyond the page itself.
Output is always a PNG. That is the only common format with real transparency, and it is why a cutout of a large photo can be a substantial file. If you replaced the background with a solid colour, there is no transparency left to preserve — running the result through PixSqueeze as a JPEG will often cut the size by most of it.
Work in the order the controls are laid out: get the rough cutout from whichever mode suits the image, then feather, then refine by hand. The feather slider blurs the transparency edge and then pulls the faint outer halo back, recomposing those pixels from the original photo so they blend rather than fringing. One or two pixels is right for most subjects; push higher for hair, where a slightly soft edge reads better than a jagged one. Leave it off for hard-edged products where crispness is the point.
The refine brush is where a good cutout becomes a convincing one, and the trick is brush size. Use a large brush to clear big leftover regions quickly, then drop to something small — around ten to twenty pixels — for the edges themselves, switching between erase and restore as you go. Restore paints back from the original image, so you can recover an ear or a strap the model trimmed off without starting over.
When compositing onto a new background, the cutout is rarely what gives it away. Mismatched lighting direction, colour temperature and the absence of any contact shadow are what make a composite read as fake. The blurred-background option sidesteps all of that by keeping the original scene, which is why it looks natural on portraits with almost no effort. If you need a specific backdrop, pick one photographed in similar light.
For product listings, transparent PNG is the flexible choice, but check the marketplace's requirements first — many specify a pure white background rather than transparency, which the colour swatch produces directly. Then resize to the required dimensions and compress. If you only need to remove a distracting edge rather than the whole background, cropping is faster and lossless; the full set is on the image tools page.
Yes — completely free, no sign-up, and no watermark. You download a full-resolution transparent PNG. The by-colour and by-object modes preview on a reduced-size copy for speed, but the exported file is scaled back to your original resolution rather than being capped.
No. The AI model runs directly in your browser, so your image never leaves your device. The model is downloaded once and then cached for next time.
The first time, your browser downloads the AI model (a few seconds depending on your connection). After that it's cached and runs much faster.
If your image already has a plain, solid-colored background, the By color mode removes it instantly — just click the background to pick the color and adjust the tolerance. It compares each pixel's distance from that colour in RGB, with no understanding of the subject, so it needs a genuinely uniform background and will punch holes in a subject that shares the key colour.
Yes. After removing the background, keep it transparent or choose white, black or a custom color, then download. You can also load your own photo as a background, or blur the original scene — which is usually the most natural-looking option for portraits, since the lighting already matches.
Because a single strand is thinner than a pixel, so that pixel is genuinely part hair and part background — there is no correct hard edge to find, and every cutout tool approximates it. Two things help: raise the feather slider a few pixels so the transition softens instead of stepping, and use the refine brush at a small size to restore the strands that matter. Glass, veils, smoke and motion blur have the same underlying problem.
AI mode uses a general foreground-subject model and works on nearly any photo. By object uses DeepLab v3, which classifies pixels into 21 fixed categories — person, car, cat, dog, chair, bottle and so on — and lets you tap which of the detected things to keep, which is useful when a photo contains several subjects. If what you want to cut out is not one of those categories, it will not be offered, and AI mode is the one to use.