Canvas · Free Tool

Background Remover

Drop in the images for your mood board, get transparent cutouts back. The model runs on your own device, so nothing is ever uploaded.

Loading tool…

Mood boards are made of cutouts. A chair floating free of the showroom it was photographed in, a jacket without the model, a plant with no wall behind it — once the backgrounds are gone the objects can sit next to each other and actually be compared, instead of each dragging its own lighting and setting into the frame. This tool takes a batch of images and gives you back transparent PNGs, one per image, ready to arrange.

The cutting happens on your own device. An open-source segmentation model downloads into your browser once and runs there, which means the images you feed it — client photos, unreleased product shots, pictures of your own house — are never transmitted anywhere. That is the difference from the well-known paid services: not the price, the fact that nothing is uploaded at all.

Give your mood board somewhere to live

OmniCanvas puts your cutouts, references and notes on one infinite canvas you can keep rearranging — not a folder of PNGs you never open again. Free to start.

Sign up free

No credit card · 30 days free

How to remove a background here

Drag in up to twelve images at once, or choose them from a file picker. Pick a model — the detailed one is better on objects and interiors but needs WebGPU and a larger one-time download; the fast one is a fifth of the size and runs anywhere, and is strongest on people. Then press the button once and the whole batch is processed in order, with progress for each image so you can see what is actually happening.

Each finished cutout appears on a checkerboard so you can see where the transparency actually falls, with a toggle to flip back to the original and compare. Download them one at a time, or take the whole set as a single zip. None of that requires an account, and there is no watermark on anything.

What these models do well, and where they fail

Background removal models predict, for every pixel, how much of the subject is there. They are genuinely good when there is one clear subject with an edge you could trace: a piece of furniture, a bottle, a shoe, a person standing against a wall, a plant on a plain floor. On that kind of image the result is usually better than you would cut by hand in the same amount of time.

They struggle in predictable places. Flyaway hair and fur produce a hard edge where the real one is soft and wispy. Anything semi-transparent — glass, water, smoke, a net curtain, a veil — comes out either fully solid or fully gone, because a binary-ish matte cannot express "70% there". Motion blur, low contrast between subject and background, and images with several equally prominent objects all reduce accuracy, and the model may quietly pick the wrong subject entirely. Check every cutout before you build something on top of it; when one comes out badly, a tighter crop around the subject is the single most effective fix.

Why nothing is uploaded

The usual architecture for a free background remover is: your image goes to a server, a GPU there runs the model, an image comes back. It works, and it means a third party holds your picture for at least as long as the request takes — with whatever retention, logging and training policy they have chosen this quarter. For an interiors project, an unreleased product, or a photograph of someone who did not agree to be in a vendor's logs, that is a real consideration rather than a theoretical one.

Here the model itself is what travels. It downloads once from the model host, your browser caches it, and every image after that is processed locally — on the second run there is no network request at all. Two consequences worth knowing: the first cutout is slower than a server-based tool because of that download, and after it, the tool works with the network off entirely.

Making a mood board that is actually useful

A pile of attractive images is not a mood board; it is a mood. The boards that change what gets made have an argument in them — a stated direction at the top, images chosen because they answer it, and a note under each one saying what it is contributing. "Warm, worn, cheap to build" is a direction. "Nice kitchens" is a search query. Write the direction before you collect anything, and be willing to reject images you like because they are pulling somewhere else.

Size is signal: whatever is biggest on the board reads as the main reference whether you meant that or not, so put the anchor image first and large, and let the rest support it. Group things that agree with each other and leave a real gap between groups, because a board that is evenly tiled edge to edge tells the viewer nothing about which parts matter. Add a couple of anti-references — this, but explicitly not that — since knowing what a board is ruling out is usually more useful than another example of what it likes.

Then finish it. Add a line at the bottom that says what you are going to do because of the board, and revisit it after the first draft of the real thing. A mood board that is never returned to was decoration; one that gets argued with, pruned and re-ordered is doing its job.

When to cut out, and when to leave the background alone

Cut out when you are comparing objects. Silhouettes, proportions, colours and materials only compare cleanly when each item is free of its own setting — five sofas from five different showrooms are impossible to judge until the showrooms are gone. Cutouts also let you overlap and cluster things tightly, which is how a board starts to read as one composition rather than a grid of separate photographs.

Leave the background when the context is the point. A room, a shop front, a street, a lighting setup, a way of composing a photograph — for those the environment carries the reference, and cutting it out throws away the very thing you saved the image for. Full-bleed images also make good anchors: one large uncut photograph setting the overall tone, with cutouts arranged around it, usually reads better than a board made entirely of floating objects.

Sourcing images responsibly

One thing worth being plain about: removing a background does not give you any rights to the image. A cutout is a derivative of someone else's photograph, and a product shot lifted from a retailer or a photograph found through image search remains theirs after the background is gone. For a private board used to think with, that is normally fine under ordinary reference use. For anything that gets published, presented to a client as your own work, or built into something you sell, it is not — and "I edited it" does not change that.

The practical route is to note where each image came from as you collect it — a link under the tile takes seconds and saves an hour later — and to source from places whose licence you have actually read: stock libraries you have a licence for, photographs you took yourself, supplier images used with permission, or public-domain and permissively licensed collections. Photographs of identifiable people carry a separate question about their consent that no licence answers for you.

From cutouts to a canvas

Cutouts in a downloads folder are not a board yet. "Start the mood board" hands a labelled layout to OmniCanvas — a direction card, one tile per cutout, and a closing prompt asking what you will do because of it — and creates the note before it asks you to sign in. Drag the PNGs you downloaded onto the tiles and the board is arranged on an infinite canvas you can keep adding to.

The images themselves stay on your device and are not part of that hand-off, which is deliberate: the whole point of processing locally would be undone by quietly shipping the results somewhere to build a layout. The canvas is a place for them to live, not a copy of them being taken.

Frequently asked questions

Is my image uploaded to a server?

No. The segmentation model downloads into your browser once and runs on your own device. Your images are decoded, processed and discarded locally, and no request in this tool carries image data. After the model is cached, the tool works with the network off.

How many images can I do at once?

Twelve per batch. Each image holds a decoded copy and a cutout in memory alongside the model, and that limit keeps a normal laptop comfortable rather than running the tab out of memory. Process a batch, download it, and start another.

Why does the first image take so long?

The model has to download before anything can run — around 44 MB for the fast model or 219 MB for the detailed one. Your browser caches it, so the second and subsequent images are much quicker, and later visits skip the download entirely.

Which model should I pick?

The detailed model is better on objects, products and interiors, but it needs WebGPU (recent Chrome, Edge or Safari) and a larger download. The fast model runs in any modern browser, downloads a fifth as much, and is strongest on people — it just leaves more behind on cluttered object shots. The tool picks a default based on what your device supports and you can switch at any time.

Why does the edge of my cutout look wrong?

Almost always hair, fur, glass or something else with a soft or semi-transparent edge. These models produce a fairly hard matte, so wispy detail is either kept as a solid clump or cut away. Cropping tighter around the subject before you run it is the most reliable improvement; some images simply need a manual mask instead.

Is there a watermark or a paid tier?

No watermark and no paid tier. The cutouts are full-resolution transparent PNGs and both the individual downloads and the zip are free with no account. Saving a mood-board layout to a canvas is the only step that involves an OmniCanvas account, and it is optional.

Does removing the background mean I can use the image?

No. Editing someone else's photograph does not give you rights to it. A cutout is a derivative of the original, so for anything published, sold or presented as your own you still need a licence or permission — stock you have licensed, your own photographs, or permissively licensed sources. For a private reference board, ordinary reference use is normally fine.

What image formats work?

PNG, JPG, WEBP, BMP and GIF, up to 25 MB each. Very large images are scaled down to 1600px on the long edge before processing, which is more than enough detail for a mood-board tile and keeps memory use sane.

More free tools