Background Blur Tool

Keep the background, throw it out of focus behind the subject. One image at a time.

Drag & drop an image here

or tap to browse (PNG, JPEG, WebP, SVG)

One image at a time, up to 25 MB

Background

Blur

One image at a time

Load a single photograph

This works best on images that have a real background worth keeping: an environment, a room, a street. A subject already shot against a plain wall gains very little from blurring it.

Tip: Some depth in the original scene sells the effect. A flat backdrop blurred still looks like a flat backdrop.

Let the model find the subject

The network runs locally and separates subject from background. The first run of a session downloads the weights, after which the browser cache handles it.

Tip: Subjects with a clear outline give the cleanest separation. Overlapping people or objects are harder.

Process and download

The background is blurred, the sharp subject is composited over it, and the finished image comes back as a PNG in your download.

Tip: Look closely at the boundary. Blur exaggerates a bad mask far more than a flat colour fill does.

Context is preserved

Unlike a cut-out, the setting stays in the picture. The image still reads as a photograph taken somewhere real, just with attention pushed firmly onto the subject.

Subject stays genuinely sharp

Only the masked background is blurred. The subject is composited back at full detail rather than being softened along with everything else, which is what separates this from a global blur.

Local segmentation and compositing

Both the model and the blur run in your browser using canvas filters. Your photograph is never transmitted and the tool keeps working with no network connection once the model is cached.

One image per run

The engine allows exactly one segmentation at a time. Concurrent runs exhaust the memory a browser tab is allowed, and on phones they fail outright rather than merely slowing down.

Twelve pixel Gaussian

The blur radius is set to a value that reads as shallow depth of field at normal viewing sizes without dissolving the background into unrecognisable colour fields.

Half precision on mobile

Phones and tablets automatically receive a smaller model and a reduced working resolution, so the tool remains usable on hardware that could not run the full network.

Frequently Asked Questions

How is this different from blurring the whole image?+

A global blur softens the subject too, which just looks like a mistake. Here the subject is masked out, only the background is blurred, and the sharp subject is composited back over it. The contrast between the two is the entire effect.

Is this the same as portrait mode on a phone?+

Similar in intent, different in method. A phone often has depth information from a second camera or a dedicated sensor. This works from the single image alone using segmentation, which is why it depends on the subject having a recognisable outline.

Why can I only process one image at a time?+

The segmentation model holds large intermediate tensors in memory alongside tens of megabytes of weights. Running two concurrently reliably exhausts what a browser tab can allocate, so the engine enforces a single slot.

The model stays cached between runs, so processing a series one after another is only slow for the first image.

Can I change the blur strength?+

Not currently. The radius is fixed at twelve pixels, chosen to read as shallow depth of field at typical viewing sizes without turning the background into abstract colour.

Why does the edge look wrong on my image?+

Blur is unforgiving of an imperfect mask, because a sharp fragment of background left inside the subject outline sits against a soft field and stands out immediately. Busy backgrounds and fine hair are the usual causes.

Does my photograph leave my device?+

No. Segmentation, blurring and compositing all happen in your browser. Nothing is uploaded, nothing is logged, and no copy exists anywhere but your own machine.

Format Comparison

Feature Specification	background blur	Alternative
What happens to the background	Blurred, still recognisable	Removed entirely
Subject sharpness	Full detail, composited last	Full detail
Sense of place	Preserved	Lost
Tolerance for mask errors	Low, errors are visible	Moderate
Output file size	Larger, the background is still there	Smaller, most pixels are transparent

Technical Overview & Specifications

Deep dive into background blur architecture and processing.

Show Details

Platform Overview

A blurred background does something a removed one cannot: it keeps the context. You can still tell the photograph was taken in a kitchen or on a street, the colours behind the subject still relate to the colours in front, and the result reads as a photograph rather than a cut-out pasted onto a card. Optically this is what a wide aperture gives you, and it is the single most recognisable difference between a picture taken on a proper lens and one taken on a phone. Doing it after the fact means separating the subject from the background and blurring only what is behind, which is exactly the same segmentation problem the other background tools solve.

The model here is ISNet, the same network used for transparent cut-outs, running in WebAssembly in your browser with WebGPU acceleration where it is available. It produces a per-pixel mask, the background is blurred with a twelve pixel Gaussian, and the sharp subject is composited back over the top. Nothing is uploaded and nothing is stored. As with every background tool on this site, exactly one image is processed at a time, because the network needs a large and contiguous chunk of memory to run and competing for it is how you get a tab killed halfway through.

Masking, blurring and recombining

The pipeline runs in three stages. First ISNet produces a per-pixel confidence mask separating subject from background, the same output the transparent and solid colour tools use. Then the original image is drawn to a canvas with a twelve pixel Gaussian blur applied through the canvas filter property, which delegates to the browser's own optimised implementation rather than convolving pixels by hand in JavaScript. Finally the masked subject is drawn over the blurred copy at full sharpness. Because the mask carries partial confidence rather than a hard boundary, the transition between sharp subject and soft background is itself gradual, which is what stops the composite looking like a sticker.

The order matters. Blurring the whole image before masking means the blur samples pixels from the subject and drags subject colour into the background, producing a faint bright halo around the outline. This is the most common artefact in naive implementations of the effect and it is entirely avoidable by compositing the sharp subject last. Where the environment does not expose a canvas filter, the code detects the missing capability and skips the blur rather than silently producing an unblurred result that looks like a failed run. As with the other background tools, every segmentation call passes through a single concurrency slot, so a second image waits for the first to finish rather than allocating memory beside it.