Background Blur
Uses a local AI model (the same one behind the background remover) to recognize the main subject in your photo, then blurs everything around it by an amount you control — a portrait-style depth-of-field (bokeh) effect, entirely in your browser with nothing ever uploaded to a server.
Your image is processed entirely on your device by a local AI model — it is never uploaded to any server.
Drag an image here or choose one
JPG, PNG, or WebP, up to 20 MB
What this tool does
Blurs an image's background while keeping the main subject sharp, using an image-segmentation model that runs entirely inside your browser and identifies the salient object in the picture.
When to use it
- A portrait has a distracting background and you want attention on the person.
- A photo for a profile or a call needs to look cleaner.
- You want a depth-of-field look from an ordinary phone photo.
How to use it
- Put the image on the drop area or choose it.
- Wait for the model to load the first time — it downloads once and is then cached.
- Set the blur amount.
- Apply the processing and download the result.
A practical example
Input
A photo of someone standing in front of a busy street.
Output
The person as sharp as before, with the street behind them blurred by the amount you chose.
Limitations
- The model identifies the salient object, not people specifically — an image with no clear main subject can produce an unexpected result.
- Fine boundaries such as stray strands of hair may not be separated perfectly.
- The limits are 20 MB per file and 4096 pixels on the longest side.
- It needs a recent browser capable of running the model locally; otherwise you get an explicit message suggesting one.
- The first run fetches roughly 7 MB of model and runtime files, so it is slower than subsequent ones.
Privacy and processing
The model runs entirely inside your browser: neither the image nor any pixel of it leaves your device, and no server is called during processing.
Technical notes
- The model is u2netp, a compact version of the U-2-Net salient-object segmentation network, about 4.6 MB.
- It predicts an alpha mask at 320×320 which is then upscaled and applied to the original image — genuine image segmentation, not a generative model.
- Both the runtime and the model are self-hosted on this site and are fetched only when this tool is opened, so they add no weight to any other page.
Frequently Asked Questions
No — the recognition model runs entirely in your browser, and your photo never leaves your device at any stage of processing.