Blur Background (Portrait Mode)

About the Blur Background (Portrait Mode)

Give any photo the soft, out-of-focus background of a portrait lens. Choose A person or people and an AI portrait model (MODNet) finds them, keeps them sharp and blurs everything behind them evenly. Choose Whatever is in focus for pets, products, food and scenes: a depth model (Depth Anything V2) estimates how far away everything is, and the blur grows with the distance from the point you tap — like a real lens, with near and far both going soft.

The blur is made in linear light with a round-aperture (disc) kernel, so small lights bloom into bokeh circles instead of smudges. Adjust the strength, the focus and how much depth stays sharp, fix the edge with Keep sharp and Blur brushes, and save a full-resolution JPG or PNG. The models run on your device with ONNX Runtime in WebAssembly — your photo is never uploaded.

Common uses

• Portrait-mode photos from a phone or camera that has no portrait mode.
• Profile pictures and headshots with a calm, blurred backdrop.
• Product, food and pet photos where the subject should stand out.
• Thumbnails, posters and social posts that need a clean, out-of-focus background.

Two ways to find the subject

• A person or people uses MODNet, a portrait-matting model (Z. Ke et al., “MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition”, AAAI 2022; Apache-2.0). The matte is refined along the photo’s own edges, the background is blurred with the person left out so their colours do not smear into it, and the person is put back sharp.
• Whatever is in focus uses Depth Anything V2 Small (L. Yang et al., “Depth Anything V2”, NeurIPS 2024). The Small model is published under Apache-2.0; this site serves an 8-bit quantised ONNX export of it (about 27 MB, stored in two parts). It estimates relative depth for every pixel — no special camera needed — which is aligned to the photo’s edges at the working size.

Both run on your device with ONNX Runtime Web (MIT) in WebAssembly. A depth estimate takes a few seconds on a computer and longer on a phone.

How the blur is made

For a camera lens of focal length f and aperture A focused at distance z_f, a point at distance z is drawn as a blur circle of diameter c ≈ A·f·|1/z − 1/z_f| (the thin-lens model; M. Potmesil and I. Chakravarty, “A lens and aperture camera model for synthetic image generation”, SIGGRAPH 1981) — proportional to the difference in inverse depth from the focus. Depth Anything estimates relative inverse depth, so the blur radius here grows linearly with that difference beyond the sharp range, both behind and in front of the focus; Strength sets the unknown scale (like the lens’s aperture).

The photo is split into depth layers, each blurred with a disc the size of its blur circle (larger blurs are computed at a reduced size, which looks the same), and the layers are composited from far to near, so a blurred foreground softly covers what is behind it. Blurring happens in linear light, and with Bright highlights, small points brighter than their surroundings are boosted first — the camera clipped them — so they bloom into bokeh discs. Smooth uses a Gaussian-like blur instead of a disc.

Getting a natural result

• Real portrait lenses blur gradually: a strength between 30 and 60 usually looks most natural; very strong blur exaggerates any mistake on the edge.
• Zoom in on the edge of hair and hands; paint Keep sharp where the subject went soft and Blur where background stayed sharp.
• Glass, reflections, fences, fine hair against a busy background and very thin objects are hard for both models.
• The preview is drawn at a reduced size so it stays quick; the download is rendered again from the full-size photo, with the sharp edge refined along the photo’s own detail.

Limitations

• The models can miss parts of the subject or keep bits of background sharp — check the edges and fix them with the brushes.
• Depth is estimated from one photo: it is relative, not measured, and can be wrong on glass, mirrors, reflections and flat pictures of scenes.
• Photos over 16 megapixels (8 on phones and tablets) are reduced to that size; the preview is shown at a smaller working size.
• The first run downloads 27–42 MB; a depth estimate takes a few seconds, more on older phones.
• Edits are kept only while the page is open.

Privacy

Your photo is processed in your browser and never sent to a server. The first blur downloads ONNX Runtime and a model from MySmartCoPilot — not from a third party — about 27 MB for portraits and 42 MB for scenes; after that your browser’s cache serves them. Saved photos contain no EXIF data from the original, such as the camera or GPS location.