Blur Behind a Subject Without the Halo

How it works

Your image is resampled to about two megapixels. A segmentation model produces a matte of the subject, and that matte is then grown outwards by more pixels than the blur radius that follows. The whole frame is blurred, and your original sharp image is composited back over it through the grown mask.

There is no generative model. The subject in the output is the subject you uploaded and the background is your background with a blur applied to it, so nothing in the picture has been invented.

The order matters and the growth amount matters, and between them they remove the artifact this page is named after.

The halo, and why it exists

A blur replaces each pixel with a weighted average of its neighbors. That is the whole operation, and it has no idea what a subject is.

So consider a pixel one step outside the subject's outline. Its neighborhood includes a good deal of subject, which means the average includes the subject's colors, which means the blurred value there is partly the subject. The same is true a few pixels out, less so, fading with distance until the blur radius is exhausted.

Paste the sharp subject back over exactly its own outline and all of that contaminated ring remains visible in the background, tracing the silhouette in a soft version of the subject's own color. A red jumper leaves a pink glow. Hair leaves a dark aura.

The fix is not a better blur, it is a wider paste. Grow the mask past the radius and the contaminated band is covered by original pixels. Everything that could have been polluted is simply not shown.

What a real lens does that this does not

Two differences remain, and being straightforward about them tells you when this page suits a photograph.

Blur varies with depth. Defocus grows the further something is from the plane of focus, so a real photograph of a corridor has a slightly soft near wall and an extremely soft far end. This applies one strength everywhere behind the subject, which is nearly correct for a background at a single distance and clearly wrong for a receding one.

Highlights become shapes. An out-of-focus point of light in a real lens is projected as a disc the shape of the aperture — round, or polygonal with visible straight sides. A gaussian blur has no aperture, so it turns a bright point into a fading smudge. Any scene with fairy lights, street lamps, or specular sparkle shows this immediately.

Neither is fixable without depth information and a different blur model, which is what a phone's portrait mode uses its multiple cameras to obtain.

Choose the photograph, not the setting

The practical consequence is that this works beautifully on some images and poorly on others, and you can tell which before you upload.

Good: a subject in front of a wall, a hedge, a crowd at one distance, a cluttered desk, a curtain. Everything behind them is at roughly one depth, so one blur is a fair approximation of what a fast lens would have produced.

Poor: a subject at the end of a long room, on a street stretching away, or in a landscape. The scene has depth and a single blur strength flattens it into a backdrop, which reads as a cutout on a soft picture rather than as a photograph with shallow focus.

Publication gate

This page ships once the workflow has been run against a subject in front of a wall, a subject in a receding corridor, a scene containing bright point lights, and a subject with loose hair, with every silhouette inspected at full size for a colored halo.