Remove Dust and Specks from a Photo

How it works

Your image is scaled so its total area is around one megapixel, keeping the original proportions. An image-editing model reads that copy along with a fixed instruction: clean small dust spots and scanning specks, while preserving the people, objects, edges, texture, text, lighting, and composition, without smoothing meaningful detail or adding anything new.

A second, negative instruction names what the result must not contain — changed identity, altered composition, plastic skin, added or missing objects, text, watermarks.

The model then repaints the image at 60 per cent strength. The seed is a fixed number, so the same file always produces the same repair.

Dust is not a visual category

This is the thing to understand before judging any result.

Every other kind of image defect can be described by how it looks. Blur is a loss of high frequencies. Compression blocking is a repeating eight pixel pattern. Noise is random pixel-to-pixel variation. A machine can be taught to find them because they have a signature.

Dust has no signature. A speck of dust is a small mark that contrasts with what is around it — and so is a mole, a beauty spot, a freckle, a distant bird in a sky, a star, a snowflake, a stray sequin, a pigeon on a roof, a full stop in printed text. Physically identical. The only thing separating them is that you wanted some of them in the photograph.

So the model is not detecting dust. It is guessing at intent from context: marks that sit on no surface, that cast no shadow, that appear in a region where nothing similar appears elsewhere, that break the grain of the print rather than belonging to it. That reasoning is good and it is not reliable, and it is weakest exactly where the mark is ambiguous.

A local problem solved globally

A speck occupies a few hundred pixels of an image containing a million of them.

The obvious way to fix it is locally: select the spot, sample nearby pixels, paint over it, leave everything else untouched. That is what a healing brush in an editor does, and for a handful of visible spots it is both faster and safer than anything here.

This workflow does the opposite. It repaints the entire image and the specks happen to be absent from the repainting. Every pixel is reconsidered, including the ninety nine point nine per cent that had nothing wrong with them.

That is a genuinely poor trade for three spots on a clean photograph, and a good one for a scan carrying hundreds of them across a complicated surface, where selecting each by hand is an afternoon. Choose accordingly rather than treating this as a general-purpose cleanup.

Sixty per cent, set for the worst case

The strength here is slightly higher than on the noise page, which surprises people who assume dust is the smaller job.

It is set for the hard case rather than the typical one. A heavily flecked scan of an old print needs enough freedom for the model to rebuild surface it cannot see through the damage. A photograph with four specks does not, and at this strength it gets rebuilt anyway.

The consequence is that a lightly affected image is changed more than it needed to be. If your source is nearly clean, an editor's healing brush will preserve more of it than this page will.

The resize may get there first

There is an honest observation here that is easy to leave out.

A dust speck is small by definition — often three or four pixels on a large scan. Scaling that scan to one megapixel is roughly a three-fold linear reduction, which turns a four pixel speck into something close to a single pixel, blended with its neighbors.

So a share of the improvement you see happened during the resize, before the model ran at all. On very large scans that share can be substantial. Downscaling your own copy and looking at it is a fair comparison to make, and occasionally it is the whole answer.

Publication gate

This page ships once the workflow has been run against a flatbed scan with glass dust, an old print with paper flecks, a portrait with visible moles, and a night sky containing stars, with each output compared against the source at full size to confirm what was removed alongside the dust.