Remove Noise from a Photo
A traditional denoiser subtracts noise from your photo. This does something different — it repaints the picture with the noise absent, at just over half strength, and hands you the repainting. That distinction explains both why it clears grain a filter cannot touch and why fine texture sometimes comes back invented rather than recovered. It also gives the same answer every time you run it, which for a repair tool is a deliberate choice.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeDrop an image here, or click to browse
Up to 25 MB. Uploaded temporarily, deleted after processing.
What it does
- Partial repaint at 55 per cent strength
- Fixed seed, so runs are reproducible
- Identity and composition explicitly guarded
- Roughly one megapixel output
How to use Remove Noise
- 1
Keep the original
This returns a new image at a different size rather than editing yours. The source stays the reference for judging whether real texture survived.
- 2
Run the pass
The photo is scaled to about one megapixel and repainted at 55 per cent strength with an instruction to clear grain without smoothing.
- 3
Check texture, not noise
The noise will be gone. The question is what replaced it — inspect skin, fabric, hair, and foliage at full size against the original.
How it works
Your photo 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: reduce visible digital noise and compression grain, while preserving natural texture, facial identity, fine edges, text, color, and the composition, and without over-smoothing or inventing detail.
A second, negative instruction lists what the result must not contain — changed identity, altered composition, plastic skin, added or missing objects, text, watermarks.
The model then runs a partial repaint at 55 per cent strength, and the decoded image is what you download.
Repainting rather than filtering
This is the whole difference, and it explains both the strengths and the costs.
A conventional denoiser is arithmetic. It examines a pixel, looks at its neighbors, decides how much of the difference between them is signal and how much is noise, and subtracts accordingly. It never invents anything, and it never really recovers anything either — the information the noise destroyed stays destroyed, so aggressive settings trade grain for mush.
This works the other way round. It starts from your photo, adds controlled uncertainty, and redraws towards an image that matches the instruction. Where a filter can only remove, this can reconstruct: an edge buried in shadow noise gets redrawn as an edge, because the model has seen a great many edges.
The cost sits in the same mechanism. Reconstruction is inference. What comes back is a plausible version of what was there, not a measured one.
What 55 per cent strength means
The strength setting decides how far the repaint is allowed to travel from your photograph.
At low values almost nothing changes and the noise stays. At high values the model produces a clean, confident image that is increasingly its own — the sort of result where the composition is right and the person is subtly someone else.
Fifty-five per cent sits deliberately in the middle: enough freedom to rebuild texture that noise had swallowed, not enough to redesign the picture. It is a compromise rather than an optimum, and on a lightly noisy image it will feel like too much, while on a badly degraded one it will feel like too little.
The same answer twice
The seed is a literal number rather than a random one, so this workflow is reproducible by design.
That is unusual for a generative tool and correct for a repair tool. Restoration is something you check — you open the result beside the original, look at the face, look at the text, decide whether it is faithful. A tool that returned a different image each run would make that check meaningless, because the file you approved would not be the file you get next time.
Repeated runs will not give you a better result. If the output is wrong, it will be wrong the same way again, and the answer is a different tool or a different source rather than another attempt.
Judging the result
Do not look at the noise. It will be gone; that part is easy and it is not what can go wrong.
Look at what took its place. Skin should have pores rather than a smooth gradient. Fabric should have weave. Hair should have separate strands rather than combed bands. Foliage should have leaves rather than green shapes suggesting leaves. Any text inside the photograph should still say what it said.
Where those hold, the repaint recovered something. Where they do not, it replaced something, and only you can say whether that matters for what the picture is for.
Publication gate
This page ships once the workflow has been run against a high-ISO night photograph, a heavily compressed download, a portrait with visible skin texture, and an image containing small text, with each output compared to the source at full size.
Examples
Night photo from a phone
The best case. High-ISO shadow noise is structureless, so there is little real detail to confuse with it and little to lose.
Over-compressed social download
Compression blocking is a repeating pattern, and repainting handles it better than filtering does. Text inside the image is the thing to check first.
Frequently asked questions
Why is my result smaller than the photo I uploaded?
The image is scaled to about one megapixel before the model reads it, because that is the working size the graph is built around. A 12 megapixel phone photo therefore returns at roughly 1152 by 864. If you need the original dimensions, run an upscale afterwards, and expect the two-step result to differ from a single pass.
Does the resize do some of the work by itself?
Yes, and it deserves saying rather than hiding. Downscaling averages neighboring pixels, and noise is random from pixel to pixel, so it averages away while real detail survives. On a very large noisy source a noticeable share of the improvement happens before the model runs at all. Downscaling your own copy is a fair thing to compare against.
Will I get the same result if I run it twice?
Yes. The seed is fixed at a literal value rather than randomised, so the same input produces the same output every time. That is intentional for a repair pass — if you compare a result against the original and decide it is right, you need to be able to reproduce it, which a tool that rolls a new number each run cannot promise.
Why does the skin look slightly plastic?
Because noise and fine texture are close to indistinguishable at the pixel level, and a model told to remove one will sometimes take the other. The instruction explicitly asks it not to over-smooth and the negative prompt names plastic skin directly, which reduces the failure without eliminating it. Portraits are the case that most needs checking at full size.
Is the AI Remove Noise tool free, and do I need an account?
It is free and there is nothing to register for. No account is created, no address is collected, and no reduced-quality preview is withheld for a plan that does not exist. The constraints that do apply come from the workflow itself rather than from billing: 25 MB in, a shared GPU queue, and a thirty-minute window to collect the output.