No sign-up · no watermark · unlimited · runs on your device
or drop it here
Checking what this device can do…
More resolution and cleaner edges.
Turns blocks and stair-steps back into gradients.
Resamples the pixels, then writes a real DPI.
Tightens soft edges, keeps your dimensions.
2× or 4×, photos and illustrations.
2×, 3× or 4× on your own GPU, frame by frame.
Crisper edges, without the tell-tale halo.
Tightens soft edges, frame by frame.
Cleans up compressed, soft footage.
Three cases the evaluation run can speak to, including the one the paid tier does better.
BeforeAfter
BeforeAfter
BeforeAfterDrag one picture or a dozen onto the box above, or click to browse. Leave the scale at 1× to clean without changing the size — that is what it loads with.
Light is the default because it measured closer to the clean original than Medium on nine of ten realistic-noise photographs. If grain is still visible after a run, that is what the dial is for.
Zoom the before/after slider into fabric, hair or foliage rather than a flat sky — texture is what strength costs, so that is where to check it. No watermark, no sign-up, no limit.
Texture — and here is how much, one notch apart.
Light — the default · denim weave and hair strands intact
Medium · smoother, and the weave is going
The clean original · the reference both are scored against
The default is Light, and that is a measurement rather than timidity. Two noise recipes were run — one synthetic, one built to imitate a real camera sensor. On the realistic one, Light was closer to the clean original than Medium on nine photographs out of ten, and the realistic recipe is the one that decides: nobody arrives holding synthetic noise.
And the ceiling is the method, not the setting. Fine texture is, to this filter, indistinguishable from noise — both are small variations that do not form an edge. So texture is what you spend, at every strength.
Every cut from this picker is a measurement rather than a preference.
| Method | What happened when it was measured | Offered◆ |
|---|---|---|
| Soft — the one on offer | Closest to the clean original of everything that runs here, and instant. 0.1 s for a four-megapixel photo, nothing to download. | Yes |
| Balanced | Was the first entry in this picker and could not run at all: it ships with noise reduction at zero, so every one of 20 test requests was refused by the planner. | No — it was broken |
| Anime Lines | The same filter with sharpening tuned for line art. Lost to Soft on all ten inputs, in both noise recipes, at both strengths. | No |
| Denoise (Trained) | A 682 KB learned denoiser. Lost 19 of 20 by four to eight decibels, at 25× the processing time — it smooths photographic texture away wholesale. | No |
| Photo · Photo (Grainy) · Anime (Noisy) | All three only clean while enlarging. This page keeps your dimensions, so the planner greys them out with a reason rather than hiding them. | No — cannot hold the size |
The Balanced row is the uncomfortable one, and it is the reason this page was held back rather than shipped. It looked like a working option, it was the first thing in the list, and it would have refused every job. It was found by running this page’s whole test set through the real planner rather than by reading the code — which is the argument for doing that on every page.
The trained option is untested rather than beaten, and the difference matters. What it was built for is film grain and scan dirt — noise that is part of the original photograph. Both recipes here add noise to a clean master, which is not its home ground. Measuring it fairly needs a set of genuinely grainy originals, which has no clean reference to score against and therefore does not exist yet. When it does, this option gets retested and this table gets rewritten.
On heavy high-ISO noise this filter starts trading texture for smoothness. Learned denoisers were trained on real sensor noise and do not have that limit. These upload the image and run on our servers for credits.
| Model | Type | Best for | Per photo |
|---|---|---|---|
| AI Restore | Generative | Cleans and enlarges in one pass | 1 credit |
| Topaz Standard | Precision | Denoises, removes compression and enlarges in one pass | 6 credits |
Read the middle column before you pick: these clean while rebuilding at a larger size, which is the one thing the method on this page deliberately does not do. Billed once per photo rather than by the pixel, so the figure on the Run button is the whole price before anything is sent — no subscription.
Drop the photo into a browser tool that processes it locally. FreeUpscaler runs an edge-preserving bilateral filter on your own graphics card and hands the picture back at its original size — no account, no watermark and no limit.
The trade is texture, and the amount is published above: strength buys smoothness and spends fine detail, which is why the page loads at its lowest working setting rather than its strongest. When the noise is heavier than a formula can separate from texture, a learned denoiser is the right tool, and this page carries two of them in the cloud tier.
The measured edge step does not move at any strength — that is a property of the filter rather than a tuning choice.
The last two are limits no filter gets around, and one of them is where the paid tier is genuinely the right answer.
Indoor shots at a middling ISO, phone photos in mixed light, scans of old prints. This is what the page exists for and what the numbers above were measured on: the grain goes and the edge step does not move.
The mottled colour patches a heavily compressed photo carries, especially in skies and shadows. Chroma noise in low-frequency blotches is exactly the shape the realistic test recipe imitates.
City lights and lit signage are themselves high-frequency, so a filter looking for high-frequency variation cannot fully separate them from noise. The night skyline in the run scored lowest of its group — 0.9226 against a group average of 0.9445. Keep the strength low on this material.
A hand-held shot in a dark room. Turning the strength up does not solve it: an extra notch past Strong measured further from the clean original, not closer, so the ceiling is the method rather than the setting. This is where a learned denoiser earns its price, and it is why Topaz is a row in the cloud table above.
Flat stepped bands in a gradient are not noise — they are the opposite, too few distinct values rather than too many. Smoothing makes the steps softer without making them go away. That needs dithering, which is a different tool.
Grain has one cause almost all of the time, and knowing it tells you whether software or a retake is the answer.
If the photograph is already taken, drop it in above and start at Light. If you can retake it, more light beats every denoiser there is, including the paid ones.
Drop the photo onto the box above and leave the scale at 1×. It is decoded in your browser, cleaned on your own graphics card by an edge-preserving bilateral filter, and handed back at its original size — no account, no watermark and no limit. A four-megapixel photo takes about a tenth of a second, and nothing has to be downloaded first.
A bilateral filter. Each pixel is replaced by a weighted average of its neighbours, where the weight depends on two things: how close the neighbour is, and how similar its colour is. Pixels on the far side of an edge differ too much in colour to contribute, so edges survive while random variation inside a flat region averages away. It is a real edge-preserving denoiser and it needs no model download.
Not the free method. The image is decoded and processed inside your browser tab, and you can verify it in ten seconds: load this page, disconnect from the internet, and clean a photo anyway. A cloud model is the exception and says so before it runs — it uploads the image for that job, and the result is kept for 30 days unless you delete it.
Not on heavy noise. Learned denoisers are trained on what real sensor noise looks like, so they can separate noise from fine texture in a way a mathematical filter cannot. This one is very effective on moderate grain, JPEG noise and colour blotches, and it is free with no limit. On heavy high-ISO noise from a night shot it starts trading fine texture for smoothness — and Topaz is a row in this page's own cloud tier for exactly that case.
Start with Light, which is what the page loads with, and go up only if noise is still visible. That default is a measurement rather than caution: on photos carrying realistic sensor noise, Light came out closer to the clean original than Medium on nine of ten test images. Strength buys smoothness and it is paid for in texture — one notch up is already enough to start losing a denim weave and individual hair strands.
The strength is too high for the amount of noise present. Drop a notch. Denoising is always a trade: past a certain point the filter can no longer tell fine detail from noise, and it removes both. There is no setting, in any tool, that removes all noise and keeps all detail.
No. It cleans the photo at the size you gave it — the scale stays at 1× and the output has the same dimensions as the input. If you also want it larger, do that separately, and in that order: rebuilding first spreads each noise speck over several pixels, turning small grain into larger blobs that are much harder to remove.
It would, which is why the sharpening is scaled back automatically as you raise the noise reduction. Without that, the sharpen pass would simply reinstate the grain the filter just removed. It is also why the gentler of the two available sharpening profiles is the one used here — measured against the line-art profile it was closer to the clean original on all ten inputs, in both noise recipes and at both strengths.
No, and past a point it makes things worse rather than better. Testing an extra notch beyond Strong moved the result further from the clean original, not closer — the ceiling here is the method, not the setting. A bilateral filter has a limit, and that limit is exactly where a learned denoiser earns its price.
Credits pay for the cloud models, because those runs cost real money to perform. Signing in is the wallet for them and nothing else — the local method never asks, and nothing about it is gated. A new account starts with 3 free credits, enough to put one photo through the cheapest cloud model and see what it does.