How to read the three settings before you upscale
Mode, scale and denoise are the only choices there are. Mode decides the result, and x2 is not the faster option.
Steps
- Set the mode to match the picture
Choose Photo for photographs and Illustration for drawings and line art. This is the setting that changes the result most. - Pick the scale you actually need
Choose x4 unless the output would overshoot the size you want, then x2. It is a size decision, not a speed one. - Set denoise by where the file has been
Leave DN on Low for an ordinary photo, raise it to Strong for a heavily compressed JPEG, and use Off when the grain is part of the picture. - Look at the result at 100% before downloading
Compare the before and after at full size, change one setting at a time, and run it again.
There are three controls above the image and no hidden ones: mode, scale, and DN. The defaults — Photo, x4, DN Low — are right for an ordinary photograph, so if that is what you have, open Pixmend and press the button. The rest of this page is for when the defaults are not right, and it is mostly about two things: mode changes the result more than anything else, and x2 is not the fast option.
The three controls
| Control | Options | Default | What it decides |
|---|---|---|---|
| Mode | Photo / Illustration | Photo | Which model runs. The largest single difference |
| Scale | x2 / x4 | x4 | How big the output is. Not how long it takes |
| DN | Off / Low / Strong | Low | How much compression damage and grain is removed. Photo mode only |
DN disappears when you switch to Illustration. That is not a gap in the interface: the illustration model is a single model with no denoise variants, so there is nothing to choose. When you switch back to Photo, your DN setting is still there.
Mode is the setting to get right
Photographs and drawings break in different ways, and the two models are trained accordingly. A photograph is grain, gradients, and soft edges. A drawing is flat colour bounded by deliberate lines, where the line is the picture — and a model that treats a hard edge as something to smooth will hand back artwork that looks slightly melted.
Judge by what the image is made of, not by its subject. A drawn portrait belongs in Illustration. A photograph of a painting is a photograph. Screenshots of an interface, comics, logos, and pixel-style art all behave like line art, so Illustration is the better starting point for them.
If you are unsure, run it both ways. Everything happens on your own machine, there is no per-run cost, and comparing two results at 100% settles it faster than reasoning about it.
Scale is a size decision, never a speed decision
x2 does not run a smaller model. The x4 model runs first and its output is scaled back down, so the amount of work — and the time — is identical to x4. Choosing x2 to finish sooner does not work, and any time you are told that a lower scale will get you past a size limit on the input, it will not, for the same reason.
So choose x2 for one of these reasons only:
- x4 overshoots the size you want. A 1500 px photo at x4 is 6000 px across, which is more file than most uses need.
- The output is refused as too large. The result has a per-side limit, and the interface will say so before you run it.
- You prefer how it looks. Shrinking any image tends to make small artefacts less visible, so the x2 result can look cleaner than the x4 one seen at the same size.
Otherwise x4, and scale it down yourself later if you want to. Time is decided by the pixel count of the input, not by the scale you choose.
DN: set it by where the file has been
The right denoise setting follows from the file's history, not from its content.
| Setting | Use it when | What it does |
|---|---|---|
| Off | The grain is part of the picture — film, a deliberate texture | Removes nothing, and enlarges the grain along with the image |
| Low | Ordinary photos, camera files, anything you are unsure about | The default. Cleans mild compression damage without flattening texture |
| Strong | The file has been through chat apps, social platforms, or repeated saving | Removes blocks and the shimmer around edges, at the cost of fine texture |
Off is not "the setting that preserves detail". It preserves everything that is in the file, which includes the compression damage, and enlargement makes that damage larger too. Reach for it when you know the grain is wanted; leave it alone otherwise.
Strong is aimed at real damage, and on a clean file it will smooth texture that you would rather have kept — skin, fabric, foliage. If a result comes back looking waxy, that is the setting to move first.
What no setting can change
None of the three invents detail. Nothing here draws what the file does not contain, which is the point of the tool and also its ceiling: a small, heavily compressed image comes back clean and larger, not restored.
The input limit is 12.6MP, and it applies to what you put in. The first run downloads about 11.6MB of models and runtime, kept on your device afterwards, and a 1024×1024 photograph then takes 7.1 seconds, an illustration of the same size 3.9 seconds. Nothing about the image leaves the browser at any setting — you can watch the Network tab in DevTools stay silent while it runs.
Changing settings sensibly
Change one control at a time and look at the result at 100%, not fitted to the screen. Two changes at once and you cannot tell which one helped. The usual order is mode first, since it moves the result most, then DN, then scale last because it does not change the character of the output at all — only its size.