Removing JPEG noise from an image
The blocks and shimmer in a JPEG are compression damage, and they have to be cleaned as the image is enlarged.
Steps
- Open the image
Drop the JPEG onto Pixmend. Processing happens in your browser. - Raise the noise reduction setting
Compression artefacts need more reduction than a clean photograph does. - Choose a scale factor
Even x2 helps, because the cleanup happens during the same pass. - Check edges at full size
Look at hard edges and flat gradients, where artefacts are most visible.
The blocky squares in a gradient and the faint shimmer clinging to hard edges are not noise in the camera sense. They are what JPEG compression left behind, and removing them works differently from removing grain.
What you are actually looking at
JPEG saves space by discarding information in blocks of eight pixels. Push it far enough and the blocks stop being invisible:
- Blocking. Flat squares in what should be a smooth gradient — skies and shadows show this first.
- Ringing. A halo of noise along high-contrast edges, most obvious around text and line art.
- Colour bleeding. Colour is stored at lower resolution than brightness, so saturated edges smear sideways.
None of this was in the original scene. It is damage introduced by saving, which is why it gets worse every time the file is re-saved rather than staying constant.
Why cleanup and enlargement belong in the same step
The intuitive order — clean it up, then enlarge it — is the wrong one, and so is its reverse.
Enlarging first magnifies the artefacts along with the image, turning eight-pixel blocks into sixteen-pixel blocks that are now unmistakable. Cleaning first, with a conventional blur-based filter, removes the artefacts by removing detail, and hands the upscaler a softer image with less to work with.
Doing both at once avoids the problem: the model reconstructs edges while suppressing the artefacts around them, because it was trained on images that had both characteristics at the same time. That is why the noise-reduction setting sits next to the scale factor rather than in a separate menu.
Settings that work
Start with a moderate noise-reduction setting and go up only if artefacts remain. Over-reduction flattens real texture — skin, fabric, foliage — and the result reads as plastic.
x2 is enough for cleanup. If you do not need a larger image, x2 still runs the same reconstruction and cleans the artefacts. There is no separate "clean only" mode because there does not need to be.
Match the model to the content. Photographs through the photo model, drawings through the illustration model. Ringing around line art is particularly visible, and the illustration model handles those edges better.
Checking the result
Look at three places at 100%:
- A smooth gradient — sky, a wall, a shadow. Blocks should be gone, and the gradient should not have turned into a smooth-but-flat area.
- Text or a hard edge. The halo should be gone without the edge itself softening.
- An area of real texture. This is where over-reduction shows. If skin or fabric has lost its grain, lower the setting and run it again.
Since processing is free and unlimited, running the same image at two settings and comparing is the fastest way to find the right one.
What cannot be recovered
Compression removed information permanently. What the model does is reconstruct plausible detail in its place — not retrieve the original pixels, because those are gone.
For a heavily compressed image, this means the result will look clean but will not match what the original photograph contained. That is usually what you want. It is worth knowing the difference, though, particularly if the image is evidence of something rather than decoration.
The practical habit that follows: keep the largest, least-compressed original you have, and avoid re-saving JPEGs repeatedly. Each save is a small permanent loss, and no amount of processing afterwards reverses it.