Compression artifacts: blocking, ringing and banding

The image breaks into visible squares, most obviously in areas that should be smooth — a clear sky, a wall, a shadow.

The cause is structural. JPEG processes the image in independent 8 × 8 blocks, and when quantization is aggressive enough, each block is reduced to so few surviving coefficients that its average colour drifts from its neighbours. The block boundaries stop lining up, and you see the grid.

The fix: raise the quality. Blocking is the clearest signal that you have gone too far — below about 55 on most photographs. It cannot be fixed by resizing or by changing format after the fact, because the damage is already in the pixels.

Faint ghost echoes appear either side of a hard edge — around text, around a logo's outline, along a roofline against bright sky. It is sometimes called mosquito noise, because it shimmers around edges in video.

The cause is a mismatch between the content and the codec. Describing an instantaneous jump from black to white requires an infinite series of wave patterns; the encoder keeps a finite number, and the truncated series overshoots and undershoots around the edge.

The fix: usually not a quality setting at all. Ringing on text means the image should not be a JPEG. Move it to PNG or lossless WebP and it disappears completely, often at a smaller file size too.

A gradient that should fade smoothly instead steps through visible stripes — most often a sunset, a clear blue sky, or a studio backdrop.

Two things cause it. Quantization can leave too few distinct values to represent a slow transition, so neighbouring regions snap to the same value. And 8-bit colour only has 256 levels per channel to begin with, which is marginal for a very gradual fade across a wide frame.

The fix: raise the quality — banding appears earlier than blocking, so a gradient-heavy image may need 85 where a textured one is fine at 75. Some encoders can add a small amount of dither, which trades faint noise for a smooth appearance.

Red text on white looks muddy and orange at the edges. Thin saturated lines lose their colour or bleed sideways. Raising the quality barely helps.

This is chroma subsampling, and it is a separate lossy step from quantization. The encoder stores colour at half resolution in each direction, which is invisible on photographs and destructive on saturated fine detail.

The fix: use a format and encoder that keeps colour at full resolution (4:4:4), or — far more practically — recognise that an image of coloured text belongs in PNG.

Look at the image at the size it will actually be displayed, not zoomed to 400% in a viewer. Artefacts that are obvious under magnification are frequently invisible at display size, and optimising against a zoomed view wastes bytes on nobody's behalf.

Then do look deliberately at the places faults appear first: large flat areas, gradients, skin tones, and any hard boundary between contrasting colours. If those are clean, the image is clean.

The most reliable way to judge is a direct comparison against the original at the same scale — which is what the before-and-after slider on the tool pages is for.

If an image looks smeared and blotchy in a way that does not match any single description above, it has probably been through several lossy saves. Each one re-quantises data that was already quantised, treating the last pass's errors as real detail and adding new errors on top.

This is generation loss, and no setting undoes it. Go back to the original file and compress it once.