One discards detail permanently to reach a much smaller file. The other guarantees every pixel survives. Which you want depends on what is in the picture.
Lossless rewrites the file so it takes less space while every pixel value survives exactly — decode it and you get the original back, byte for byte.
Lossy throws information away on purpose, choosing what to lose based on the limits of human vision, and cannot be reversed.
That is the whole difference, and everything else follows from it.
A photograph is mostly noise. Sensor grain means neighbouring pixels differ slightly even in a flat expanse of sky, and lossless algorithms have almost nothing repetitive to collapse. That is why a 12-megapixel photograph saved as PNG is commonly 15–25 MB.
A lossy encoder looks at the same image and asks a different question: which of this detail would anyone actually notice? The answer is very little of the finest detail, and none of the colour resolution. Discarding it produces the same photograph at 1.5 MB.
Ten times smaller, no visible difference. On photographic content the argument is not close.
Now invert the content. A screenshot is large flat areas, hard edges and text — enormously repetitive, which lossless compression exploits brilliantly.
Lossy encoders, meanwhile, are built on the assumption that sharp transitions are rare. Give one a black letter on a white background and it cannot represent the edge cleanly, so it produces ringing: faint ghost echoes either side of every stroke. Chroma subsampling compounds it, smearing coloured text into a dirty fringe.
So for graphics the lossless file is frequently both smaller and sharper. There is no trade-off to weigh.
Ask one question: does this image have large areas of flat, uniform colour and hard edges?
Be wary of tools advertising "lossless" savings of 70–80% on photographs. That figure is not achievable losslessly, and what is usually happening is palette quantisation — reducing the image to 256 colours — which is a lossy operation whatever the marketing calls it.
Equally, "lossless JPEG compression" almost always means one of two narrower things: stripping metadata without touching the image, or rearranging the existing JPEG data with a tool such as jpegtran. Both are legitimate and both save far less than the phrase implies. JPEG itself has no lossless mode — the quantisation step that discards information is the format, not a setting.
Everything below runs in your browser, so you can test any of this on your own files without uploading them.
Zero quality loss, every pixel preserved exactly. Honest about how modest the saving is.
Open the toolShrink JPG and JPEG photographs with a quality slider or an exact target size.
Open the toolLossless by default, with transparency carried through untouched.
Open the toolTypically 25–35% smaller at matched quality. The standard web win.
Open the toolQuestions
Every answer below is present in the page source, expanded, so it can be read without opening anything.
You can change the container, but not recover the detail. Saving a JPEG as PNG produces a large, lossless copy of the already-degraded image — every artefact preserved perfectly. The original detail was discarded at the first save and no longer exists in the file.
The PNG format is, but tools that produce PNG-8 are not being lossless with your image — they reduce it to a 256-colour palette first, which is a lossy step done before a lossless save. If a tool offers a big PNG saving, check whether it is quantising.
Either — it supports both modes, which is unusual and useful. Lossy WebP beats JPEG by roughly 25–35% at matched quality; lossless WebP beats PNG by roughly 25% on the same image. Browser canvas encoders produce the lossy variety.
At quality 80 on a photograph, essentially nothing you would see at normal viewing size — you have to zoom to 200–400% and look at flat areas such as sky or skin to find differences. Below about quality 60 it becomes visible without looking for it.
Lossy for photographs, lossless for interface graphics and logos, and WebP for both if you can — it covers each case in one format and is smaller than the alternative either way.
Neither, ideally: keep the camera original untouched. If you must re-encode for archiving, use a lossless format so that future copies start from something undamaged.
From the blog
Longer, more practical guides from the blog.
PNG stored the same photograph in 1274.1 KB where JPEG needed 67.5 KB — nineteen times larger. On a flat graphic PNG won. On a smooth gradient PNG lost to JPEG by a factor of two, which surprises most people.
9 min read
WebP produced a 27.0 KB file against MozJPEG's 41.1 KB on the same photograph at slightly higher measured fidelity — 34% smaller. It also produced a larger file than JPEG on random noise. The advantage is real and content-dependent.
9 min read
Quality 80 is not 80% of anything. It is an index into a quantisation table. Measured on one photograph: dropping from 100 to 90 removes 77% of the file and 4.2 dB of fidelity; dropping from 40 to 30 removes 17% and costs almost as much.
8 min read