Yes, but far less than most people fear, as long as nothing changes between saves. The first save as JPEG throws away the most. Opening the same JPEG and saving it again with the same settings costs a little more each time for a few saves, then stops: on 20 images we saved 50 times in a row, the loss had leveled off by about the tenth save. What really wears an image down is changing something between saves, such as cropping, resizing or switching the quality. Then every save loses more, and the damage becomes plain to see.
Why does a JPEG lose quality when you save it?
JPEG is a lossy format. The JPEG committee describes its core system, standardized in 1992, as a "DCT based lossy image format". A JPEG encoder splits the image into small blocks of 8 by 8 pixels, describes each block as a mix of patterns, and rounds those numbers off. The rounding is where detail goes, and it cannot be undone.
When you open a JPEG, the program decodes it back to pixels. When you save it, those pixels are encoded and rounded again. Only the save does this. Opening, viewing, copying, renaming, emailing a file as an attachment or backing it up leaves its bytes exactly as they were.
How much quality is lost with each save?
We measured it. We took 20 images from the Heart Photos wall tagged Made by AI, each stored as a lossless PNG or lossless WebP, so we had exact original pixels to compare with. They were 1.5 megapixels at the median. With Python's Pillow 12.1.1 (which uses libjpeg-turbo 3.1.3), we saved each one as a JPEG at quality 85, opened that JPEG and saved it again, 50 times, and compared every generation with the original pixels.
We scored closeness with PSNR, in decibels: the higher, the closer to the original. A drop of 3 dB means the average error has doubled.
| Saves at quality 85 | Median PSNR | Median loss since the first save | Largest loss of the 20 |
|---|---|---|---|
| 1 | 37.61 dB | ||
| 2 | 37.43 dB | 0.11 dB | 0.61 dB |
| 5 | 37.24 dB | 0.25 dB | 1.57 dB |
| 10 | 37.22 dB | 0.29 dB | 2.28 dB |
| 50 | 37.22 dB | 0.29 dB | 2.92 dB |
Three things stand out:
- The first save is the big one. The next 49 saves together added a median of 7% to the error the first save made (at worst 96%, just under double).
- The loss stops. In 17 of the 20 images, a save eventually produced a file identical, byte for byte, to the one before it, at the 12th save by the median (between the 6th and the 38th). From then on, saving again changes nothing at all. The other three were still changing slightly, but lost at most 0.25 dB more between the 10th save and the 50th.
- Most of what is left comes from color. Pillow stores color at half resolution by default (called 4:2:0 subsampling). When we ran the same test keeping full color resolution (
subsampling=0), 30 saves lost a median of 0.02 dB against the first, which is nothing.
We also looked. Next to the first save, the 50th save of the castle image at the top of this post looked the same to us, even magnified four times.
What makes repeated saves worse?
Changing anything between saves. That is what usually happens in real life: you crop a little, resize for a website, or a different app saves with its own settings.
| What happened before each save | Median loss after 10 saves | After 30 saves |
|---|---|---|
| Nothing (same quality, 85) | 0.29 dB | 0.29 dB |
| Quality switched between 80 and 90 | 0.68 dB | 0.79 dB |
| 16 pixels cropped off the top and left | 0.82 dB | 1.71 dB |
| 1 pixel cropped off the top and left | 6.48 dB | 11.69 dB |
The 1-pixel crop is the worst case, and it is a common one. A crop that is not a multiple of the block size moves every pixel to a different place in its 8-by-8 block, so the encoder rounds the image in a new way every time, and nothing settles. After 30 such saves, the castle image had lost its color and turned grainy: you could see it without zooming. Keeping full color resolution did not help (12.13 dB lost by the 30th save), so the cause is the block grid, not color. A crop of 16 pixels lines up with the blocks, and lost much less.
Switching the quality cost more than twice as much as keeping it, and it was still creeping up at the 50th save (0.80 dB). Saving at the higher setting brought nothing back: every save at 90 came after one at 80, and the losses stayed.
How do you avoid losing quality?
- Keep the original, and edit from it. Make every new version from the first file, not from the last JPEG you exported. Editors that keep your edits as instructions, and leave the original alone, do this for you.
- Work in a lossless format while editing. Save working copies as PNG, TIFF, lossless WebP or lossless JPEG XL, and make a JPEG only at the end. Every one of our 20 images survived 50 PNG saves with identical pixels.
- Rotate and crop losslessly when you can. libjpeg-turbo's
jpegtranrearranges the compressed data "without ever fully decoding the image", and its manual says "its transformations are lossless: there is no image degradation at all." There are limits: a crop's top-left corner must fall on a block boundary, rotation works cleanly only when the width and height are multiples of the block size (usually 8 or 16 pixels), and-perfectmakes jpegtran stop with an error rather than do an imperfect job. - If you must re-save a JPEG, reuse its settings. Pillow's
quality="keep"keeps "the original image quality level, subsampling, and qtables", which is the same-settings case that settled in our test. - Do not save at 100 to be safe. Pillow's documentation says values above 95 should be avoided, as 100 "results in large files with hardly any gain in image quality". Our JPEG quality post measured what each setting costs.
Frequently asked questions
Does opening a JPEG reduce its quality?
No. Only saving it as a new JPEG does. Viewing, copying, moving, renaming or uploading the file as it is leaves every byte the same.
Does rotating a JPEG lose quality?
It depends on the app. A lossless rotation, like jpegtran's, moves the compressed blocks around without decoding them. Many apps instead decode, turn and save the image again, which is one more lossy save. Some only change a tag that tells viewers which way to turn it; our post on why photos come out sideways explains that tag.
How many times can you save a JPEG before you see a difference?
With the same settings and no edits, we saw no difference after 50 saves, and most files stopped changing at all after about a dozen. With a 1-pixel crop before every save, the second save already lost a median of 1.48 dB, and the damage was clearly visible by the 30th.
Is it better to save as PNG instead?
For working copies, yes: PNG is lossless, so saving it again never changes a pixel. It is usually much bigger than a JPEG, though. For keeping a smaller lossless copy, see our PNG vs WebP comparison, and AVIF vs JPEG XL vs WebP for which format to serve on the web.
Does converting a JPEG to another format lose quality?
Converting to a lossy format, such as lossy WebP or AVIF, is one more lossy save. Converting to a lossless one, such as PNG, keeps the JPEG's pixels exactly but cannot bring back what the JPEG already lost.
What Heart Photos does
Heart Photos never saves your JPEG again. The file you upload is your original: a JPEG is stored as lossless JPEG XL, and the download gives you back the very JPEG you uploaded, byte for byte. The copies visitors see are AVIF, each made from that original rather than from another copy, so they never pile up generations.
Turning a photo in the viewer saves the turn without touching the file. If you want the turn written into the pixels, Save as new makes a separate image from your untouched original, once.
To keep your originals safe in one place, open Heart Photos and upload them. Our post on the 3-2-1 backup rule covers how many copies to keep, and where.