Repeated JPEG Compression vs Exporting from the Original: 1, 2, 5 and 10 Saves

Three distinctions to make first

Two charts, fixed dimensions and three save paths

1. 1

   On September 26, 2026, we generated two 1,200×800 PNG charts in Chrome 153 on Windows. One has small text and lines; the other includes deterministic noise and gradients. This was not an evaluation of a real-photo collection.

2. 2

   A separate local experiment page reproduced Toolit's manual JPEG quality processing using browser Canvas. Resizing, cropping, rotation and target-size mode were excluded. It was not an end-to-end test of every tool UI step.

3. 3

   The repeated path decoded the previous JPEG and saved it at quality 80, recording saves 1, 2, 5 and 10. The original path made fresh quality-80 exports from the same PNG at each checkpoint. We also tested 80→60→80.

4. 4

   We measured file size, byte identity and mean absolute RGB difference from the PNG source. The difference is averaged over all color channels and pixels on a 0–255 scale; lower is closer to the reference. It is not a quality percentage or a human preference score. Identical crops were enlarged three times for inspection.

The same quality 80 still drifted further from the source

Sizes use 1KB=1,000 bytes. Text and texture have different characteristics: compare save paths within each chart, not the absolute scores as a quality ranking between the charts.

• The text chart grew and then shrank; the texture file shrank slightly. Saving again does not guarantee a smaller file.
• Differences increased at the measured checkpoints, but not by a fixed amount per save. We cannot extend this pattern to every image or beyond ten saves.
• JPEG discards some information during encoding. Feeding the previous output back in makes the next file from already-changed pixels. Simply copying, renaming or viewing a file does not itself perform this recompression.

What happened when each export started from the original?

Fresh quality-80 exports from the source PNG were byte-identical at checkpoints 1, 2, 5 and 10. The text chart remained 88,877 bytes with difference 1.0946; the texture chart remained 245,156 bytes with difference 4.9885. Earlier output changes did not accumulate because those outputs were not used as inputs.

In Toolit, changing quality and compressing again while keeping the same selected input rereads that original input. Moving the slider repeatedly is not the ten-save chain in this experiment. Downloading a result and loading that file as the next input creates the repeated-output path.

Our reference is a generated PNG. A user's camera-original JPEG may already contain JPEG loss. Keeping the earliest original avoids additional generations; it does not reconstruct information lost during capture or earlier encoding.

Did lowering to 60 and returning to 80 undo the loss?

For both charts, 80→60→80 differed more from the source than a direct quality-80 export. A higher quality setting describes how to save the current input; it does not recover detail discarded by previous saves.

• A smaller result is not automatically a more efficient file of equal quality. Here, smaller files came with larger differences from the source.
• If only a low-quality output remains, raising its quality setting cannot accurately restore the original letters or texture. Return to the first received file or camera original when available.

Inspect the enlarged crops, but also check the intended viewing size

Each comparison runs left to right: source PNG, one quality-80 save, ten quality-80 saves, and 80→60→80. The top crops show small text; the lower crops show colored lines and surrounding texture. Identical regions are enlarged three times without smoothing.

These outputs show changes around small letters and increased spreading around red and blue lines after repeated saves. A difference visible in an enlarged crop is not necessarily equally noticeable in a small phone preview. Check the image at its actual intended size too.

• View enlarged text and line comparisons
• View enlarged texture comparisons

A better workflow for meeting a file-size limit

Keep originals and submission outputs in separate folders so it is clear which file to reuse. The goal is not to insist on maximum quality; it is to compare candidate outputs made from the same starting point.

Reproduction files and what this experiment cannot establish

These are two controlled charts in one browser. We did not measure damage across portraits and landscapes, all JPEG encoders, or a universal number of saves that makes an image unusable. Other encoders or inputs may change less or reach a plateau under particular conditions.

Mean color difference does not assess shifts, color-management differences or perceptual importance. Local problems around small letters can be hidden by an image-wide average, so the enlarged crops accompany the numbers. Download the inputs and measurements to repeat the same-size, same-quality comparison.

• Original text-chart PNG
• Original texture-chart PNG
• Full measurements (JSON)