BENCHMARK / 2026

Seven tools. Four images.
Every byte counted.

An image compression benchmark at 20 KB, 50 KB and 100 KB. Exact file sizes, dimensions and similarity metrics, with human review and the files to check them.

Collected September 30–October 1, 2026 · Publication dataset v2 · Four synthetic originals

Planned cases
84
Completed outputs
82
Unavailable
2
Tools tested
7

01 / THE SAMPLE

What we tested

Seven tools received the same four original files at each of three UI targets: 20, 50 and 100 KB. That gives 84 planned cases. These are controlled synthetic workloads, not a representative sample of camera photographs. Inputs are JPEG and PNG; this study does not test WebP inputs.

Photo-like JPEG

A synthetic photo-like scene with smooth color transitions, moderate detail and seeded grain. It tests a photographic workload without using a private photo.

JPEG · 1600 × 1200 · 763,501 bytes

Download original

Texture JPEG

Dense, repeated color texture and high-frequency detail. It makes dimension reduction and fine-detail loss visible.

JPEG · 1600 × 1200 · 3,476,856 bytes

Download original

Document / UI PNG

A synthetic document and interface with small text, a table and colored graphics. Useful for checking whether a small file remains readable.

PNG · 1280 × 900 · 522,230 bytes

Download original

Transparent PNG

A synthetic graphic with clear and partially transparent pixels. It exposes format changes, matte backgrounds and alpha-edge differences.

PNG · 1200 × 900 · 1,077,941 bytes

Download original

Why 82 completed, not 84? The original September 30 dataset had 79 downloads and five unavailable cases. Three Simple Image Resizer (SIR) 20 KB cases produced files in a documented follow-up on October 1. The normal result UI exposed Download; an advertising interstitial could be dismissed using its ordinary visible Close control. The original download-event failures remain archived. The follow-up files are separate later observations, not the missing original downloads.

Two ExactKB cases remain unavailable because the UI showed a minimum-scale refusal and offered no Download. No measurements are invented for them. The publication overlay records the version change, dates and approved follow-up hashes.

02 / THE METHOD

Original in. Actual download measured.

  1. Use the original fixture. Select the requested size through each tool’s normal UI. Leave unrelated options at their recorded defaults.
  2. Save the actual result. Measure its exact bytes and SHA-256 independently, then decode format, orientation, dimensions and alpha.
  3. Keep measurements separate. Record both byte caps, retained pixels, SSIM and RGB PSNR. Check transparency independently and inspect usability at native size.

No CAPTCHA, rate-limit or protection bypass, private API retrieval, source pre-conversion or selection of a favorable output was used. Original collection and the later SIR verification used Chrome 154.0.8037.59 on Windows. Egress country was not measured. Tool defaults and dynamic sites can change after collection.

See the complete methodology, metric definitions and dated environment record.

03 / OBSERVATIONS

Tool-by-tool results

Tools appear in the fixed study order. Coverage means completed / planned; each cap count uses completed outputs only. None of these columns is an overall quality score.

Seven tools, 12 planned cases each
ToolCompletedUnavailableWithin decimal capWithin binary equivalentResizedFormat changed
CompressMyPhotos12/12012/1212/1210/120
Pi7 Image Tool12/1208/1212/127/126
JPEG Optimizer12/1208/1212/1211/120
Simple Image Resizer12/12012/1212/1212/120
ExactKB10/1227/1010/106/100
KBSize12/1207/1212/122/126
SmallJPG12/12012/1212/1212/126

CompressMyPhotos

12 of 12 downloaded files kept their source format. 10 changed dimensions. 3 transparent-fixture outputs completed: 3 retained alpha and 0 lost it.

Pi7 Image Tool

6 of 12 downloaded files kept their source format. 7 changed dimensions. 3 transparent-fixture outputs completed: 0 retained alpha and 3 lost it.

JPEG Optimizer

12 of 12 downloaded files kept their source format. 11 changed dimensions. 3 transparent-fixture outputs completed: 3 retained alpha and 0 lost it.

Simple Image Resizer

12 of 12 downloaded files kept their source format. 12 changed dimensions. 3 transparent-fixture outputs completed: 3 retained alpha and 0 lost it.

ExactKB

10 of 10 downloaded files kept their source format. 6 changed dimensions. 2 transparent-fixture outputs completed: 2 retained alpha and 0 lost it. The two unavailable cases are documented below.

KBSize

6 of 12 downloaded files kept their source format. 2 changed dimensions. 3 transparent-fixture outputs completed: 0 retained alpha and 3 lost it.

SmallJPG

6 of 12 downloaded files kept their source format. 12 changed dimensions. 3 transparent-fixture outputs completed: 0 retained alpha and 3 lost it.

Results at 20 KB, 50 KB and 100 KB

The smallest UI target has two unavailable cases. They are excluded from cap-compliance and similarity distributions. At 20 KB, 22 of 26 completed files were resized; compare dimensions and readability before treating a small download as useful.

Target-level counts from the 82-output publication view
UI targetDecimal capBinary equivalentCompletedUnavailableWithin decimalWithin binaryResized
20 KB20,000 B20,480 B26/28223/2626/2622/26
50 KB50,000 B51,200 B28/28024/2828/2820/28
100 KB100,000 B102,400 B28/28019/2828/2818/28

Try the local 20 KB, 50 KB or 100 KB compressor with your own original. A destination may also require a particular format and minimum dimensions.

A “KB” label can mean two different limits

A decimal kilobyte is 1,000 bytes. Some software labels 1,024 bytes as “KB” (properly KiB). We entered the same numbers in each UI, then classified the actual downloads against two independent caps.

Strict decimal benchmark: 20,000 / 50,000 / 100,000 bytes.

Binary equivalents: 20,480 / 51,200 / 102,400 bytes.

66/82 completed files fit the decimal caps; 82/82 fit the binary equivalents. All 16 decimal overages fit the corresponding binary cap.

CompressMyPhotos explicitly documents decimal bytes. KBSize explicitly states “1 KB = 1024 bytes.” The other five tools have unspecified unit semantics in preserved evidence. That does not establish decimal semantics. A file that exceeded the strict decimal benchmark cap did not necessarily miss a tool’s own promised limit.

Declared units and exact supporting statements

CompressMyPhotos: decimal. Targets are best-effort upper bounds in decimal bytes (1 KB = 1,000 bytes). Preserved publisher documentation; not historical per-case UI capture.

Pi7 Image Tool: unspecified. No explicit bytes-per-KB definition found in preserved evidence. No explicit definition found in preserved evidence.

JPEG Optimizer: unspecified. No explicit bytes-per-KB definition found in preserved evidence. No explicit definition found in preserved evidence.

Simple Image Resizer: unspecified. No explicit bytes-per-KB definition found in preserved evidence. No explicit definition found in preserved evidence.

ExactKB: unspecified. No explicit bytes-per-KB definition found in preserved evidence. No explicit definition found in preserved evidence.

KBSize: binary. Sizes are measured in binary kilobytes (1 KB = 1024 bytes), the same way your operating system reports them. Captured explicit site statement.

SmallJPG: unspecified. No explicit bytes-per-KB definition found in preserved evidence. No explicit definition found in preserved evidence.

Every output exceeding the strict decimal benchmark cap; exact bytes
CaseDecimal capActual bytesOver decimalBinary capWithin binary
pi7-photo-100kb100,000100,352352102,400Yes
pi7-texture-100kb100,000100,352352102,400Yes
pi7-document-100kb100,000100,352352102,400Yes
pi7-transparent-100kb100,000100,352352102,400Yes
jpegopt-photo-20kb20,00020,23623620,480Yes
jpegopt-document-50kb50,00050,51651651,200Yes
jpegopt-transparent-50kb50,00051,0891,08951,200Yes
jpegopt-transparent-100kb100,000100,761761102,400Yes
exactkb-photo-20kb20,00020,0333320,480Yes
exactkb-photo-100kb100,000102,1192,119102,400Yes
exactkb-document-100kb100,000100,212212102,400Yes
kbsize-photo-100kb100,000102,1392,139102,400Yes
kbsize-document-50kb50,00050,47747751,200Yes
kbsize-document-100kb100,000102,2872,287102,400Yes
kbsize-transparent-20kb20,00020,26526520,480Yes
kbsize-transparent-50kb50,00050,0616151,200Yes

Smaller files can also mean fewer pixels

60 of 82 outputs changed dimensions. Pixel retention compares output width × height with the original pixel count. Width or height alone can hide how much pixel resolution was lost.

Median retained dimensions among completed outputs
UI targetSample sizeWidth retainedHeight retainedPixels retained
20 KB2647.42%47.28%22.55%
50 KB2874%74%54.77%
100 KB2893.75%93.75%87.89%
Synthetic texture original beside SIR 20 KB output: 301 by 226 pixels, 19,672 bytes, 3.543 percent of original pixels retained.
SIR texture-20KB: human review found the coarse image recognizable, with substantial loss of fine texture detail. Output enlargement is nearest-neighbor for inspection, not recovered detail.

SSIM measures similarity to a resized reference

The original reference is resized to the output’s dimensions. Both are composited over neutral gray before SSIM and RGB PSNR are calculated. A tiny output can therefore score highly against a tiny reference while losing information from the full-size original. Similarity is not legibility, usable resolution or a calibrated human-quality rating.

SSIM within each fixture and UI target; completed observations only
FixtureUI targetSSIM nMinimumMedianMaximumMedian pixels retained
Photo-like JPEG20 KB70.7459920.8689580.9766235.61%
Photo-like JPEG50 KB70.7816930.8405940.97779587.89%
Photo-like JPEG100 KB70.8043660.8435380.96654292.64%
Texture JPEG20 KB70.1456570.3539570.90618219.1%
Texture JPEG50 KB70.2080460.3184540.73711844.75%
Texture JPEG100 KB70.2226370.397860.71167681%
Document/UI PNG20 KB60.5718140.8801910.992712.31%
Document/UI PNG50 KB70.6863480.8820580.99203535.5%
Document/UI PNG100 KB70.9015040.9406130.98671287.92%
Transparent PNG20 KB60.6796010.8275650.993221%
Transparent PNG50 KB70.7194410.8526170.9937148.53%
Transparent PNG100 KB70.7437220.882180.9888766.15%

Compare observations within the same fixture, and read retained dimensions alongside SSIM. RGB PSNR and the full descriptive distributions are in the downloadable data. We do not combine these metrics into a score.

Synthetic photo-like original and CompressMyPhotos 50 KB output with exact dimensions, bytes and normalized SSIM labels.
One photo-like workload, one recorded result. Fitted previews illustrate the case; the exact original and output are available for native inspection.

PNG inputs sometimes became JPEG

18 completed PNG inputs changed to JPEG: six each from Pi7, KBSize and SmallJPG. Each format change is recorded as observed. We did not convert outputs back, repair transparency or correct filenames manually. The other tools retained the source format in their completed cases.

JPEG cannot preserve an alpha channel. A small byte count and high RGB similarity do not describe what happened to transparent pixels. For your own files, see PNG compression, JPG compression and WebP compression; this study’s inputs remain JPEG and PNG only.

Alpha is an independent observation

20 of 21 transparent-fixture cases completed. Eleven retained alpha and nine lost it. One ExactKB case is unavailable and has no transparency outcome. Alpha preservation records clear and partially transparent pixels; it does not promise an identical alpha channel. Human appearance review was consistent with the stored alpha measurements.

Transparent original beside SIR 20 KB PNG on a checkerboard; the output retains alpha at 114 by 85 pixels and 3,830 bytes.
SIR transparent-20KB: the graphic remained recognizable and transparency was visually preserved, consistent with the separate alpha measurement.
Transparent original beside Pi7 20 KB JPEG; the measured PNG-to-JPEG conversion loses alpha, so the output background is opaque.
Pi7 transparent-20KB: PNG became JPEG, with recorded alpha loss. The checkerboard belongs to the presentation; the exact JPEG’s opaque background is part of its pixels.

Under the limit. Still unreadable.

The SIR 20 KB document output met the byte limit at 3,314 bytes, measured 117 × 82 pixels and scored 0.992700 normalized SSIM. In the approved human review it was not readable as a normal document at its native output size.

Original synthetic document with small text beside an enlarged SIR 117 by 82 PNG, showing the loss of document-text readability despite high normalized SSIM.
Nearest-neighbor enlargement reveals the recorded output; it adds no detail. Meeting a file-size cap does not establish document legibility. View the native 117 × 82 output and original document.

Check the smallest text at the intended display size. A larger limit or a tighter source crop may be needed, subject to the receiving form’s rules. The compressor can report bytes and dimensions; it cannot certify readability or acceptance.

04 / THE CASE LEDGER

All 84 planned cases

Exact bytes stay visible. “Within” and “Over” refer to the indicated benchmark cap, not a tool’s promised semantics. A dash means the output was unavailable. Every row is present without JavaScript; filters only narrow this view.

2026-v2: 82 measured outputs, two unavailable
Case / toolUI targetExact bytesInput → outputOutput dimensionsPixels retainedSSIMAlphaDecimal capBinary capExact file
cmp-photo-20kbCompressMyPhotos20 KB19,994jpeg → jpeg931 × 69833.85%0.925864N/AWithinWithinDownload
cmp-photo-50kbCompressMyPhotos50 KB49,935jpeg → jpeg1600 × 1200100%0.781693N/AWithinWithinDownload
cmp-photo-100kbCompressMyPhotos100 KB95,535jpeg → jpeg1600 × 1200100%0.804366N/AWithinWithinDownload
cmp-texture-20kbCompressMyPhotos20 KB19,977jpeg → jpeg438 × 3287.48%0.768111N/AWithinWithinDownload
cmp-texture-50kbCompressMyPhotos50 KB49,768jpeg → jpeg680 × 51018.06%0.617351N/AWithinWithinDownload
cmp-texture-100kbCompressMyPhotos100 KB98,476jpeg → jpeg800 × 60025%0.711676N/AWithinWithinDownload
cmp-document-20kbCompressMyPhotos20 KB19,393png → png139 × 971.17%0.972475N/AWithinWithinDownload
cmp-document-50kbCompressMyPhotos50 KB47,699png → png254 × 1783.92%0.969532N/AWithinWithinDownload
cmp-document-100kbCompressMyPhotos100 KB81,630png → png350 × 2467.47%0.969841N/AWithinWithinDownload
cmp-transparent-20kbCompressMyPhotos20 KB19,098png → png151 × 1131.58%0.9932RetainedWithinWithinDownload
cmp-transparent-50kbCompressMyPhotos50 KB43,715png → png239 × 1793.96%0.99371RetainedWithinWithinDownload
cmp-transparent-100kbCompressMyPhotos100 KB87,494png → png343 × 2578.16%0.98887RetainedWithinWithinDownload
pi7-photo-20kbPi7 Image Tool20 KB18,432jpeg → jpeg1500 × 112587.89%0.80381N/AWithinWithinDownload
pi7-photo-50kbPi7 Image Tool50 KB49,152jpeg → jpeg1500 × 112587.89%0.840594N/AWithinWithinDownload
pi7-photo-100kbPi7 Image Tool100 KB100,352jpeg → jpeg1500 × 112587.89%0.862273N/AOverWithinDownload
pi7-texture-20kbPi7 Image Tool20 KB18,432jpeg → jpeg1500 × 112587.89%0.145657N/AWithinWithinDownload
pi7-texture-50kbPi7 Image Tool50 KB49,152jpeg → jpeg1500 × 112587.89%0.310132N/AWithinWithinDownload
pi7-texture-100kbPi7 Image Tool100 KB100,352jpeg → jpeg1500 × 112587.89%0.4543N/AOverWithinDownload
pi7-document-20kbPi7 Image Tool20 KB18,432png → jpeg922 × 64851.86%0.827163N/AWithinWithinDownload
pi7-document-50kbPi7 Image Tool50 KB49,152png → jpeg1280 × 900100%0.873099N/AWithinWithinDownload
pi7-document-100kbPi7 Image Tool100 KB100,352png → jpeg1280 × 900100%0.930086N/AOverWithinDownload
pi7-transparent-20kbPi7 Image Tool20 KB18,432png → jpeg1200 × 900100%0.679601LostWithinWithinDownload
pi7-transparent-50kbPi7 Image Tool50 KB49,152png → jpeg1200 × 900100%0.719441LostWithinWithinDownload
pi7-transparent-100kbPi7 Image Tool100 KB100,352png → jpeg1200 × 900100%0.743722LostOverWithinDownload
jpegopt-photo-20kbJPEG Optimizer20 KB20,236jpeg → jpeg955 × 71635.61%0.868958N/AOverWithinDownload
jpegopt-photo-50kbJPEG Optimizer50 KB46,857jpeg → jpeg1302 × 97666.19%0.866626N/AWithinWithinDownload
jpegopt-photo-100kbJPEG Optimizer100 KB80,750jpeg → jpeg1444 × 108381.45%0.866325N/AWithinWithinDownload
jpegopt-texture-20kbJPEG Optimizer20 KB19,342jpeg → jpeg700 × 52419.1%0.353957N/AWithinWithinDownload
jpegopt-texture-50kbJPEG Optimizer50 KB44,417jpeg → jpeg907 × 68032.12%0.260463N/AWithinWithinDownload
jpegopt-texture-100kbJPEG Optimizer100 KB98,221jpeg → jpeg1115 × 83648.55%0.222637N/AWithinWithinDownload
jpegopt-document-20kbJPEG Optimizer20 KB19,423png → png477 × 33513.87%0.571814N/AWithinWithinDownload
jpegopt-document-50kbJPEG Optimizer50 KB50,516png → png763 × 53635.5%0.686348N/AOverWithinDownload
jpegopt-document-100kbJPEG Optimizer100 KB92,172png → png1280 × 900100%0.985833N/AWithinWithinDownload
jpegopt-transparent-20kbJPEG Optimizer20 KB19,062png → png613 × 45826%0.897789RetainedWithinWithinDownload
jpegopt-transparent-50kbJPEG Optimizer50 KB51,089png → png836 × 62748.53%0.887441RetainedOverWithinDownload
jpegopt-transparent-100kbJPEG Optimizer100 KB100,761png → png976 × 73266.15%0.88218RetainedOverWithinDownload
sir-photo-20kbSimple Image Resizer20 KB18,229jpeg → jpeg301 × 2263.54%0.974707N/AWithinWithinDownload
sir-photo-50kbSimple Image Resizer50 KB40,388jpeg → jpeg476 × 3578.85%0.977795N/AWithinWithinDownload
sir-photo-100kbSimple Image Resizer100 KB74,676jpeg → jpeg673 × 50517.7%0.966542N/AWithinWithinDownload
sir-texture-20kbSimple Image Resizer · Oct 1 follow-up20 KB19,672jpeg → jpeg301 × 2263.54%0.906182N/AWithinWithinDownload
sir-texture-50kbSimple Image Resizer50 KB34,327jpeg → jpeg476 × 3578.85%0.737118N/AWithinWithinDownload
sir-texture-100kbSimple Image Resizer100 KB79,151jpeg → jpeg673 × 50517.7%0.61413N/AWithinWithinDownload
sir-document-20kbSimple Image Resizer · Oct 1 follow-up20 KB3,314png → png117 × 820.83%0.9927N/AWithinWithinDownload
sir-document-50kbSimple Image Resizer50 KB7,953png → png186 × 1302.1%0.992035N/AWithinWithinDownload
sir-document-100kbSimple Image Resizer100 KB14,047png → png263 × 1854.22%0.986712N/AWithinWithinDownload
sir-transparent-20kbSimple Image Resizer · Oct 1 follow-up20 KB3,830png → png114 × 850.9%0.986783RetainedWithinWithinDownload
sir-transparent-50kbSimple Image Resizer50 KB7,047png → png180 × 1352.25%0.985935RetainedWithinWithinDownload
sir-transparent-100kbSimple Image Resizer100 KB12,443png → png254 × 1914.49%0.985189RetainedWithinWithinDownload
exactkb-photo-20kbExactKB20 KB20,033jpeg → jpeg1600 × 1200100%0.749249N/AOverWithinDownload
exactkb-photo-50kbExactKB50 KB49,936jpeg → jpeg1600 × 1200100%0.781693N/AWithinWithinDownload
exactkb-photo-100kbExactKB100 KB102,119jpeg → jpeg1600 × 1200100%0.8071N/AOverWithinDownload
exactkb-texture-20kbExactKB20 KB17,520jpeg → jpeg936 × 70234.22%0.326577N/AWithinWithinDownload
exactkb-texture-50kbExactKB50 KB46,936jpeg → jpeg1198 × 89856.03%0.407842N/AWithinWithinDownload
exactkb-texture-100kbExactKB100 KB94,518jpeg → jpeg1600 × 1200100%0.39786N/AWithinWithinDownload
exactkb-document-20kbExactKB20 KB—png → ————N/A——Unavailable
exactkb-document-50kbExactKB50 KB46,686png → png315 × 2226.07%0.828524N/AWithinWithinDownload
exactkb-document-100kbExactKB100 KB100,212png → png478 × 33613.94%0.901504N/AOverWithinDownload
exactkb-transparent-20kbExactKB20 KB—png → ———————Unavailable
exactkb-transparent-50kbExactKB50 KB48,411png → png240 × 1804%0.852617RetainedWithinWithinDownload
exactkb-transparent-100kbExactKB100 KB95,348png → png346 × 2598.3%0.922976RetainedWithinWithinDownload
kbsize-photo-20kbKBSize20 KB19,196jpeg → jpeg1600 × 1200100%0.745992N/AWithinWithinDownload
kbsize-photo-50kbKBSize50 KB49,935jpeg → jpeg1600 × 1200100%0.781693N/AWithinWithinDownload
kbsize-photo-100kbKBSize100 KB102,139jpeg → jpeg1600 × 1200100%0.807099N/AOverWithinDownload
kbsize-texture-20kbKBSize20 KB17,158jpeg → jpeg1517 × 113889.91%0.146325N/AWithinWithinDownload
kbsize-texture-50kbKBSize50 KB32,717jpeg → jpeg1600 × 1200100%0.208046N/AWithinWithinDownload
kbsize-texture-100kbKBSize100 KB94,532jpeg → jpeg1600 × 1200100%0.397851N/AWithinWithinDownload
kbsize-document-20kbKBSize20 KB19,287png → jpeg1059 × 74468.39%0.811689N/AWithinWithinDownload
kbsize-document-50kbKBSize50 KB50,477png → jpeg1280 × 900100%0.882058N/AOverWithinDownload
kbsize-document-100kbKBSize100 KB102,287png → jpeg1280 × 900100%0.935179N/AOverWithinDownload
kbsize-transparent-20kbKBSize20 KB20,265png → jpeg1200 × 900100%0.684076LostOverWithinDownload
kbsize-transparent-50kbKBSize50 KB50,061png → jpeg1200 × 900100%0.72518LostOverWithinDownload
kbsize-transparent-100kbKBSize100 KB98,789png → jpeg1200 × 900100%0.771316LostWithinWithinDownload
smalljpg-photo-20kbSmallJPG20 KB14,111jpeg → jpeg460 × 3458.27%0.97662N/AWithinWithinDownload
smalljpg-photo-50kbSmallJPG50 KB43,173jpeg → jpeg1170 × 87853.5%0.901059N/AWithinWithinDownload
smalljpg-photo-100kbSmallJPG100 KB93,744jpeg → jpeg1540 × 115592.64%0.843538N/AWithinWithinDownload
smalljpg-texture-20kbSmallJPG20 KB13,448jpeg → jpeg400 × 3006.25%0.800408N/AWithinWithinDownload
smalljpg-texture-50kbSmallJPG50 KB43,644jpeg → jpeg1070 × 80344.75%0.318454N/AWithinWithinDownload
smalljpg-texture-100kbSmallJPG100 KB94,146jpeg → jpeg1440 × 108081%0.338159N/AWithinWithinDownload
smalljpg-document-20kbSmallJPG20 KB18,510png → jpeg420 × 29510.76%0.933219N/AWithinWithinDownload
smalljpg-document-50kbSmallJPG50 KB47,546png → jpeg1090 × 76672.48%0.90176N/AWithinWithinDownload
smalljpg-document-100kbSmallJPG100 KB93,351png → jpeg1200 × 84487.92%0.940613N/AWithinWithinDownload
smalljpg-transparent-20kbSmallJPG20 KB18,893png → jpeg480 × 36016%0.757341LostWithinWithinDownload
smalljpg-transparent-50kbSmallJPG50 KB43,284png → jpeg1140 × 85590.25%0.733976LostWithinWithinDownload
smalljpg-transparent-100kbSmallJPG100 KB93,947png → jpeg1180 × 88596.69%0.757431LostWithinWithinDownload

The two unavailable cases

exactkb-document-20kb and exactkb-transparent-20kb showed a visible warning that the target could not be reached within the default minimum scale. No Download was offered. They were not rerun; byte size, dimensions, format, alpha and similarity remain null.

These are unavailable observations for this dated workflow, not zero-byte successes or measured over-cap files. Their reasons are in unavailable-cases.csv.

What this study cannot establish

  • Four synthetic images and two input formats do not represent all photos, documents, devices or encoders.
  • Tool defaults differ. Sites can change. Three SIR outputs have a later collection date, explicitly retained in the data.
  • Resized-reference SSIM and RGB PSNR are not readability, human quality ratings or exhaustive alpha-appearance tests.
  • Historical CompressMyPhotos case screenshots and browser-suggested filenames were not captured. That gap remains disclosed.
  • Pi7 has repeated output-byte patterns; CMP and KBSize photo-50KB share a SHA-256 and 49,935-byte length. Reused filenames are also recorded. These are factual observations; their causes are not established.
  • Marketing claims and limited network observations do not establish processing location or retention. This page makes no competitor privacy comparison.

05 / OPEN TO INSPECTION

Data, files and reproduction

Publication version 2026-v2 contains the same approved bytes used for the tables. Original fixtures are synthetic and CC0. Outputs are derived from those originals; no private user photos are included. Article graphics are labeled derivatives, separate from exact files.

Use Node and Sharp 0.35.4 as described in the README. The script verifies exported hashes, decodes all 82 outputs, reproduces measurements and checks CSV/JSON parity. Public download files are separate from page rendering; none of the raw image archive is preloaded.

Return to the image compressor · Inspect and remove image metadata