Every claim here comes from reproducible inputs, outputs, and implementation facts — no guesses about TinyPNG's private algorithms. Each comparison measures its black-box output on a stated date.
Default JPEG output sits at essentially the same size–quality point as TinyPNG: across 30 public photos it is 0.30% larger in total, with a similar median. MediaSlim's lowest S2 is 74.47 versus 68.64 for TinyPNG — a steadier per-image quality floor.
Across 8 mixed-content images in Simplified Chinese, Traditional Chinese, English, and Spanish, default output is 11.03% smaller in total with median S2 0.55 higher; 7 of 8 match or beat TinyPNG image by image.
PNG offers higher compression coverage: 66.4% saved across 79 images vs 39.5%. Excluding the 9 TinyPNG returns as-is, sizes are roughly level and MediaSlim's median perceptual quality is slightly higher.
Privacy and failure behaviour are the certain wins: fully local, originals not overwritten, lossless truecolor fallback when risk is high. Text / UI screenshots remain a clear weak spot.
Default JPEG: level on size, steadier at the quality floor
JPEG quality is an encoder parameter, not final visual quality: one q72 across 30 photos ranges from S2 58.7 to 76.6. The desktop path therefore runs a per-image loop — encode → SSIMULACRA2 score → adjust quality — and returns the smallest file that meets its quality floor. The table compares each product's default output across 30 public photos; it is not the experiment that makes MediaSlim chase TinyPNG's score image by image.
MetricMediaSlimTinyPNG
Total bytes9,030,3309,003,755
Relative size+0.30%baseline
S2 median / P0575.36 / 74.8175.69 / 69.43
Lowest S274.4768.64
SSIM median / P050.9575 / 0.91950.9647 / 0.8973
Banding median / P95 (lower better)0.630 / 1.3650.738 / 1.217
Total size differs by just +0.30% and median S2 by −0.33, putting the two defaults at a similar operating point
MediaSlim's S2 P05 is 74.81 and its minimum is 74.47; TinyPNG measures 69.43 and 68.64 respectively
20/30 are smaller and 15/30 match or beat TinyPNG on S2; this is not a claim that every image is both smaller and clearer
A separate experiment dated 2026-07-28 first measures TinyPNG's S2 for each image, then has MediaSlim search for that same target. Across 30 frozen public photos from Kodak, CLIC 2020, and DIV2K, MediaSlim is 3.26% smaller in total. That figure demonstrates efficiency for this corpus and method only; it is not the product's current default output and does not generalize to every JPEG.
Mixed content: raise the floor when edge risk appears
When AV1-style block risk and local edge loss in the q68 probe are detected, MediaSlim raises the S2 floor to 81 within the same search. The table covers default output for 8 guide images in Simplified Chinese, Traditional Chinese, English, and Spanish.
MetricMediaSlimTinyPNG
Total bytes729,869820,351
Relative size−11.03%baseline
S2 median / P0576.78 / 76.2476.23 / 74.87
SSIM median / P050.98970 / 0.989220.98993 / 0.98968
Median S2 is 0.55 higher; 7 of 8 images match or beat TinyPNG, with the remaining image 0.65 lower
Total output is 90,482 bytes smaller (−11.03%); automatic output hashes match an explicit floor 81 on all 8 images
PNG first tries a ≤256-color indexed PNG and inspects the actual output for perceptual banding; where risk is found it adds local gradient colors, rebuilds the index and dithers. The quality tier outputs a lossless truecolor PNG when serious risk remains. The table is the default delivery across 79 real-world images.
MetricMediaSlimTinyPNG
Total output55.8 MB100.4 MB
Saved vs original66.4%39.5%
S2 median / P0578.4 / 67.079.7 / 66.0
SSIM median / P050.9946 / 0.97390.9943 / 0.9616
Banding median / P95 (lower better)0.52 / 1.480.56 / 2.26
The total gap includes product policy: TinyPNG returns 9/79 as-is (declining to compress to preserve quality), while MediaSlim still compresses the same batch 58–90%. Excluding those 9, MediaSlim is only 1.9% larger with a +0.81 paired-median S2 — sizes level, quality slightly ahead.
Results vary by content type
The same PNG batch, broken down by content type. The total-size lead comes mostly from line art and icons — it doesn't generalize to every category. Text / UI screenshots are the clearest weak spot.
Content typenRel. sizeS2 wins
Pet furUse as-is
6+0.3%6/6
Flat skyOK, spot-check gradients
4−5.8%4/4
High textureUse as-is
6+3.4%5/6
Photo + textCheck text edges
6+6.3%5/6
Low-lightGenerally OK
6+6.3%4/6
FacesReview key portraits
6−28.0%3/6
Text / UI shotsUse quality tier or keep original
6+0.1%0/6
Line art / iconsincl. 9 TinyPNG returned as-is
39−60.3%16/39
Four things to know
Default JPEG does not win every image.
Of the 30 public photos, 20 are smaller, 15 match or beat TinyPNG on S2, and only 5 do both. The value is a tighter quality floor across images, not guessing the competitor's target for each one.
This is not an encoder-efficiency blowout.
On a specified corpus at matched quality, the pure encoder-efficiency edge is only about 5%. Most of the product value comes from choosing quality per image and enforcing an explicit floor.
Text / UI screenshots still lag.
0/6 S2 wins on the sample; the typical difference is dithered fringing on colored icon edges. For design review and pixel-level UI QA, use the quality tier or keep the original.
A snapshot, not a standing claim.
TinyPNG is an evolving cloud black box; these numbers match the frozen output on the measurement date. The repo records filename maps and SHA-256 for independent replay.
How it was measured
1JPEG default output: send the same 30 lossless-source Kodak, CLIC 2020, and DIV2K photos to each product and compare each product's default file size, S2, SSIM, and banding risk.
2JPEG specified-corpus matched quality: measure TinyPNG's output S2, have MediaSlim search for the same target, then compare size. This experiment is reported separately from default output.
3Mixed-content default output: use 8 multilingual guide images to validate the automatic block-risk tier and compare each image with the frozen TinyPNG output.
4PNG default delivery: the same 79 real-world images to each product; compare default output size, S2, SSIM and banding risk.
5Cross-checked metrics: S2 primary, with SSIM, PSNR and a banding score, plus a manual flicker blind test.
6Input files, output hashes and evaluation scripts live in the repo for independent replay.
Check these numbers on your own images
All features are free. Processing stays on-device, originals are kept, and MediaSlim flags any result that isn't actually smaller.