Run a reproducible 12-scene quality matrix, 24-scene matched-quality ladder or balanced 100-scene PNG, JPEG, WebP and optional AVIF benchmark locally in your browser.
Direct answer · browser image benchmark png jpeg webp
What does the browser image benchmark measure?
It generates deterministic scenes locally and measures encoded file size, repeated browser encode time, sampled PSNR and transparency behavior for PNG, JPEG, WebP and supported AVIF.
Research mode: the 100-scene browser image benchmark balances photo-like detail, interface/text, transparent graphics and flat illustrations.
Evidence on this page
The methodology, environment, raw exports and known limitations are published with the tool.
Important limit
Synthetic scenes, one device and sampled PSNR do not establish a universal format winner.
Published reference run · 30 July 2026
Verified results from Brave 150 on Windows
These static results come from the exported v2.1 JSON reports, not from a marketing estimate. The full run measured 100 balanced deterministic scenes. The matched run sampled five quality levels across 24 scenes. Both used one warm-up and three measured encodes per output.
100balanced scenes in the full run78.2%median WebP saving versus PNG at quality 8257.1%smaller WebP files at near-matched sampled PSNR across 18 opaque scenes17.0 msmedian WebP encode time versus 5.8 ms for JPEG on this device
100-scene results by category
Scene category
PNG median
JPEG 82 saving
WebP 82 saving
Interpretation
Photo-like detail
527.9 KB
90.2%
92.8%
Both lossy formats compared with the same-scene PNG baseline.
Interface and text
10.1 KB
-95.9%
54.1%
Both lossy formats compared with the same-scene PNG baseline.
Transparent graphic
266.4 KB
91.2%
83.6%
JPEG flattened alpha; WebP preserved it.
Flat illustration
55.4 KB
51.5%
72.4%
Both lossy formats compared with the same-scene PNG baseline.
Matched-quality JPEG versus WebP
For opaque photo-like, interface/text and flat-illustration scenes, the closest sampled PSNR pairs had a 0.07 dB median gap. At that near-matched sampled quality, WebP files were 57.1% smaller in the median. Transparent scenes are excluded from this headline because JPEG removed alpha while WebP preserved it.
Reference hashes: full run 7ecff7e69b8c0e1b…; matched run ccd7755cf5efbeb8….
Measure this browser—not a hidden server
Every scene is generated from fixed drawing rules and deterministic seeds. Each output receives one warm-up encode and three measured encodes; the reported time is their median. Quick mode compares three quality levels across 12 scenes, matched mode uses five levels across 24 scenes, and full mode measures 100 balanced scenes at quality 82.
Benchmark interface 20260731-3 ready. No benchmark has been run yet.
Median values reduce the effect of a single unusually easy or difficult scene. Timing is the median of three measured encodes after one warm-up. Savings are measured against the PNG generated from the same scene.
Format
Quality
Outputs
Median size
Median encode time
Average sampled PSNR
Median saving vs PNG
Noisy timings
Results by scene category
Category summaries prevent photo-like scenes from hiding different behavior in interface graphics, transparency and flat illustrations.
Category
Format
Quality
Outputs
Median size
Median encode time
Average sampled PSNR
Median saving
Approximate JPEG versus WebP comparisons
The tool selects the closest available sampled PSNR pair and the closest available file-size pair for each scene. These comparisons are limited to the quality values in the selected mode and are not claims that encoder quality controls are equivalent.
Method
Scene pairs
Closeness
Median result
Raw measurements
The page displays up to 240 rows to protect responsiveness. Every row is included in CSV, JSON and standalone HTML exports, including all three timing samples and timing-noise flags.
Scene
Category
Format
Quality
Dimensions
File size
Median encode time and range
PSNR
Saving
Notes
Balanced scene corpus
The full run creates 25 scenes in each category: photo-like detail, interface and text shapes, transparent graphics and flat illustrations. Dimensions vary within each category.
Repeated timing
One warm-up encode is discarded. Three timed encodes are retained and their median is reported; the full timing range and robust noise flag are exported.
Lossless baseline
PNG is shown as an infinite/lossless PSNR baseline rather than using an artificial finite value such as 99 dB.
What this proves
It records Blob size, repeated browser encoding time, sampled pixel fidelity and transparency behavior produced by the current browser and device.
What this does not prove
Synthetic scenes and discrete quality samples do not represent every production image or establish that one format or encoder is universally superior.
How to reproduce it
Run the same mode in another browser, export JSON or HTML and compare the environment block, benchmark version, timing samples and raw scene identifiers.