Finding 1
WebP inputs had the largest observed mean reduction
Across priority and metadata settings, the two WebP inputs averaged a 81.2% reduction. The JPEG inputs averaged 12.3%. These are averages for this small corpus, not expected savings for every file of that type.
- WebP
- 81.2%
- 2.62 sec
- GIF
- 27.1%
- 1.55 sec
- SVG
- 18.2%
- 9.34 sec
- PNG
- 17.1%
- 14.17 sec
- JPEG
- 12.3%
- 4.14 sec
Finding 2
The priority tradeoff appeared when metadata was retained
With metadata retention enabled, compression-ratio averaged a 44.6% reduction and 6.64 sec of execution. Conversion-speed averaged a 25.6% reduction and 5.72 sec. With metadata removed, the average size-reduction difference narrowed to 0.3 percentage points.
Compression-ratio
44.6%
mean reduction · metadata kept
Conversion-speed
25.6%
mean reduction · metadata kept
Finding 3
Output size varied sharply across JPEG quality settings
Quality 25 reduced size by 77.6% on average, quality 50 by 60.6%, and quality 75 by 31.5%. At quality 100, outputs were 126.4% larger than inputs on average. Mean execution ranged from 2.02 sec to 3.14 sec, but the single-run timings are too noisy to establish a clean speed trend.
- Quality 25
- 77.6%
- Quality 50
- 60.6%
- Quality 75
- 31.5%
- Quality 100
- -126.4%
Method
How the measurements were collected and recalculated
The source repository contains 12 test images and 3 Python scripts. Each script creates a Transloadit Assembly, waits for completion, and records input size, output size, and the Assembly execution_duration value.
The format dataset contains 48 runs: 12 files multiplied by 2 priorities and 2 metadata settings. The quality dataset contains 40 runs: 10 files multiplied by 4 JPEG quality values. The factorial dataset contains 48 runs across 4 images, 3 quality values, 2 priorities, and 2 metadata settings.
For this report, mean size reduction is calculated as 100 × (1 − output size ÷ input size). Positive values mean a smaller output; negative values mean the output grew. Means are unweighted across rows, so every run contributes equally regardless of input size.
Read before generalizing
This is an exploratory historical benchmark
- The corpus contains twelve files, and the source does not document a sampling method for representing production image traffic.
- Each configuration appears once per source file, so timing variance and confidence intervals cannot be estimated.
- The CSV files do not record run date, region, Robot stack version, hardware, or Assembly identifiers.
- The results describe the 2022 experiment. They should not be treated as a benchmark of current Transloadit stacks.
- No perceptual metric such as SSIM or VMAF was recorded, so smaller files cannot be ranked by visual fidelity from this dataset alone.