Transloadit Research

# Measurements for media decisions that matter

We publish the measurements behind practical choices in upload reliability, encoding quality, file size, latency, cost, and compute. Every report separates what the data shows from what it cannot prove.

Open methodsDownloadable dataExplicit limitations

Published report

August 2022 dataset

## Image optimization: speed, size, and quality

A transparent re-analysis of 136 recorded Assembly runs across optimization priority, metadata retention, source format, and JPEG quality.

136

recorded runs

3

open CSV datasets

12

source files

[Read the benchmark report →](/research/image-optimization.md)

What the first dataset shows

Largest format reduction

81.2%

Average reduction for the two WebP inputs across priority and metadata settings.

Quality 25 reduction

77.6%

Average reduction at quality 25 across ten JPEG and PNG source files.

Research program

## Questions worth measuring, not guessing

Our research program focuses on decisions where a benchmark can replace a vague best practice with a measurable tradeoff.

Video delivery

### HLS encoding time, quality, and cost

Measure wall-clock encoding time, visual quality, output size, and processing cost for practical adaptive-streaming ladders.

Codec efficiency

### VMAF and SSIM versus file size

Compare perceptual quality and structural similarity against bytes delivered for modern and established codecs.

Upload reliability

### Uppy and tus on unreliable networks

Test completion rate, transferred bytes, and recovery time through packet loss, disconnects, and browser restarts.

Format coverage

### Document preview compatibility

Run a documented corpus through preview pipelines and report successful rendering, fidelity issues, and failure classes.

Recognition quality

### OCR and transcription accuracy

Score text and speech outputs against reviewed references, broken down by language, media conditions, and error type.

Delivery architecture

### Preprocessing versus Smart CDN

Compare first-request latency, cache behavior, bytes processed, and cost for eager and on-demand transformations.

Compute efficiency

### Savings from avoiding redundant work

Quantify avoided compute and estimated emissions when derivatives are reused instead of transformed again.

Publication standard

## Useful evidence needs context

Reproducible inputs

Publish the corpus or describe how another team can reconstruct it.

Named environment

Record relevant stack versions, parameters, regions, and run dates.

Comparable units

Define every metric and keep the raw observations available for re-analysis.

Limits in plain sight

State sample limits, missing controls, and when results should not be generalized.