Image optimization: size, speed, and visual quality

Source files

The optimization study uses 30 files from Wikimedia Commons: for each of JPEG, PNG, GIF, WebP, and SVG, the first six search results inside a fixed file-size band that Commons labels CC0 or public domain. The re-encoding study uses the 24 lossless PNG photographs of the Kodak image set, a common reference set for codec comparisons. The corpus manifest lists every file with its URL, license, dimensions, byte size, and SHA-256 checksum, together with the exact search queries, so the same bytes can be fetched again.

Assemblies

Every measurement is one Assembly that uploads one file and runs one Step. The optimization study varies priority and metadata handling for JPEG, lossy optimization and metadata handling for PNG, and metadata handling for WebP and SVG. Parameters that the Robot ignores for a format are left out, because they would repeat an identical job. The upload Step sets output_meta to false: no Step here reads the metadata of the uploaded file, and extracting it would add to every Assembly duration without changing any result. These are the Assembly Instructions, with the varied parameters shown at one of their values:

{
  "steps": {
    ":original": {
      "robot": "/upload/handle",
      "output_meta": false
    },
    "measured": {
      "robot": "/image/optimize",
      "use": ":original",
      "priority": "compression-ratio",
      "preserve_meta_data": true
    }
  }
}

{
  "steps": {
    ":original": {
      "robot": "/upload/handle",
      "output_meta": false
    },
    "measured": {
      "robot": "/image/resize",
      "use": ":original",
      "format": "webp",
      "quality": 75,
      "strip": true
    }
  }
}

Runs and environment

All Assemblies ran on October 17, 2026 against the production API at api2.transloadit.com, on the same shared machines that serve customer traffic. The full matrix of 294 file and parameter combinations was executed 3 times in sequence, so repeats of one combination are spread over the session, with 3 Assemblies in flight at a time.

Measurements and aggregation

Sizes and timings come from the Assembly Status: the upload size, the result size, the result’s Robot execution and queue time, and the Assembly execution duration. For SSIM, the result of the first run is downloaded, both images are flattened onto white and converted to 8-bit luma, and structural similarity is averaged over 8 × 8 pixel windows at a stride of 4 pixels. Output bytes were identical across the repeated runs of every combination, so only timings vary between repeats. Each table row first takes the median of the repeated runs per file, then the median across files. The 90th percentile is taken over all runs in the row.