Executive Summary
Document type dominated the result. A text-first PDF fell only 16.2–16.7% at every setting. Image-heavy fixtures fell 83.6–99.7%, but the smallest outputs also had the lowest raster dimensions and most visible detail loss. All 12 outputs kept their page counts, and all existing text layers were retained exactly.
- Text-first: use Light. Maximum provided no meaningful extra saving.
- Scanned pages: Balanced cut 7.23 MB to 571 KB; Maximum reached 192 KB but visibly softened small text.
- Photos: Light retained the most detail; Balanced removed deliberate high-resolution excess; Maximum discarded fine grain.
- Slides: vector labels stayed exact at every level, while raster artwork became progressively softer.
File-size reduction by document type
Each bar is the measured percent reduction from the original fixture. Direct labels and the complete table below provide the exact values.
Source: GoPDFConverter PDF Compression Preset Benchmark CSV, production build v84, July 10, 2026. These intentionally oversized synthetic fixtures are stress cases, not promised reductions for other PDFs.
Decision aid
Which compression preset should I use?
All 12 benchmark results
| Fixture | Preset | Original | Output | Reduction | Output max image | First-page PSNR | Validation |
|---|---|---|---|---|---|---|---|
| Text-first | Light | 12,598 B | 10,555 B | 16.2% | No images | Pixel-identical | 8 pages; exact text |
| Balanced | 12,598 B | 10,496 B | 16.7% | No images | Pixel-identical | 8 pages; exact text | |
| Maximum | 12,598 B | 10,496 B | 16.7% | No images | Pixel-identical | 8 pages; exact text | |
| Scan-first | Light | 7.23 MB | 1.18 MB | 83.6% | 1551 × 2007 | 29.87 dB | 4 pages; image-only |
| Balanced | 7.23 MB | 571 KB | 92.1% | 1261 × 1632 | 26.85 dB | 4 pages; image-only | |
| Maximum | 7.23 MB | 192 KB | 97.3% | 838 × 1084 | 22.79 dB | 4 pages; image-only | |
| Photo-heavy | Light | 22.97 MB | 1.66 MB | 92.8% | 2007 × 1368 | 31.58 dB | 4 pages; exact text |
| Balanced | 22.97 MB | 417 KB | 98.2% | 1632 × 1112 | 29.87 dB | 4 pages; exact text | |
| Maximum | 22.97 MB | 65 KB | 99.7% | 1084 × 739 | 29.82 dB | 4 pages; exact text | |
| Slide deck | Light | 5.12 MB | 585 KB | 88.6% | 1460 × 821 | 42.57 dB | 6 pages; exact text |
| Balanced | 5.12 MB | 264 KB | 94.8% | 1187 × 667 | 37.89 dB | 6 pages; exact text | |
| Maximum | 5.12 MB | 105 KB | 97.9% | 789 × 444 | 37.17 dB | 6 pages; exact text |
Download all measured fields as CSV
What changed as files got smaller?
Text-first PDFs reached a structural floor
The 8-page text fixture contained selectable text and vector tables but no raster images. Light removed 16.2%; Balanced and Maximum removed 16.7%. All three rendered pixel-for-pixel like the source. Choosing Maximum cannot create large savings when there are no large images to downsample.
Scans showed the clearest quality tradeoff
Light produced 1551 × 2007 page images and the highest measured similarity. Balanced produced 1261 × 1632 images and remained readable in the synthetic invoice inspection. Maximum fell to 838 × 1084 and PSNR dropped to 22.79 dB; small characters were visibly softer. For OCR, fine print, or reprinting, use Light or test Balanced carefully.
Vector slide content stayed sharp
The slide fixture mixed a raster hero image with vector titles, charts, and labels. Raster artwork softened at stronger settings, but exact text extraction, page count, and vector labels survived all three presets. That separation is why slides can shrink substantially without making every element blurry.
Fixtures and reproducibility
All source documents were generated deterministically with seed 20260710. They contain no personal, customer, or copyrighted source material.
| Fixture | Construction | Pages | Source SHA-256 |
|---|---|---|---|
| Text-first | Selectable text and vector tables | 8 | 6cbc7a14…9db63f4 |
| Scan-first | Four image-only synthetic document scans | 4 | 9c769d51…e98d49f |
| Photo-heavy | High-resolution synthetic photographs plus vector captions | 4 | 73ab4096…1780d8a |
| Slide deck | 16:9 mixed vector and raster slides | 6 | ecb5f67f…02b400c |
Method
- Generate controlled inputs.
Four deterministic PDFs isolate text/vector, scanned-page, photographic, and mixed slide content.
- Run the production algorithm.
Chrome executed the same v84 PDF-LIB traversal, MozJPEG encoder, preset definitions, metadata behavior, and output fallback used by Compress PDF. Light used JPEG quality 82, Balanced 72, and Maximum 55, with progressively lower dimension caps.
- Validate independently.
Poppler 26.05.0 checked PDF structure, page sizes, text extraction, image counts and dimensions. First pages were rendered at 120 DPI and compared with NumPy for PSNR, mean absolute error, and changed pixels.
- Inspect the visible output.
Source and result renders were reviewed for legibility, clipping, missing content, background corruption, and raster softness.
- Publish exact evidence.
The CSV includes output checksums and unrounded measurements. See the broader testing methodology for release gates.
Limitations
These are synthetic stress cases designed with unusually large, noisy source images, so their reductions can be much larger than an already optimized real-world PDF. PSNR is a full-page pixel metric, not a guarantee of perceived quality, OCR accuracy, accessibility, print quality, or preservation of every PDF feature. The fixtures do not test forms, annotations, signatures, unusual color profiles, damaged files, or every image filter.
Runtime was recorded during testing but omitted from the public comparison because it depends heavily on hardware, browser state, and thermal conditions. Results describe production build v84 and should be re-run when the compression code or codecs change.
Further questions
- How does each preset affect OCR word accuracy on small scanned text?
- At what print size do Balanced and Maximum scan artifacts become noticeable?
- How do CMYK, transparency masks, monochrome fax scans, and unusual image filters behave?