How the compressor works

Smart mode (default)

We parse the PDF into its objects, walk the object graph from the document catalog and drop everything that isn't reachable — leftovers from previous edits. We remove metadata and embedded page thumbnails, compress any uncompressed streams with zlib, then look at each image. Supported images (JPEG, and 8-bit RGB or grayscale Flate images, including PNG-predicted ones) are decoded with the browser's own decoder, resized so the longest side fits the chosen limit, and re-encoded as JPEG. If the new version isn't smaller, the original image is kept. CMYK, indexed-colour and masked images are left untouched to avoid colour shifts.

Exact-size mode

Given a target like 100 KB, we first try lossless cleanup. If that isn't enough, we binary-search an 8-step ladder from high quality (JPEG 82%, 2600 px) to low (20%, 620 px), keeping the best result under the target. If even the lowest step is too large — for example because the PDF contains many vector-heavy pages — we render each page with pdf.js and binary-search resolution (150 → 44 DPI) and quality. The result tells you which path was used.

Safety

We never return a file larger than your original. PDFs with security restrictions are flattened rather than modified, because rewriting them in place would break their encryption.