Before compressing anything, identify what is consuming the space. PDF files contain several distinct types of content: Images — By far the most common cause of large PDFs. A single full-page photograph exported at 300 DPI for print can be 10–20 MB on its own. PDFs intended for screen viewing rarely need more than 150 DPI, and often 96 DPI is sufficient. Embedded fonts — Each embedded font adds 50–500 KB to a PDF, depending on the typeface and whether the full font or a subset is embedded. A PDF with 10 embedded fonts and no subsetting can add several megabytes of font data for glyphs that never appear in the document. Unreferenced objects — When content is deleted from a PDF in an editing session, the old objects often remain in the file as unreferenced "garbage" — no longer pointed to by anything in the document structure, but still occupying space. This is especially common in PDFs that have been edited multiple times. Uncompressed streams — Some older PDFs, and PDFs generated by certain software, leave their content streams uncompressed. The Flate (zlib/DEFLATE) compression built into the PDF specification can reduce text content streams by 50–80% with no quality loss. Embedded thumbnails and metadata — Some creation tools embed per-page thumbnail images and extensive XMP metadata that modern viewers do not use. These are safely removable. A quick inspection with a tool like ExifTool or Apache PDFBox reveals the dominant content types before you start.

Images are the primary driver of PDF size in most real-world documents. The PDF specification supports multiple image compression formats, and choosing the right one for each image type makes a substantial difference: JPEG compression for photographs — Photographic content compresses best with JPEG (DCT in PDF terminology). Quality settings of 70–80% are visually indistinguishable from 100% in a typical PDF workflow while reducing image data by 60–80%. For images that will only ever be viewed on screen, quality settings of 60% are often acceptable. JBIG2 for black-and-white scans — Scanned text documents in black and white compress dramatically with JBIG2, which is purpose-built for bi-level (1-bit) content. JBIG2 can achieve 5–10× better compression than CCITT G4 for typical document pages with clean type. Flate (PNG-style) for screenshots and diagrams — For images with large areas of flat colour, sharp edges, or transparency, lossless Flate compression preserves quality while still achieving meaningful size reduction. Downsampling resolution — A document image intended for screen reading does not need 300 DPI. Downsampling to 150 DPI (print) or 96 DPI (screen) before compression reduces the raw pixel count — and therefore the compressed file size — in proportion to the square of the resolution ratio. Halving resolution from 300 to 150 DPI reduces image data by 75%. The key rule: do not apply lossy JPEG compression to images that have already been JPEG-compressed. Each re-compression cycle amplifies artefacts. If you have access to the original source image, re-export at the target resolution rather than re-compressing the embedded PDF version.

When a PDF embeds a font, it includes the font program — the mathematical descriptions of every glyph in the typeface. A full embedding of a Latin typeface typically adds 100–500 KB. Many PDF creation tools embed a subset of the font: only the glyphs actually used in the document. If your document uses 200 unique characters from a 65,000-glyph typeface, subsetting embeds only those 200 glyphs, reducing font overhead by over 99%. If your creation tool does not subset fonts automatically, or if you have merged PDFs with multiple copies of the same font embedded from different sources, a post-processing step to subset and deduplicate fonts can recover significant space. For PDFs where you need to minimise size at the cost of portability — documents intended for a specific viewer or system where the fonts are known to be available — fonts can be removed entirely and replaced with standard PDF system fonts (Helvetica, Times Roman, Courier). This is rarely advisable for public distribution: text rendering will change if the recipient's system substitutes a different font.

Every significant data structure in a PDF is stored as a stream object — a byte sequence with optional compression applied. The PDF specification's preferred compression filter is Flate, the same algorithm used by gzip and PNG. Any stream not already Flate-compressed can be compressed with no quality loss and typically 50–70% size reduction for text content. Content stream objects that are already compressed cannot be compressed again meaningfully. But other objects — page content descriptions, form XObjects, embedded ICC colour profiles — may be stored uncompressed in older PDFs and benefit from Flate compression. Cross-reference stream compression — PDF 1.5 introduced cross-reference streams (replacing the older plaintext xref table). Cross-reference streams store the object offset table in a compressed binary format, saving space in large PDFs with many objects. Object deduplication — Merged PDFs often contain duplicate font programs, ICC profiles, or form XObjects referenced by different objects but containing identical data. Deduplication identifies and collapses these to a single shared object. Tools like Ghostscript and QPDF perform this analysis. Garbage collection — Incremental PDF updates (used by editing tools to avoid rewriting the entire file on each save) append new versions of modified objects while leaving the old versions in place. Over many edit cycles, a document accumulates significant unreferenced data. A linearise-and-rewrite pass strips all unreferenced objects and produces a clean, minimal file.

Several apparent "compression" approaches are either ineffective or harmful: Re-printing to PDF — Printing a PDF to a virtual PDF printer and saving the output often increases file size or silently degrades quality. The print pipeline rasterises all vector content (text, shapes) into images, which are then re-compressed — typically losing text sharpness and font embedding entirely. Excessive JPEG quality reduction — Dropping JPEG quality below 60% for normal documents produces visible artefacts — blocking, colour banding, smearing around text — that are irreversible. The file size saving from quality 60 to quality 30 is smaller than from 100 to 60, but the quality loss is proportionally greater. Removing embedded fonts from documents with unusual typefaces — Removing fonts that are not available on the recipient's system results in font substitution, which changes text layout, line breaks, and in extreme cases obscures meaning. Compressing already-archived PDF/A files — PDF/A has specific requirements about embedded resources. Reprocessing a PDF/A file for size reduction can inadvertently remove required elements (embedded colour profiles, font programs) that are needed for archival conformance. Always re-validate conformance after modifying a PDF/A document.