Optimize draw.io SVG exports

What it does

Drop rasterised label fallbacks

When draw.io cannot guarantee a fallback font will render — notably for CJK text — it rasterises the label and embeds it as a base64 PNG. On a typical export those images are 91% of the file, and around 99% of the achievable gzip win, because already-compressed data is something gzip cannot touch. This single pass is most of the value. An <image> anywhere else is real diagram content and is never touched.

Collapse <switch> onto draw.io’s own text

Where the export already carries a faithful <text> fallback, the HTML branch is dropped and the text hoisted at its exact position. The text-equivalence check is a hard precondition rather than a formality: draw.io truncates the fallback with an ellipsis when a label overflows its shape, and collapsing onto that would silently corrupt the label.

Convert HTML labels to SVG text

The only way to remove a <foreignObject> from an export that has no text fallback — and the only way to make those labels render outside a browser at all. In this app the shaping is done by your browser’s own layout engine rather than a reimplementation, and the result is checked against draw.io’s own coordinates.

Collapse the group scaffolding

draw.io wraps every shape in several groups that carry nothing. Removing them is asserted to be geometrically lossless, not merely to look right: every path’s absolute bounding box is resolved through its transforms and required to match exactly.

Strip editor bookkeeping and unused ids

data-cell-id attributes, empty style elements, the redundant transparent-background declaration. Referenced ids are collected from url(), href, ARIA attributes, animation timing and CSS before anything is removed — a reference that lives only in an @supports rule is still a reference.

Sanitize

These files usually come from someone else, so scripts, event handlers, external references, SMIL animation and CSS url() are removed by default. Two different risks are being prevented: that the document makes this tool fetch something, and that the output turns a static picture into a beacon that leaks a viewer’s IP and referrer.

The passes and the reasoning behind each are documented in full in the drawio-svg-optimizer repository, which is also available as a command-line tool.