Performance
Criterion measures complete, in-process conversion to PDF bytes:
| Document | Representative median | Approx. conversions/sec |
|---|---|---|
Simple HTML (<h1> + <p>) |
0.93 ms | 1,080 |
| Styled HTML (CSS, lists, links) | 3.5 ms | 285 |
| Table (5 rows, styled headers) | 5.9 ms | 170 |
| Markdown (headings, code, lists) | 7.0 ms | 143 |
| Full report (tables, flex, progress bars) | 15.9 ms | 63 |
| Full report + header/footer | 15.9 ms | 63 |
| 200 CJK text-emphasis spans | 310 ms | 3.2 |
Measured on macOS ARM (Apple M2) with Rust 1.94.0 at commit
600aadb. The benchmark profile uses opt-level=3, fat LTO, and one codegen
unit. Results depend on hardware and document content; conversions/sec is not a
page-throughput claim. Reproduce the measurements with:
cargo bench --bench conversionQuick start
use ironpress::html_to_pdf;
let pdf = html_to_pdf("<h1>Hello</h1><p>World</p>").unwrap();
std::fs::write("output.pdf", pdf).unwrap();let pdf = ironpress::markdown_to_pdf("# Hello\n\nWorld").unwrap();CLI
cargo install ironpress
ironpress input.html output.pdf
ironpress document.md output.pdf
ironpress --page-size letter --landscape --margin 54 input.html output.pdf
ironpress --header "Report" --footer "Page {page} of {pages}" input.html output.pdf
echo '<h1>Hello</h1>' | ironpress --stdin output.pdfBuilder API
use ironpress::{HtmlConverter, Margin, PageSize, RasterQuality};
let pdf = HtmlConverter::new()
.page_size(PageSize::LETTER)
.margin(Margin::uniform(54.0))
.raster_quality(RasterQuality {
background_dpi: 144.0,
..RasterQuality::default()
})
.header("My Document")
.footer("Page {page} of {pages}")
.convert("<h1>Custom page</h1>")
.unwrap();RasterQuality keeps source-image, filter, and flattened-background resolution
in one physical-DPI policy. Its default preserves sharp source/filter output
while using 192 DPI for flattened synthetic backgrounds; lowering one field
does not change page geometry. The CLI exposes the same controls through
--image-dpi, --filter-dpi, and --background-raster-dpi.
Features at a glance
| Area | Highlights | Details |
|---|---|---|
| HTML | 50+ elements: headings, tables, lists, forms, media, <img>, inline <svg>
|
Layout Engine |
| CSS | Flexbox, grid, multi-column, calc(), variables, @media, @page, @font-face
|
CSS Support |
| Fonts | Base-14 PDF fonts, custom TTF embedding with subsetting, system font discovery, Unicode/CJK fallback | Font System |
| Math | LaTeX via $...$ / $$...$$: fractions, roots, matrices, Greek, operators |
Math Engine |
| SVG | Vector rendering: path, shapes, gradients, transforms, clip paths, viewBox
|
Layout Engine |
| Images | JPEG + PNG, data URIs, local files, remote URLs (remote feature) |
Architecture |
| PDF 1.4, bookmarks, link annotations, headers/footers, gradients, streaming output | PDF Rendering | |
| WASM |
npm install ironpress - runs 100% client-side in the browser |
WASM & Playground |
| Testing | 3500+ unit tests, property-based tests, 6 fuzz targets, 1,662-fixture parity corpus | Testing Strategy |
Custom fonts
let pdf = HtmlConverter::new()
.add_font("Inter", std::fs::read("Inter.ttf").unwrap())
.convert(r#"<p style="font-family: Inter">Shaped with HarfBuzz</p>"#)
.unwrap();Fonts are shaped with rustybuzz, subset to used glyphs only, and embedded as CIDFontType2. Characters outside WinAnsi (CJK, Arabic, emoji) are rendered via automatic Unicode font fallback. See Font System.
Math
The equation $E = mc^2$ is famous.
$$\sum_{k=1}^{n} k = \frac{n(n+1)}{2}$$Full LaTeX support: fractions, roots, matrices, Greek letters, operators, delimiters, accents. See Math Engine.
Python / Ruby
pip install ironpressimport ironpress
pdf = ironpress.html_to_pdf("<h1>Hello</h1>")gem install ironpressrequire "ironpress"
pdf = Ironpress.html_to_pdf("<h1>Hello</h1>")WASM
npm install ironpressimport init, { htmlToPdf, markdownToPdf } from 'ironpress';
await init();
const pdf = htmlToPdf('<h1>Hello</h1>');
const blob = new Blob([pdf], { type: 'application/pdf' });See WASM & Playground.
Security
HTML is sanitized by default. Scripts, iframes, event handlers, and
javascript: URLs are removed. SVG sanitization and image decoder limits also
apply.
Local files are denied unless base_path or resource_root grants a canonical
directory. Traversal and symlink escapes outside that directory are rejected.
Remote fetching is disabled unless the crate is built with remote:
cargo add ironpress --features remoteWith that feature, public HTTP and HTTPS resources are allowed by default. Loopback, private, link-local, metadata, multicast, documentation, and reserved addresses are denied. Redirects are checked again, DNS results are pinned to the connection, and response bodies are limited to 10 MB by default.
Use an allow list for a known CDN, or combine it with deny_public_ips(true)
to deny every host that was not explicitly allowed:
use ironpress::{HtmlConverter, NetworkPolicy, RemoteHost};
let cdn: RemoteHost = "cdn.example.com".parse().expect("valid host");
let policy = NetworkPolicy::default()
.with_allow_list([cdn])
.deny_public_ips(true)
.max_redirects(4)
.max_body_size(2 * 1024 * 1024);
let pdf = HtmlConverter::new()
.network_policy(policy)
.convert("<img src='https://cdn.example.com/logo.png'>")
.expect("conversion succeeds");A deny-list match always wins. An allow-list match explicitly trusts that host and bypasses its IP-class check.
Environment proxies are respected. They are operator configuration, not document input. If the proxy resolves the target hostname, Ironpress can still check target IP literals and host lists, but the proxy must enforce the final IP policy.
For a server that converts untrusted documents, also enforce egress outside the process:
- Block cloud metadata, loopback, private, and link-local networks at the host or network namespace.
- Restrict outbound DNS and traffic to required destinations or a controlled proxy.
- Apply time, memory, and process limits to conversions.
- Treat image malware scanning as a server concern. Ironpress controls resource access; it is not an antivirus.
HTML sanitization, local-file access, and remote access are independent.
Calling .sanitize(false) does not disable either resource policy.
Migration note: .sanitize(false) no longer grants implicit access to files in
the process working directory. Configure .base_path(...) for document assets,
and .resource_root(...) when those assets need a broader directory boundary.
How it works
HTML/Markdown → Sanitize → Parse (html5ever) → Style cascade → Layout engine → PDF 1.4
See Architecture for the full pipeline.
Visual parity harness
tests/parity/ is an adversarial HTML/CSS corpus with one focused fixture per
feature, value, or interaction. Ironpress produces the candidate PDF; the
declared oracle renderer's PDF is committed. At test time both PDFs go through the
same discovered pdftoppm executable with the same 300 DPI arguments. A fixture
passes only when a fixed, same-coordinate human-visibility policy finds no
visible defect. Every raw RGBA difference remains in the evidence; the harness
never translates, registers, or fixture-tunes either raster.
scripts/parity.sh # run the complete exact parity gate
scripts/parity-gen-refs.sh <category> # regenerate oracle PDFs explicitly
scripts/parity-gen-refs.sh --check # authenticate the complete corpus-
Run it:
scripts/parity.shsupplies a fresh invocation identity, renders every fixture in-process, rasterizes candidate and oracle PDFs symmetrically, and verifies that JSON, Markdown, and HTML all belong to that invocation. -
Read it: the HTML parity report
provides the complete visual evidence;
tests/parity/REPORT.mdis the compact, problem-first summary, andtests/parity/report.jsoncontains the complete machine result. -
Oracles: committed PDFs are the source of truth. Oracle-preview, candidate,
and diff PNGs are generated report evidence and are intentionally ignored.
refs.lockauthenticates each fixture, oracle PDF, manifest entry, renderer, fonts, and generator provenance. Every future oracle PDF is generated only by the pinned Chromium Fontations/Foundation launcher; authenticated historical non-Chromium PDFs are evidence-only and cannot be regenerated. -
Baseline:
tests/parity/baseline.jsonis a separately reviewed regression snapshot. Updating it is explicit; retained FAILs remain current-health failures while their exact rasters become protected against movement or worsening. -
CI:
.github/workflows/parity.ymlruns the same browser-free gate, checksrefs.lock, and uploads the current report and evidence even when defects make the gate fail.
License
MIT
