Tankōbon

Appearance

0004. PDF pages rasterized in the main process with pdf.js + @napi-rs/canvas ​

  • Status: Accepted; the rendering now runs in the decoder process (ADR 0012)
  • Date: recorded 2026-09-30 (the decision predates this record)

Context ​

The reader, the zoom and the continuous mode all work on page images returned by a ComicArchive. PDF pages are vector content, with no native resolution. Rendering them in Node needs a Canvas implementation, which Node doesn't have.

Decision ​

  • PdfArchive (src/main/services/pdf-archive.ts) implements ComicArchive like CBZ/CBR, rasterizing each page on demand to PNG at a fixed ~200 DPI (RENDER_SCALE = 200 / 72).
  • It uses pdfjs-dist's legacy build, with @napi-rs/canvas providing DOMMatrix, Path2D, ImageData and Image (installed once on globalThis) and the canvas itself. @napi-rs/canvas is pdf.js's own optional dependency for Node and ships prebuilt per-platform binaries (no compilation step).
  • Both packages are kept external in the main bundle and copied into the packaged app by the packageAfterCopy hook of forge.config.ts; AutoUnpackNativesPlugin unpacks the native binary out of the asar archive.

Consequences ​

  • The rest of the reader doesn't distinguish a PDF page from an image page.
  • ~200 DPI is high enough not to trigger the AI upscaling at normal zoom, without the memory/CPU cost of going higher.
  • pdf.js's standard_fonts and cmaps ship with the package, with no separate copy step.
  • Only the @napi-rs/canvas-<platform>-<arch> package matching the machine that ran npm install is present, so the app must be packaged on each target platform.
  • Packaging is more fragile: see the build gotchas.

Alternatives considered ​

  • Bundling both packages with Vite: impossible for the native .node binary, and unsafe for pdf.js's legacy build (itself a webpack bundle).
  • No other rendering approach (in the renderer, other libraries) was evaluated.