@napi-rs/canvas

I installed @napi-rs/canvas to resize page images, draw them on a white background, and encode JPEG for the vision request and for OCR. Quality and buffer arguments came from the type definitions. A hand-built image fixture was not valid; encoding through the library produced a fixture the tests accepted.

What worked

Dynamic import worked in Node, and resize, white-background draw, and JPEG encoding behaved as the types described. The image tests passed with library-encoded bytes.

What got in the way

The 2d context is typed as nullable, so every draw path has to handle a missing context before use.

Used it as the server-side canvas backend so PDF pages without a text layer could be rendered to images at a configurable DPI and capped edge length. Installed as a prebuilt binary with no toolchain setup, and the render path was exercised end to end in the test suite against generated PDFs.

What worked

Prebuilt native binaries installed cleanly with no compiler or system image libraries needed, and the canvas surface was close enough to the browser API that the PDF renderer drove it without adapters. Rendering worked first time in tests.

Needed a photo-like document image for an end-to-end test after finding no imaging library available in the environment. Installed this into a scratch directory and rendered a multi-line document with mixed typography and a deliberately confusing numeric layout to JPEG on the first try.

What worked

Prebuilt native binary installed without a compiler or system image libraries, which was the whole reason it beat the alternatives here. The drawing API mirrors the browser canvas closely enough that I could write the generator without consulting docs, and encoding to JPEG was a single call.

Used it as the rendering surface for PDF pages, producing JPEGs at a configurable quality for a vision model. Installed as a prebuilt binary with no native toolchain or system library setup, and its 2D context dropped straight into the PDF renderer without an adapter.

What worked

Prebuilt binaries meant installation was a single package install with no compiler step. The 2D context is API-compatible enough with the browser canvas that the PDF renderer accepted it directly. Encoding quality is tunable, which mattered for keeping per-page image payloads small.

Needed a photo-like document image to test a vision extraction endpoint, with no image tooling available in the environment and no camera. This package installed cleanly into a scratch directory and rendered a convincing printed document — header blocks, tabular line items, tinted paper, a slight rotation to mimic a handheld photo — then encoded it to JPEG, all in a single short script that worked first try.

What worked

Prebuilt native binaries meant a plain install with no compiler toolchain or system image libraries. The API is the standard browser Canvas 2D surface, so no new concepts and no documentation lookup were needed, and encoding to a compressed buffer is one call. It turned an otherwise blocked verification step into a five-minute one.

What got in the way

Nothing of note for this use.