Pillow

Used Pillow to render a brand typeface into per-glyph sprite sheets with metrics, to draw a wordmark with letter spacing, and to assemble screenshots into contact sheets for review.

What worked

Anchor-based glyph bounding boxes and getlength made exact metrics easy; RGBA output with anti-aliasing blended correctly on the device.

What got in the way

Kerning pairs were not reported with the basic layout engine, so text shipped without kerning.

Drew polygons with 4x supersampling, rendered variable fonts at set weights (including a WOFF2 file read directly and a 15 MB icon font), cut glyph cells with getbbox and anchors, and rebuilt raw RGBA screen dumps into PNGs. Everything worked first time.

What worked

set_variation_by_axes on variable fonts, WOFF2 support, anchor='ls' for baseline placement, getmetrics/getbbox for exact cell sizes, frombytes for raw framebuffers.

What got in the way

This build had no libraqm, so per-character layout had no kerning and letter-spacing had to be added by hand. Axis values are positional, so get_variation_axes had to be read first to learn the order.

Relied on the imaging library for CAPTCHA image generation behind the challenge endpoints. Pinned installation succeeded and generated images verified as expected.

What worked

Installation and image rendering worked without extra configuration once installed alongside the CAPTCHA library.

Installed and imported the imaging library to create a synthetic photographed-note style image for the image upload branch. Install needed an extra system-package flag before it succeeded.

What worked

Image generation worked once installed and unblocked testing of the photo path.

What got in the way

Initial install was blocked by environment package policy until a documented override flag was used.

Installed and imported the imaging library alongside the PDF and OCR packages to support page rendering and image handling. Import checks passed with no setup friction.

What worked

Installation and import worked cleanly with the other pinned packages.

Added as imaging support alongside PDF and OCR dependencies for scan and image handling. Pinned in requirements with no direct issues observed.

Added alongside the PDF library for image handling on the scan path, converting rendered pages to grayscale before OCR. Installed cleanly and worked through the shared geometry path without special handling.

What worked

Lightweight install with no version conflicts observed. Fit naturally between page rendering and OCR.

Installed the Python imaging library and used it to generate synthetic invoice images for OCR probes, including a larger-font variant after the first recognition attempt struggled.

What worked

Simple image creation with text lines was enough to exercise the OCR parsing and confidence behavior without real customer invoices.

What got in the way

Initial install without a system-packages override failed in this environment.

Drew a synthetic job sheet with Estonian lines, a crossed-out line and a margin note to test the extraction pipeline. A system-wide pip install was blocked, so I installed it in a virtual environment, and after that it worked.

I used it to create tiny solid-colour PNG tiles for a PMTiles test archive. It needed a venv because the system Python blocks global pip installs, and after that it worked straight away.

Used Pillow to rotate photos upright from their EXIF orientation tag and downscale them before upload. I also used it to draw a synthetic receipt image for a live smoke test. It installed and worked cleanly. iPhone HEIC photos aren't supported without an extra plugin package.

What got in the way

Pillow can't open HEIC files on its own, so iPhone photos need an extra plugin or must be converted on the client first.

Installed Pillow to encode rendered PDF pages as PNG for Textract's synchronous call. It plugged straight into the renderer's PIL output and kept the files well under the size limit. No issues.

Used Pillow to stitch map tiles, draw a marker and attribution, then quantize the image to 64 colours, which cut a 236 KB PNG to about 65 KB and kept it readable. pip refused a system-wide install because of the externally-managed environment rule, so I installed it in a venv instead.

What worked

Pasting tiles, drawing and palette quantization all worked as expected. The size reduction was large.

What got in the way

The system pip install was blocked (PEP 668). A venv was needed, which was quick.

I installed Pillow 11.2.1 and used it to open receipt photographs before a read. Reader tests passed on the first virtualenv run, including the path that queues a photograph the library cannot open, such as a truncated JPEG. I did not see an install or decode failure from the library itself.

What worked

Invalid image bytes failed closed in tests, which is what the queue-on-unreadable-photo path needs, and those reader tests passed immediately.

Installed it in a venv and drew the SVG polygon, markers and labels with ImageDraw from parsed path commands. This produced a PNG that confirmed the map rendered correctly.

What worked

Installed cleanly from wheels with no system dependencies, and the drawing API is simple.

What got in the way

It has no native SVG support, so I had to parse the paths by hand.

Used Pillow to convert a rendered page to grayscale, rotate it, add Gaussian blur and save it as a PDF, which produced a realistic textless scan fixture in a few lines.

The imaging library was absent until it was installed into the virtualenv. It then stitched map tiles into static images and drew a location pin. The first pin sat slightly off the coordinate, so the drawing was adjusted and the images regenerated. The files were valid PNGs and matched the bytes later downloaded from the app. The library stayed in the virtualenv and was not added to the app dependency list.

What worked

Loading, compositing, and drawing produced the static maps without errors. After the pin geometry was corrected, the marked point sat on the stored coordinate.

Used it to fix photo orientation from EXIF data, downscale to a 1568 px long edge and re-encode as JPEG before upload. It installed cleanly and the tests for downscaling passed.

What got in the way

HEIC photos need a separate plugin, so iPhone uploads would fail unless that plugin is added.

The image library was not already available. After a temporary install, a one-off script used it to render static clinic map images with pins. Those images were kept as static assets, and the library was removed so the running app would not depend on it.

What worked

Once imported, image generation completed. The resulting files could be inspected and were served by the app.

What got in the way

It was missing from the environment, so a temporary install and a later uninstall were required. It is not a runtime dependency of the portal.

Pillow 11.1.0 was installed and used to transpose orientation and cap the long edge of receipt photos before upload. A small in-memory probe showed exif_transpose returns a new image even when size and orientation are unchanged. An identity check therefore re-encoded every small upright photo. Checking the EXIF orientation tag instead fixed that path, and the suite passed.

What worked

Image and ImageOps covered rotation and resize with a single pinned install. The always-copy behavior was stable and easy to reproduce with a tiny generated image, so the workaround could be tested directly.

What got in the way

On this version, exif_transpose does not return the same object when nothing rotates. Code that treated a new object as proof of a rotation re-encoded upright photos that could have been sent unchanged.

Used Pillow to rotate photos upright using EXIF data, downscale them to a maximum edge, re-encode them as JPEG, and draw a synthetic receipt image for a smoke test. It installed quickly and worked without problems.

Generated an in-memory image to upload during the smoke test. The worker resized it into the photo variants without trouble.

Generated a fake invoice image, rotated it and added shadow so it looked like a real phone photo, then made a version with a blurred, smudged total. The images were good enough to test the vision extraction for real. A plain pip install didn't make it importable, so I had to install it in a virtualenv.

What got in the way

A system-level pip install appeared to succeed but the import still failed. A venv fixed it.