Pillow-AllSIMD performance

Pillow-AllSIMD 13 (the simd/pillow-13 branch), a fork of Pillow-SIMD and Pillow, against Pillow 13.0, the last Pillow-SIMD release, OpenCV, libvips and ImageMagick, on one x86-64 and one arm64 machine. Each library runs single-threaded on the same 2560×1600 images. Bars show throughput: longer is faster. The second number is the speed relative to Pillow 13.0.0. The branch's builds run in exact mode (PILLOW_ALLSIMD_FAST=0, Pillow's arithmetic); fast is its AVX2 or NEON build again in fast mode, the default since October 2026: kernels whose pixels may differ from Pillow's by a few levels (the fast kernels); operations without one run the exact kernels, as in the exact build above it.

Jump to

How it's measured

Every library runs in its own process, pinned to one CPU core of the container (--cpuset-cpus), with one thread: cv2.setNumThreads(1), VIPS_CONCURRENCY=1, MAGICK_THREAD_LIMIT=1. Pillow never uses more than one.
The inputs are a 2560×1600 photo-like image (smooth gradients with a little noise), a second one for two-image operations, an RGBA version with a soft alpha, its L version, and the photo as JPEG q90, PNG and WebP q90. They're made once and loaded into each library's own image type before timing.
The libraries take turns suite by suite, for three rounds. In each round a case runs once to warm up, then at least 3 times and for at least 0.2 s (at most 4 s). The fastest run of all rounds counts. Throughput is megapixels of the source image per second; drawing cases count shapes or lines per second.
On the x86 laptop, turbo is off (intel_pstate/no_turbo) and the power profile is performance, so the core runs at its 2.9 GHz base clock. With turbo on, the clock moved between 2.9 and 5.4 GHz during a run, so whichever library ran during a slow spell looked about twice as slow. Under the balanced profile it kept dropping to 2.3 GHz even with turbo off. The Jetson runs at a fixed clock.
libvips is lazy, so its cases end by computing the result into memory, and its operation cache is off. ImageMagick modifies images in place, so each of its runs works on a clone, which costs about what Pillow spends allocating its output.
Not every library has every operation, and some differ from Pillow's. Those differences are footnoted under each chart. A missing bar means the library has no such operation.
Pillow 13.0.0 is the PyPI wheel with its bundled codecs. The Pillow-AllSIMD builds are compiled in the container with GCC against Debian's codec libraries. Codec timings therefore also compare library builds, not only Pillow's code.

Versions

Run it yourself

The benchmark and the page live in desmoteo/pillow-allsimd-perf. With Docker on the target machine:

git clone https://github.com/desmoteo/pillow-allsimd
git clone https://github.com/desmoteo/pillow-allsimd-perf
cd pillow-allsimd-perf
git -C ../pillow-allsimd checkout simd/pillow-13
./run.sh local                       # this machine
./run.sh --cpu 2 user@arm64-host     # another one, over SSH
python3 tools/publish.py             # copy results/ into docs/data/