Benchmarks

Selected operations, one thread, a 2560×1600 RGB image. Each cell shows Pillow 13.0.0’s time, then Pillow-AllSIMD’s, then the speed-up:

Codecs gain little: decoding and encoding are mostly libjpeg’s, libpng’s and libwebp’s own work. The machines are an Intel Core Ultra 9 285H laptop (with turbo off) and an NVIDIA Jetson Orin (Cortex-A78AE). Each time is the best of 3 interleaved rounds, in Docker with Python 3.12 and GCC 14, measured on 2026-10-31 on the same x86-64 and arm64 code as this release. The benchmark page has all 111 operations, in exact and fast mode. It also compares Pillow-SIMD, OpenCV, libvips and ImageMagick, with notes where another library’s operation differs.

Installing

Pillow-AllSIMD is a source package, so pip compiles it on installation. Install the libraries that building Pillow from source needs first. It installs the PIL package, as Pillow does. Install your other packages first, then replace Pillow with it:

python3 -m pip uninstall -y pillow pillow-simd
python3 -m pip install -C simd=avx2 pillow-allsimd

On arm64, leave out -C simd=avx2: the default is NEON there. Uninstall Pillow before installing, not after: both own the files in PIL. To rebuild with another simd option, add --no-cache-dir --force-reinstall. Otherwise pip reuses the wheel it built the first time. With -v, pip shows the line --- SIMD kernels: ... from the build, which names the kernels you got.

Exact and fast modes

Fast mode is on by default. The Gaussian blur, RGB ↔ YCbCr, RGB → HSV, the affine, perspective and quad transforms, rotate(), blend(), the resize of RGB, RGBA and other 4-byte images and large downscales then use faster kernels. Their results differ from Pillow’s by a few levels at most and are the same on every build and architecture. Exact mode follows Pillow’s arithmetic for every operation. Use it to compare pixels with images made by Pillow:

PILLOW_ALLSIMD_FAST=0 python3 app.py

Or, in code:

from PIL import Image, ImageFilter

Image.set_fast(False)  # every operation from now on

with Image.fast_mode(False):  # one block
    blurred = im.filter(ImageFilter.GaussianBlur(2))

Fast mode in the documentation lists each operation and how far its results may differ.

Testing

Pillow’s test suite runs on every build. Consistency tests check each SIMD kernel against the scalar code, and per-kernel output hashes compare the builds with each other. The hardware tests run on x86-64 (SSE4, AVX2, AVX-VNNI) and on arm64 (NEON). The portable build also passes the suite on riscv64, ppc64le, s390x, armv7 and 32-bit x86, under QEMU. Status: beta.

Links

• Repository and readme

• Documentation

• Benchmarks

• Changelog

• Issues: SIMD kernels, builds and fast mode. Report other bugs to Pillow.

Pillow is by Jeffrey ‘Alex’ Clark and contributors, on PIL by Fredrik Lundh and Secret Labs AB. Pillow-SIMD is by Alexander Karpinsky and contributors, supported by Uploadcare. All are licensed under the MIT-CMU License, and so is Pillow-AllSIMD.