↓Skip to main content

Speeding up Pillow's open and save

Python python3.15 pillow tachyon performance
Pillow and Tachyon logos
Table of Contents

Tachyon #

I tried out Tachyon, the new “high-frequency statistical sampling profiler” coming in Python 3.15, to see if we can speed up the Pillow imaging library. I started with a simple script to open an image:

import sys
from PIL import Image

im = Image.open(f"Tests/images/hopper.{sys.argv[1]}")

Then ran:

$ python3.15 -m profiling.sampling run --flamegraph /tmp/1.py png
Captured 35 samples in 0.04 seconds
Sample rate: 1,000.00 samples/sec
Error rate: 25.71
Flamegraph data: 1 root functions, total samples: 26, 169 unique strings
Flamegraph saved to: flamegraph_97927.html

Which generates this flame graph:

Flame graph for opening a PNG with Pillow

The whole thing took 40 milliseconds, with half in Image.py’s open(). If you visit the interactive HTML page we can see open() calls preinit(), which in turn imports GifImagePlugin, BmpImagePlugin, PngImagePlugin and JpegImagePlugin (hover over the <module> boxes to see them).

Do we really need to import all those plugins when we’re only interested in PNG?

Okay, let’s try another kind of image:

$ python3.15 -m profiling.sampling run --flamegraph /tmp/1.py webp
Captured 59 samples in 0.06 seconds
Sample rate: 1,000.00 samples/sec
Error rate: 22.03
Flamegraph data: 1 root functions, total samples: 46, 256 unique strings
Flamegraph saved to: flamegraph_98028.html

Flame graph for opening a WebP with Pillow

Hmm, 60 milliseconds with 80% in open() and most of that in init(). The HTML page shows it imports AvifImagePlugin, PdfImagePlugin, WebpImagePlugin, DcxImagePlugin, DdsImagePlugin and PalmImagePlugin. We also have preinit importing GifImagePlugin, BmpImagePlugin and PngImagePlugin.

Again, why import even more plugins when we only care about WebP?

Loading all the plugins? #

That’s enough profiling, let’s look at the code.

When open()ing or save()ing an image, if Pillow isn’t yet initialised, we call a preinit() function. This loads five drivers for five formats by importing their plugins: BMP, GIF, JPEG, PPM and PNG.

During import, each plugin registers its file extensions, MIME types and some methods used for opening and saving.

Then we check each of these plugins in turn to see if one will accept the image. Most of Pillow’s plugins detect an image by opening the file and checking if the first few bytes match a magic prefix. For example:

  • GIF starts with b"GIF87a" or b"GIF89a".
  • PNG starts with b"\211PNG\r\n\032\n" (reference).
  • JPEG starts with b"\xff\xd8\xff", where \xff\xd8 means “Start of Image”, and \xff is the start of the next marker (reference).

If none of these five match, we call init(), which imports the remaining 42 plugins. We then check each of these for a match.

This has been the case since at least PIL 1.1.1 released in 2000 (this is the oldest version I have to check). There were 33 builtin plugins then and 47 now.

Lazy loading #

This is all a bit wasteful if we only need one or two image formats during a program’s lifetime, especially for things like CLIs. Longer running programs may need a few more, but unlikely all 47.

A benefit of the plugin system is third parties can create their own plugins, but we can be more efficient with our builtins.

I opened a PR to add a mapping of file extensions to plugins. Before calling preinit() or init(), we can instead do a cheap lookup, which may save us importing, registering, and checking all those plugins.

Of course, we may have an image without an extension, or with the “wrong” extension, but that’s fine; I expect it’s rare and anyway we’ll fall back to the original preinit() -> init() flow.

After merging the PR, here’s a new flame graph for opening PNG (HTML page):

Much less compressed flame graph showing less work

And for WebP (HTML page):

Much less compressed for WebP

The flame graphs are scaled to the same width, but there’s far fewer boxes meaning there’s much less work now. We’re down from 40 and 60 milliseconds to 20 and 20 milliseconds.

The PR has a bunch of benchmarks which show opening a PNG (that previously loaded five plugins) is now 2.6 times faster. Opening a WebP (that previously loaded all 47 plugins), is now 14 times faster. Similarly, Saving PNG is improved by 2.2 times and WebP by 7.9 times. Success! This will be in Pillow 12.2.0.

See also #