libvips 8.9 is now done, so here’s a quick overview of what’s new. Check the ChangeLog if you need more details.

Credit to lovell, kleisauke, deftomat, omira-sch, meyermarcel, kalozka1, kayarre, angelmixu, pvdz and others for their great work on this release.

True streaming

This is the biggest change to libvips in years: it now supports true streaming.

Previously, libvips supported file and memory data sources and sinks. If you wanted to process images on systems like AWS, you were forced to read to memory first, then process back to memory again, then send the result to the output.

It looked something like this:

Processing with current libvips

libvips now lets you connect pipelines to any source or destination, so you can do something more like this:

Processing with libvips streams

There’s no buffering, so there should be a useful drop in latency.

It’s really easy to use. For example:

aws s3 cp s3://mybucket/input.jpg - | \
  vips thumbnail_source [descriptor=0] .jpg 128 | \
    aws s3 cp - s3://mybucket/output.jpg

In ruby-vips you can make a source like this:

require 'vips'

source = Vips::Source.new_from_file "some/file/name"
image = Vips::Image.new_from_source source, "", access: "sequential"

You can also make sources from file descriptors and memory areas.

You can make custom sources like this:

file = File.open "some/file/name", "rb"
source = Vips::SourceCustom.new
source.on_read { |length| file.read length }
image = Vips::Image.new_from_source source, "", access: "sequential"

And you can do anything in the read handler. You can define a seek handler as well, if your source supports it.

Write is just as simple:

target = Vips::Target.new_to_file "some/file/name"
image.write_to_target target, ".png"

And again you can define custom targets:

dest = File.open ARGV[1], "w"
target = Vips::TargetCustom.new
target.on_write { |chunk| dest.write(chunk) }
image.write_to_target target, ".png"

There’s an optional finish handler.

A post a few weeks ago introducing this in more detail. pyvips, C and C++ also support this new feature.

OSS-Fuzz integration

Thanks to work by Oscar Mira (@omira-sch), libvips is now part of OSS Fuzz. This is a Google project to continously test open-source projects for vulnerabilities – whenever one of their clusters is idle, it starts analyzing our code.

It’s been working away since August and only found two serious bugs, so that’s great, and both have been fixed in 8.8. This means 8.9 should be very solid.

There was a post a few months ago with a lot more detail, if you’re curious.

Switch/case

libvips has a pair of new operations which speed up many-way if-then-else.

vips_switch() takes an array of N condition images and for each pixel finds the index of the first non-zero pixel. If no pixels are non-zero, it sets the output to N + 1.

vips_case() takes an index image plus an array of N + 1 result images and for each pixel copies the pixel selected by the index to the output.

Put these two together and you can make a quick many-way if-then-else. For example, in Python you could write:

import pyvips

texture_names = [
    "k2.jpg",
    "k4.jpg",
    "shark.jpg",
    "k110.jpg"
]
colours = [
    (232, 225, 199),
    (211, 167, 73),
    (210, 125, 60),
    (151, 189, 174)
]

main = pyvips.Image.new_from_file(sys.argv[1], access="sequential")

textures = [pyvips.Image.new_from_file(name)
                .replicate(100, 100)
                .crop(0, 0, main.width, main.height)
            for name in texture_names]

tests = [(main == each_colour).bandand()
         for each_colour in colours]

textured = pyvips.Image.switch(tests).case(textures + [main])

textured.write_to_file(sys.argv[2])

So it’s finding pixels which equal one of the colours and swapping those pixels for pixels from the matching texture.

Breaking changes

There’s one serious and breaking change: libvips now blocks metadata modification in shared images, that is, images with a reference count greater than one.

You were always supposed to use copy to get a unique image before altering metadata, but this is now enforced. If you attempt it, the change won’t happen and a warning will be issued.

There should be no performance implication, since all copy does is duplicate a few pointers.

This change prevents a range of serious race conditions and possible crashes in highly threaded programs.

Image format improvements

There are a range of useful improvements to image file handling.

Previously, libvips only supported a single delay for all frames of animated images. Thanks to work by deftomat, it now keeps an array of delays, one per frame. Additionally, the meaning of the loop parameter is now consistent between webp and gif.

Solid WebP images will be automatically written without their alpha band. heifsave has a new compression option and (thanks to work by meyermarcel) supports alpha correctly. tiffsave supports webp and zstd compression and has more flexible alpha support. dzsave has a new no-strip option and supports IIIF layout. svgload has a new unlimited option. PPM load and save has been rewritten and is now faster and uses less memory.

Other

Plus many even smaller bug fixes and improvements. As usual, the ChangeLog has more details, if you’re interested.