The Sane plugin has now been split into its own repo: https://github.com/python-pillow/Sane .
Png text chunk size limits¶
To prevent potential denial of service attacks using compressed text
chunks, there are now limits to the decompressed size of text chunks
decoded from PNG images. If the limits are exceeded when opening a PNG
ValueError will be raised.
Individual text chunks are limited to
PIL.PngImagePlugin.MAX_TEXT_CHUNK, set to 1MB by
default. The total decompressed size of all text chunks is limited to
PIL.PngImagePlugin.MAX_TEXT_MEMORY, which defaults to
64MB. These values can be changed prior to opening PNG images if you
know that there are large text blocks that are desired.
Image resizing filters¶
Image resizing methods
thumbnail() take a resample argument, which tells
which filter should be used for resampling. Possible values are:
Almost all of them were changed in this version.
Bicubic and bilinear downscaling¶
From the beginning
BICUBIC filters were based on affine transformations
and used a fixed number of pixels from the source image for every destination
pixel (2x2 pixels for
BILINEAR and 4x4 for
BICUBIC). This gave an unsatisfactory result for
downscaling. At the same time, a high quality convolutions-based algorithm with
flexible kernel was used for
Starting from Pillow 2.7.0, a high quality convolutions-based algorithm is used for all of these three filters.
If you have previously used any tricks to maintain quality when downscaling with
(for example, reducing within several steps), they are unnecessary now.
Antialias renamed to Lanczos¶
PIL.Image.LANCZOS constant was added instead of
ANTIALIAS was initially added, it was the only
high-quality filter based on convolutions. It’s name was supposed to reflect
this. Starting from Pillow 2.7.0 all resize method are based on convolutions.
All of them are antialias from now on. And the real name of the
ANTIALIAS filter is Lanczos filter.
ANTIALIAS constant is left for backward compatibility
and is an alias for
Lanczos upscaling quality¶
The image upscaling quality with
LANCZOS filter was
almost the same as
BILINEAR due to bug. This has been fixed.
Bicubic upscaling quality¶
BICUBIC filter for affine transformations produced
sharp, slightly pixelated image for upscaling. Bicubic for convolutions is
In most cases, convolution is more a expensive algorithm for downscaling
because it takes into account all the pixels of source image. Therefore
performance can be lower than before. On the other hand the quality of
BICUBIC was close to
NEAREST. So if such quality is suitable for your tasks
you can switch to
NEAREST filter for downscaling,
which will give a huge improvement in performance.
At the same time performance of convolution resampling for downscaling has been
improved by around a factor of two compared to the previous version.
The upscaling performance of the
LANCZOS filter has
remained the same. For
BILINEAR filter it has improved by
1.5 times and for
BICUBIC by four times.
Default filter for thumbnails¶
In Pillow 2.5 the default filter for
Antialias was chosen because all the other filters gave poor quality for
reduction. Starting from Pillow 2.7.0,
ANTIALIAS has been
BICUBIC, because it’s faster and
ANTIALIAS doesn’t give any advantages after
downscaling with libjpeg, which uses supersampling internally, not convolutions.
A new method
PIL.Image.TRANSPOSE has been added for the
transpose() operation in addition to
TRANSPOSE is an algebra
transpose, with an image reflected across its main diagonal.
The speed of
TRANSPOSE has been significantly improved for large
images which don’t fit in the processor cache.
Gaussian blur and unsharp mask¶
GaussianBlur() implementation has been replaced
with a sequential application of box filters. The new implementation is based on
“Theoretical foundations of Gaussian convolution by extended box filtering” from
the Mathematical Image Analysis Group. As
implementations use Gaussian blur internally, all changes from this chapter
are also applicable to it.
There was an error in the previous version of Pillow, where blur radius (the standard deviation of Gaussian) actually meant blur diameter. For example, to blur an image with actual radius 5 you were forced to use value 10. This has been fixed. Now the meaning of the radius is the same as in other software.
If you used a Gaussian blur with some radius value, you need to divide this value by two.
Box filter computation time is constant relative to the radius and depends on source image size only. Because the new Gaussian blur implementation is based on box filter, its computation time also doesn’t depends on the blur radius.
For example, previously, if the execution time for a given test image was 1 second for radius 1, 3.6 seconds for radius 10 and 17 seconds for 50, now blur with any radius on same image is executed for 0.2 seconds.
The previous implementation takes into account only source pixels within 2 * standard deviation radius for every destination pixel. This was not enough, so the quality was worse compared to other Gaussian blur software.
The new implementation does not have this drawback.
TIFF Parameter Changes¶
Several kwarg parameters for saving TIFF images were previously specified as strings with included spaces (e.g. ‘x resolution’). This was difficult to use as kwargs without constructing and passing a dictionary. These parameters now use the underscore character instead of space. (e.g. ‘x_resolution’)