Writing Your Own Image Plugin

The Pillow uses a plug-in model which allows you to add your own decoders to the library, without any changes to the library itself. Such plug-ins usually have names like XxxImagePlugin.py, where Xxx is a unique format name (usually an abbreviation).


Pillow >= 2.1.0 no longer automatically imports any file in the Python path with a name ending in ImagePlugin.py. You will need to import your image plugin manually.

Pillow decodes files in 2 stages:

  1. It loops over the available image plugins in the loaded order, and calls the plugin’s accept function with the first 16 bytes of the file. If the accept function returns true, the plugin’s _open method is called to set up the image metadata and image tiles. The _open method is not for decoding the actual image data.
  2. When the image data is requested, the ImageFile.load method is called, which sets up a decoder for each tile and feeds the data to it.

An image plug-in should contain a format handler derived from the PIL.ImageFile.ImageFile base class. This class should provide an _open() method, which reads the file header and sets up at least the mode and size attributes. To be able to load the file, the method must also create a list of tile descriptors, which contain a decoder name, extents of the tile, and any decoder-specific data. The format handler class must be explicitly registered, via a call to the Image module.


For performance reasons, it is important that the _open() method quickly rejects files that do not have the appropriate contents.


The following plug-in supports a simple format, which has a 128-byte header consisting of the words “SPAM” followed by the width, height, and pixel size in bits. The header fields are separated by spaces. The image data follows directly after the header, and can be either bi-level, greyscale, or 24-bit true color.


from PIL import Image, ImageFile
import string

class SpamImageFile(ImageFile.ImageFile):

    format = "SPAM"
    format_description = "Spam raster image"

    def _open(self):

        # check header
        header = self.fp.read(128)
        if header[:4] != "SPAM":
            raise SyntaxError, "not a SPAM file"

        header = string.split(header)

        # size in pixels (width, height)
        self.size = int(header[1]), int(header[2])

        # mode setting
        bits = int(header[3])
        if bits == 1:
            self.mode = "1"
        elif bits == 8:
            self.mode = "L"
        elif bits == 24:
            self.mode = "RGB"
            raise SyntaxError, "unknown number of bits"

        # data descriptor
        self.tile = [
            ("raw", (0, 0) + self.size, 128, (self.mode, 0, 1))

Image.register_open(SpamImageFile.format, SpamImageFile)

Image.register_extension(SpamImageFile.format, ".spam")
Image.register_extension(SpamImageFile.format, ".spa") # dos version

The format handler must always set the size and mode attributes. If these are not set, the file cannot be opened. To simplify the plugin, the calling code considers exceptions like SyntaxError, KeyError, IndexError, EOFError and struct.error as a failure to identify the file.

Note that the image plugin must be explicitly registered using PIL.Image.register_open(). Although not required, it is also a good idea to register any extensions used by this format.

The tile attribute

To be able to read the file as well as just identifying it, the tile attribute must also be set. This attribute consists of a list of tile descriptors, where each descriptor specifies how data should be loaded to a given region in the image. In most cases, only a single descriptor is used, covering the full image.

The tile descriptor is a 4-tuple with the following contents:

(decoder, region, offset, parameters)

The fields are used as follows:

Specifies which decoder to use. The raw decoder used here supports uncompressed data, in a variety of pixel formats. For more information on this decoder, see the description below.
A 4-tuple specifying where to store data in the image.
Byte offset from the beginning of the file to image data.
Parameters to the decoder. The contents of this field depends on the decoder specified by the first field in the tile descriptor tuple. If the decoder doesn’t need any parameters, use None for this field.

Note that the tile attribute contains a list of tile descriptors, not just a single descriptor.


The raw decoder

The raw decoder is used to read uncompressed data from an image file. It can be used with most uncompressed file formats, such as PPM, BMP, uncompressed TIFF, and many others. To use the raw decoder with the PIL.Image.frombytes() function, use the following syntax:

image = Image.frombytes(
    mode, size, data, "raw",
    raw mode, stride, orientation

When used in a tile descriptor, the parameter field should look like:

(raw mode, stride, orientation)

The fields are used as follows:

raw mode
The pixel layout used in the file, and is used to properly convert data to PIL’s internal layout. For a summary of the available formats, see the table below.
The distance in bytes between two consecutive lines in the image. If 0, the image is assumed to be packed (no padding between lines). If omitted, the stride defaults to 0.


Whether the first line in the image is the top line on the screen (1), or the bottom line (-1). If omitted, the orientation defaults to 1.

The raw mode field is used to determine how the data should be unpacked to match PIL’s internal pixel layout. PIL supports a large set of raw modes; for a complete list, see the table in the Unpack.c module. The following table describes some commonly used raw modes:

mode description
1 1-bit bilevel, stored with the leftmost pixel in the most significant bit. 0 means black, 1 means white.
1;I 1-bit inverted bilevel, stored with the leftmost pixel in the most significant bit. 0 means white, 1 means black.
1;R 1-bit reversed bilevel, stored with the leftmost pixel in the least significant bit. 0 means black, 1 means white.
L 8-bit greyscale. 0 means black, 255 means white.
L;I 8-bit inverted greyscale. 0 means white, 255 means black.
P 8-bit palette-mapped image.
RGB 24-bit true colour, stored as (red, green, blue).
BGR 24-bit true colour, stored as (blue, green, red).
RGBX 24-bit true colour, stored as (red, green, blue, pad).
RGB;L 24-bit true colour, line interleaved (first all red pixels, the all green pixels, finally all blue pixels).

Note that for the most common cases, the raw mode is simply the same as the mode.

The Python Imaging Library supports many other decoders, including JPEG, PNG, and PackBits. For details, see the decode.c source file, and the standard plug-in implementations provided with the library.

Decoding floating point data

PIL provides some special mechanisms to allow you to load a wide variety of formats into a mode F (floating point) image memory.

You can use the raw decoder to read images where data is packed in any standard machine data type, using one of the following raw modes:

mode description
F 32-bit native floating point.
F;8 8-bit unsigned integer.
F;8S 8-bit signed integer.
F;16 16-bit little endian unsigned integer.
F;16S 16-bit little endian signed integer.
F;16B 16-bit big endian unsigned integer.
F;16BS 16-bit big endian signed integer.
F;16N 16-bit native unsigned integer.
F;16NS 16-bit native signed integer.
F;32 32-bit little endian unsigned integer.
F;32S 32-bit little endian signed integer.
F;32B 32-bit big endian unsigned integer.
F;32BS 32-bit big endian signed integer.
F;32N 32-bit native unsigned integer.
F;32NS 32-bit native signed integer.
F;32F 32-bit little endian floating point.
F;32BF 32-bit big endian floating point.
F;32NF 32-bit native floating point.
F;64F 64-bit little endian floating point.
F;64BF 64-bit big endian floating point.
F;64NF 64-bit native floating point.

The bit decoder

If the raw decoder cannot handle your format, PIL also provides a special “bit” decoder that can be used to read various packed formats into a floating point image memory.

To use the bit decoder with the frombytes function, use the following syntax:

image = frombytes(
    mode, size, data, "bit",
    bits, pad, fill, sign, orientation

When used in a tile descriptor, the parameter field should look like:

(bits, pad, fill, sign, orientation)

The fields are used as follows:

Number of bits per pixel (2-32). No default.
Padding between lines, in bits. This is either 0 if there is no padding, or 8 if lines are padded to full bytes. If omitted, the pad value defaults to 8.
Controls how data are added to, and stored from, the decoder bit buffer.
Add bytes to the LSB end of the decoder buffer; store pixels from the MSB end.
Add bytes to the MSB end of the decoder buffer; store pixels from the MSB end.
Add bytes to the LSB end of the decoder buffer; store pixels from the LSB end.

Add bytes to the MSB end of the decoder buffer; store pixels from the LSB end.

If omitted, the fill order defaults to 0.

If non-zero, bit fields are sign extended. If zero or omitted, bit fields are unsigned.
Whether the first line in the image is the top line on the screen (1), or the bottom line (-1). If omitted, the orientation defaults to 1.

Writing Your Own File Decoder in C

There are 3 stages in a file decoder’s lifetime:

  1. Setup: Pillow looks for a function in the decoder registry, falling back to a function named [decodername]_decoder on the internal core image object. That function is called with the args tuple from the tile setup in the _open method.
  2. Decoding: The decoder’s decode function is repeatedly called with chunks of image data.
  3. Cleanup: If the decoder has registered a cleanup function, it will be called at the end of the decoding process, even if there was an exception raised.


The current conventions are that the decoder setup function is named PyImaging_[Decodername]DecoderNew and defined in decode.c. The python binding for it is named [decodername]_decoder and is setup from within the _imaging.c file in the codecs section of the function array.

The setup function needs to call PyImaging_DecoderNew and at the very least, set the decode function pointer. The fields of interest in this object are:

Function pointer to the decode function, which has access to im, state, and the buffer of data to be added to the image.
Function pointer to the cleanup function, has access to state.
The target image, will be set by Pillow.
An ImagingCodecStateInstance, will be set by Pillow. The context member is an opaque struct that can be used by the decoder to store any format specific state or options.

EXPERIMENTALWARNING, interface may change. If set to 1, state->fd will be a pointer to the Python file like object. The decoder may use the functions in codec_fd.c to read directly from the file like object rather than have the data pushed through a buffer. Note that this implementation may be refactored until this warning is removed.

New in version 3.3.0.


The decode function is called with the target (core) image, the decoder state structure, and a buffer of data to be decoded.

Experimental – If pulls_fd is set, then the decode function is called once, with an empty buffer. It is the decoder’s responsibility to decode the entire tile in that one call. The rest of this section only applies if pulls_fd is not set.

It is the decoder’s responsibility to pull as much data as possible out of the buffer and return the number of bytes consumed. The next call to the decoder will include the previous unconsumed tail. The decoder function will be called multiple times as the data is read from the file like object.

If an error occurs, set state->errcode and return -1.

Return -1 on success, without setting the errcode.


The cleanup function is called after the decoder returns a negative value, or if there is a read error from the file. This function should free any allocated memory and release any resources from external libraries.

Writing Your Own File Decoder in Python

Python file decoders should derive from PIL.ImageFile.PyDecoder and should at least override the decode method. File decoders should be registered using PIL.Image.register_decoder(). As in the C implementation of the file decoders, there are three stages in the lifetime of a Python-based file decoder:

  1. Setup: Pillow looks for the decoder in the registry, then instantiates the class.
  2. Decoding: The decoder instance’s decode method is repeatedly called with a buffer of data to be interpreted.
  3. Cleanup: The decoder instance’s cleanup method is called.