PHP, ImageMagick and Cropping to GIF: Digging into GIFs again!

Christer had an interesting case today, where he tried to resize and crop an image with the Imagick extension for PHP. Everything went as planned, the image was cropped and resized at it should be, but after writing it to disk and opening it again, the image’s size was the same as if he hadn’t done the crop. The content of the image outside the crop area was removed (simply set as transparent), but the image was still returned in it’s uncropped size.

The PHP module for binding ImageMagick is quite simple (simply marshalling between the ImageMagick methods and the PHP user space), so my guess is that this is a weird behaviour with a good enough reason somewhere down in ImageMagick. It might be a bug, but I haven’t had the time to attempt to reproduce it with convert or mogrify yet. If anyone wants to attempt that, feel free. Christer has posted the code, so simply attempt to recreate the same symptoms by using one of these two tools.

Anyways, this post was not to be about the issue itself, as Christer has done a neat analysis and write-up of that, but I’ll give a more detailed look at the issue within the GIF file itself. As chance would have it, I recently participated in a competition at the norwegian demoscene IRC hangout where the goal was to recreate the norwegian flag in an HTML page in the smallest space possible. This ended up being a competition to see who could molest and optimize GIF images the most, while browsers still were able to display them. From this experience I had a quite good knowledge of how GIF files are built internally, and I were able to do a good guess of what could be the actual issue in the resulting file.

Since GIF files can be animated, a single file may contain several “images” (which would be the frames in the animation). These images can have their own size and position within the “larger image”:

| im1     |
|    _____|
|   |     |
|   | im2 |

im1 may then represent the first image and im2 the second image in the file. The second image will only update the area that it covers, and this will leave the rest of the image “as it is”. Since a GIF image may contain a large number of these images, a “global” size is defined for the image. This global size covers all the images, and is the total area that these images will be drawn into. If an image is drawn outside of this area (in part or whole), it will be clipped against the viewport.

This should provide enough background to at least give a general feeling about what COULD be the problem here, but to actually find out what’s happening, we’ll dig into a GIF file format specification and the file that was created. This simple reference provides a general layout of the GIF file, and we’ll use that to take a look at what values the file we ended up with had:

On the left we have the actual byte values in hex and on the right we have the corresponding ASCII character represented by that value. As you can see, the first six bytes of the file (0x47 0x49 0x46 0x38 0x39 0x61 (0x is the general way of prefixing numbers that should be interpreted as hexadecimal)) corresponds to “GIF89a” (You can do this exercise yourself armed with this Ascii Table. Simply look up 47 in the Hx column, then 49, etc). Those six bytes are what we call the signature of a GIF file (although the number can be different, i.e. GIF87a, depending on the version used).

The next fields in the specification reads:

Offset   Length   Contents
  6      2 bytes  
  8      2 bytes  

So byte 6-7 and byte 8-9 should tell us the logical size of the whole gif file (which the images will be drawn onto). In our test file here, that’s represented as:

Width: 0x67 0x01
Height: 0x70 0x00

The byte order here is Little Endian, which means that the least important values are placed first. Since we have two bytes for each value, we can calculate the decimal value of the width by multiplying:

0x67 0x01 = 6 * 16 + 7 + (0 * 16 + 1) * 256 = 359
                                        ^-- Since we're in the next byte, we multiply with 256.

You can also do this with the windows calculator, by entering 167 while being in hexmode, then selecting dec (for decimal). The reason for multiplying the second byte with 256 is that this byte provides the value of the “next 8 bits”, while the first provided the value for the first 8 bits. If we see the bits themselves:

0x70 | 0x01: 0111 0000 | 0000 0001

Little Endian says that the least significant bits come first, so to get the raw bit values, we turn it around:

0000 0001 0111 0000

As you can see, the value of the second byte (0x01) can be multiplied with 256 (which is the last 8 bits).

We can also calculate the height:

0x70 0x00 = 7 * 16 + 0 + (0 * 16 + 0) * 256 = 112
                          ^-- both numbers in the second byte is zero

Alas, the global header of the GIF image that were generated says that the size of the image is 359×112, which is why the image is rendered larger than it should have been. We then take a look at the Image section of the GIF file (all GIF files should contain at least one), which is defined as:

Offset   Length   Contents
  0      1 byte   Image Separator (0x2c)
  1      2 bytes  Image Left Position
  3      2 bytes  Image Top Position
  5      2 bytes  Image Width
  7      2 bytes  Image Height

Armed with this information, we examine the area where the image section starts:

The start of the Image section is the “Image Separator”, a byte value of 0x2c, shown highlighted in the image above. This is where the image section starts, and the offsets in the table is relative to this location. The next four bytes tells us where in the global viewport the upper left corner of this image should be drawn. The values here are 0x01 0x00 twice, simply meaning (1,1), or one pixel down and out from the upper left corner (which is also related to the issue posted by Christer, but we ignore that one here now). The next values are however those we are interested in, which provides Image Width and Image Height:

0x73 0x00 = 7 * 16 + 3 + (0 * 16 + 0) * 256 = 115
0x6F 0x00 = 6 * 16 + 15 + (0 * 16 + 0) * 256 = 111

This means that the dimension of the image that’s actually supplied in the GIF file, is 115×111 pixels and should be drawn beginning one pixel down and one pixel out (as given by 0x01 0x00 in the x,y-fields above). Compare this to the reported global size of the image (359×112), and we can see where our transparent space is coming from. The browsers (and other image viewers) create a canvas the size of 359×112 pixels, while only drawing an image into the 115 leftmost pixels. The rest is left transparent, but they’re still there as the file says that’s the size of the viewport. If we manually change the size of the viewport to 0x74 0x00 in the GIF header itself, the image displays properly. To illustrate with another great ascii drawing:

|           |                         |
|  actual   |                         |
|  image    |                         |
|  drawn    |                         |
|           |                         |
|           |                         |
|           |                         |

The solution to the problem here were to call the setImagePage method of the image object, as that allows us to set the values for the global image ourselves (and we know how wide the image were supposed to be).

Bonus knowledge: This issue did not occur when saving to a JPEG file, as JPEG files does not have the same capability of storing several subimages inside one file, and does not have the same rendering subsystem as GIF files. ImageMagick knows this, and does not use the page-values when rendering the file.

Hopefully this has provided a minor introduction into how files are structured, what you can learn armed with a hex editor and a file format specification and provided a few insights into what you can do when you’re faced with a very weird problem.