Silicon Graphics Image

Silicon Graphics Image (SGI) or the RGB file format is the native raster graphics file format for Silicon Graphics workstations. The format was invented by Paul Haeberli. It can be run-length encoded (RLE). FFmpeg and ImageMagick, among others, support this format.

Background
Common file extensions are:


 * or : 3 colour channels
 * : 3 colour channels and alpha
 * or : black and white
 * : black and white and alpha

This format was originally developed for IRIX. The master files of the SVT High Definition Multi Format Test Set are SGIs. Frame 200 of the ParkJoy sequence in this set (15722.sgi, 3840×2160, 47.4 MB) was used in WebP comparisons.

Technical Details
SGI image format use big endian order. Image data scan from left to right from bottom to top.

Header
File Signature/Magic Number

Should equal    for SGI file.

Compression

if the file is uncompressed;  for RLE compression.

Bytes Per Pixel Channel

Value is  for 8 bits per channel ;   for 16 bits per channel.

Dimension

Value is  for 1 channel, 1 scanline image ;   for 1 channel, multiple scanlines image;   for multiple channels, multiple scanlines image (most common).

X Size

Scanline width in pixels.

Y Size

Number of scanlines in the image.

Number Of Channels

Number of channels in the image. Value is  for greyscale images;   for RGB;   for RGBA. SGI supports more than 4 channels.

Minimum Pixel Value

Smallest pixel value in all channels. Since SGI uses unsigned char or short types for channels, the smallest possible value is typically.

Maximum pixel value

Largest pixel value in all channels. Since SGI uses unsigned char or short types for channels, the largest possible value is typically  for 8 bit/channel and   for 16 bit/channel.

Dummy

4 bytes, unused.

Image name

Image name, 80-byte (maximum 79 characters ≠ ) C string. If the name is shorter than 79 characters, all subsequent bytes are to be  in order to fill the 80-byte space.

Color Map ID

SGI supports 4 types of color mapping:
 * 1)   normal: image has 1 channel (greyscale), 3 channels (RGB) or 4 channels (RGBA).
 * 2)   dithered: image has packed RGB. First 3 bits are for R channel, next 3 bits are G channel, and last 2 bits are for B channel.
 * 3)   screen: image uses colormap. Pixel data is index in color map and image only have 1 channel.
 * 4)   colormap: file only have data for one color map (can use for different image file). This file no have image data.

Dummy

404 bytes. Typically unused, scales the header to 512 bytes long.

Image Data
No Compression

Image data starts at byte 512 in file. Data length can be calculated using the following:

(X size) × (Y size) × (Num. of channels) × (Bytes per channel)

The data is separated for each channel. For example, if an image uses RGBA, the first set of scanlines represent the R channel, the next set of scanlines are for the G channel, then B channel and, lastly, A channel. Therefore, each set of scanlines can be accessed directly via the following formula (where C = channel index):

(X size) × (Y size) × (C ÷ Num. of channels) × (Bytes per channel)

RLE Compression

Offset tables start at byte 512. Their length can be calculated with the following:

(Y size) × (Num. of channels) × 4 bytes long

Each unsigned int in a table represent an offset (from file start) to the start of the compressed data of each scanline in each channel. The data is separated for each channel's scanline. If an image has RGBA channels, the order stays R, G, B and A.

Next is the data length table which can also be calculated with the formula:

(Y size) × (Num. of channels) × 4 bytes long

This table holds the size of the compressed data (unsigned int) of each scanline. This size should equal the difference of numbers in the offset table.

The last data block holds RLE compressed image data. If the number of bytes per channel is, the compressed data is of unsigned char type. Otherwise, if the number of bytes per channel is, the compressed data is of unsigned short type. The last byte of compressed scanline data has a value of.

(NOTE: The following paragraph could not be translated by the previous english editor)

For both case have unsigned char or unsigned short data, lowest 7 bits tell count. If bit 8 is zero, count is number times copy value after count. If bit 8 is one, count is number times copy series of values after count byte. For 2 bytes/channel image, count and value is unsigned short and top byte of count always equal zero.

(NOTE from english editor: what?)