The TGA file (.tga)

TGA is the image format of the Targa boards, early graphics and video cards for PCs. AT&T's EPICenter group defined the format in 1984; the team became Truevision in 1987, and the specification's version 2.0 followed in 1989. The name Targa stands for Truevision Advanced Raster Graphics Adapter.

The format outlived the hardware because it's easy to write, keeps every pixel exactly, and can store an alpha channel. Video, animation and 3D software adopted it for rendered frames, and game developers for textures. Today you meet TGA files as:

• Game textures and source art, often as 32-bit images with alpha, imported into engines such as Unity and Unreal Engine or converted to DDS.
• Rendered image sequences from 3D and animation software, numbered frame by frame.
• Files from older software and archives of 1990s graphics work.

For anything shown on the web or sent to people, use PNG instead: also lossless, also with alpha, much smaller, and shown by every browser. Browsers don't display TGA.

True-colour pixels are stored as little-endian numbers, so the bytes come out as blue, green, red, and then alpha in 32-bit files. 16-bit pixels pack 5 bits each of red, green and blue plus 1 attribute bit. Greyscale images use 1 byte per pixel, and colour-mapped images 1 or 2 bytes per pixel, each an index into the palette.

Bits 4 and 5 of the image descriptor (byte 17) say in which corner the pixel data starts. The default, with both bits clear, is the bottom-left corner: the first row in the file is the bottom of the picture, as in BMP. Many writers set bit 5 and store rows top to bottom instead.

A reader that ignores these bits shows half of all TGA files upside down. If a texture appears flipped in one program and not another, check byte 17: 0x20 (or 0x28 with 8 alpha bits) means top-left, 0x00 (or 0x08) bottom-left. Re-saving from GIMP, which offers a choice of origin when exporting, fixes it for picky readers.

Image types 9, 10 and 11 compress the pixels with run-length encoding. Each packet starts with a byte whose top bit says what follows and whose low 7 bits give a count from 1 to 128:

• Run packet (top bit set): one pixel value, repeated count times. 82 00 00 FF is three red pixels.
• Raw packet (top bit clear): count pixel values, stored one after another.
• The specification says a run should never cross from one row to the next. Some writers ignore that, and strict readers may then fail; most readers accept it.

RLE works well on flat areas such as UI art, masks and cartoon textures, and hardly at all on photos and noisy renders, where most packets are raw and the file can even grow slightly. It's always lossless.

For alpha, a 32-bit TGA stores 8 alpha bits per pixel and should say so in the descriptor's low 4 bits (0x08). Version 2's extension area adds an attributes-type byte saying whether the alpha is meaningful and whether colours are premultiplied by it. Many writers leave the descriptor's alpha bits at 0 or skip the extension area, so programs guess, and the guesses differ (see the problems below).

Version 2.0 (1989) kept the original layout, so old readers can still read new files and ignore what they don't understand. It added material after the image data, found through the footer:

• The footer: the last 26 bytes. Bytes 8 to 23 hold TRUEVISION-XFILE, then a full stop and a zero byte. If they don't, the file is an original-format TGA, with just a header and image data.
• The extension area: 495 bytes with an author name, comments, a date and time, a job name, the software and its version, a background colour, pixel aspect ratio, gamma, the alpha type and offsets to a colour-correction table, a thumbnail ("postage stamp") and a table of scan-line offsets.
• The developer area: tagged fields an application can use for its own data.

Many writers produce a footer with both offsets set to 0, as in our example: a version 2 file with no extra areas.

TGA's advantages today are that every graphics tool reads it and that decoding is trivial. PNG does the same job in less space for almost everything else.

• Upside down or mirrored. The origin bits weren't honoured (see above). Re-save with a top-left or bottom-left origin, or flip it in the engine's import settings.
• Transparent areas are black, white or solid. The alpha channel was ignored because the descriptor says 0 alpha bits, or the alpha type in the extension area says it isn't meaningful. Some programs also treat alpha as premultiplied when it isn't, giving dark fringes. Re-export as 32-bit with alpha.
• Red and blue are swapped. A reader took the BGR bytes for RGB.
• The file isn't recognised. With no signature at the start, a damaged or unusual header looks like any other data. Check the first 18 bytes against the table above.
• The image breaks up partway down. An RLE packet crosses a row boundary and the reader didn't expect it, or the file is truncated.
• The file is big. It's probably uncompressed. RLE helps on flat graphics; for anything else, PNG is much smaller.

• Open and edit: GIMP, Photoshop, Krita, Paint.NET, IrfanView and Affinity Photo. Visual Studio's Image Editor also opens TGA, with alpha.
• Save: GIMP's export dialog offers RLE compression and a choice of bottom-left or top-left origin. Photoshop offers 16, 24 or 32 bits and RLE.
• Convert on the command line: magick texture.tga texture.png with ImageMagick, or magick input.png -compress RLE output.tga for an RLE TGA.
• Inspect: identify -verbose texture.tga (ImageMagick) or oiiotool --info -v texture.tga (OpenImageIO) shows the type, depth, compression and orientation.

A TGA contains only pixels and a little text, no scripts. The image ID and the extension area can hold an author name, comments, the software and a date; check them before publishing a file, and remember that pixels hidden by a zero alpha are still in the file.