*.frm16 files are simple bitmap images, used for many things like UI elements, icons, minimaps, textures, portraits … In Lionheart, even more of them is actually embedded as frames in .SEQ16 and .MDL16 files. They use 16 bits per color and 8 bits for transparency (the older format probably used between 2 and 3 bits for transparency).
Supposedly they are cached form of .FRM files which in turn are created from .TGA files. To speed up loading, they do not use any JPEG-type compression like FRM files and use only RLE and 16bit colors. They reside in Cache/ subdir, which supports the theory. In some related games (Ricochet) it's allegedly possible to delete .FRM16 files and get them recreated automatically from FRM files. Unfortunately there are not any .FRM files in Lionheart.
FRM16 file consists of a header and up to five layers. Each layer consists of pixel data followed by a lookup table to first pixels of each image row.
Depending on a file type, pixels are either uncompressed (type 64), or RLE-compressed (file type 68). Only the type 68 allows transparency.
In the older incarnations of the format, layers were probably a way to encode alpha (transparency) in FRM16 files - the first layer would be opaque (possibly with fully transparent pixels in type 68), the second layer would be 80% opacity and so on with the fifth one being 20% opacity.
However in Lionheart the layers are used differently. If the image is type 68, the first layer is the same as above, but the second and third layers are never used. The fifth layer contains pixels which are semitransparent - but only their RGB colors. The alpha information for the fifth layer is in the fourth layer, but WITHOUT the RLE information. In other words, the alpha layer is unusable without the RGB one. It's stupid and makes incremental loading of type 68 images complicated.
Perhaps the 2nd and 3rd layers, unused in Lionheart, could be reused for height and normal maps for fake 3D effect (2.5D).
Header (36 bytes)
|0x06||2||Width in pixels (portraits = 90, areas = 779)|
|0x08||2||Height in pixels (portraits = 111, areas = 330)|
|0x0a||2||Flags? Either 0 (normal) or 107 (immediate)|
|0x0c||2||Type (64 or 68)|
|0x0e||2||Unknown, possibly padding, always 0 in standalone files|
|0x10||4||Bitmap size for the 1st, opaque layer|
|0x14||4||Bitmap size for the 2nd layer, unused|
|0x18||4||Bitmap size for the 3rd layer, unused|
|0x1c||4||Bitmap size for the 4th layer, alpha|
|0x20||4||Bitmap size for the 5th layer, RGB|
Each layer consists of a bitmap and a scanline lookup table (LUT).
Pixel color is encoded as a 16bit word, with a 5-6-5 RGB encoding. : How are fully transparent pixels encoded?
Type 64 files do not use any compression. Each bitmap consists of width x height pixels of 2 bytes each.
Type 68 FRM16 files use Run Length Encoding, RLE. Each bitmap line consists of variable number of pixel runs. The RLE scheme uses two highest bits in the first byte of each run. If bit 7 is set, bits 0-6 contain number of pixels to skip in the resulting image. If bit 7 is clear and bit 6 is set, bits 0-5 contain number of 16bit pixels which follow. If both bits 7 and 6 are cleared, bits 0-5 contain a number of repetitions of the following 16 bit pixel.
81 42 ff fe ff fc 05 00 00
would result in
__ __ ff fe ff fc 00 00 00 00 00 00 00 00 00 00
(First one pixel (0x81 & 0x7f) is skipped, then two (0x42 & 0x3f) whitish pixels are copied, then a black pixel is copied 5 times)
In type 68 files, the first 4 bytes in a bitmap is a DWORD containing bitmap's size in bytes. Since LUT's size is known (2 * height), it allows to read or skip the bitmap or whole layer without decoding it. However, the same number is already in the header, so its usefullness is unclear.
Array of 32bit indices into bitmap pointing to beginnings of each bitmap scanline. The first index is 0 in type 64 files, 4 in type 68 files.
|0x00||4||Offset of the 1st scanline (0 in type 64 files, 4 in type 68 files)|
|0x04||4||Offset of the 2nd scanline|
|height * 4 - 4||4||Offset of the last scanline|
[: quotes on FRM/FRM16 and JPEG-like compression] [: links to convertors by Balder, Miloch, …]