the size/quality tradeoff of BTIC is currently still a bit worse than with BTJ.
I was then left partly trying to think up ideas possibly for a format which would be cheaper than my current BTJ format (WRT decoding speed), but which still has reasonably good compression (unlike BTIC, which has both unimpressive size/quality tradeoffs as well as a problematic encoding process, from both a technical and legal POV).
one thought here would be basing it mostly on my "BTJ-NBCES" format, which is mostly a catch-all for a collection of features which basically broke JPEG compatibility (BTIC itself was partly based on this format).
however, the goal would still be mostly to focus mostly on relatively high-speed decoding into DXTn.
also, it would aim to retain most of the same basic features as BTJ, namely either lossy or lossless image storage, an alpha channel, extended components and layers, ... but, like NBCES and BTIC, would not be backwards compatible with JPEG.
uses same basic packaging strategy as BTIC, but I am still open to finding some way to avoid (or reduce) the costs of the FF escape coding (*1);
possibly use WHT (Walsh Hadamard Transform) rather than DCT (WHT can be done using mostly adds/subtracts and shifts, rather than fixed-point multiplications as is required for DCT or RDCT);
use 4:1:0 and a modified colorspace (*2);
*1: basically, a raw 0xFF byte is not allowed to appear in the output data, so generally has to be escaped into an 0xFF, 0x00 pair. the cost is then whenever reading a byte, it is necessary to check if it is 0xFF, and then eat the 0x00 byte if needed. actually the format supports a more elaborate scheme to reduce expansion due to recursive nesting, but this isn't used for directly reading/writing entropy-coded data, partly as it is expensive (it requires use of pushback and a finite-state-machine).
alternative possibilities (if dropping 0xFF escape-coding), would be moving over to an alternate (probably Matroska-like) packaging. it would loose the ability to treat 0xFF as a magic byte, but would eliminate needing to deal with escapes.
with a change to this extent, I could re-evaluate whether to use LE or BE bit-packing. I personally slightly prefer LE bit packing (JPEG uses BE bit-packing, whereas Deflate/DXT/... are LE, and LE is slightly nicer from a bit-arithmetic POV).
basically, Y doesn't necessarily mean "luminance" so much as it refers to a distance along a colorspace-vector, which would be "CoCgW" (essentially YCoCg, but representing a vector rather than a specific point).
the reason would mostly be to reduce the color degradation from the use of 4:1:0 colorspace, and also because basically this maps pretty well to what is needed for conversion to DXTn.
4:2:0 is basically a 2x2 grid of pixels with a shared UV;
4:1:0 is a 4x4 grid of pixels with a shared UV (or UVW).
note that if the basic DCT or WHT block remains as 8x8, this means a 32x32 pixel macroblock size (it would be 16x16 with a 4x4 DCT or WHT).
could possibly also use Paeth prediction for DC coefficients.
VLC scheme probably:
Z3P5: 3 bits for zeroes, 5 bits for value-prefix (followed by extra bits).
where 0<=Z<=5, with 6/7 for overflow cases.
Z=6: length/value overflow, prefix indicates zero count, followed by another value encoding the value (with a 6 or 7 bit prefix).
Z=7: meta escape (used for tags inside bitstream).
probably like in JPEG, 0x00 would encode an EOB.
remotely possible could also be considering possible use of SIMD (where applicable).
will think more on whether messing around with something like this is even really all that worthwhile...
put specs on my server for a few of my formats:
BTIC1 is the old/prior format.
BTIC2B is an incomplete spec for the new/possible format.
Spec for original BTJ / BGBTech JPEG format.
this is not NBCES, which lacks much real documentation, and may still be incomplete.