store the number of runs to avoid storing the last run value about 10% lower bitrate for -qscale 32 (forman & some music video) worst case bitrate increase <0.1% (lossless or low qscale) and now the bad news, even though this just adds a single subtraction and an if() into the medium sized unpack_coeffs() loop and the if() will only be false once per unpac_coeff() call, gcc produces 50% slower code, i didnt look at the generated asm yet, not sure if i want to ...