why not compress the residulals of ac coefficients and predicted values?

[aaaaeerr]

can we improve compression rate by compress the residulals of ac coefficients and its predicted values?

[danielrh]

possibly! I haven't had luck doing so in the past, but it might be worth trying... The place where there is likely to be the biggest improvement would be in the edge coefficients... the math there must be off by a bit or something, because simply subtracting the prediction from the result yields significantly worse compression

[xiaoxiongli]

Dear @aaaaeerr, how about your result?