[18:33:02]  <michaelni> decoding works like this: for each 8x8 block
[18:33:18]  <michaelni> 1. figure out the prediction direction and dc prediction value
[18:33:28]  <michaelni> decode the VLC for the DC coeff
[18:33:44]  <michaelni> add it with the predictor to get the first coefficient
[18:34:24]  <michaelni> then decode the up to 63 run/level pairs
[18:35:16]  <michaelni> theres some escaping cases for the VLC codes but we can ignore these and just imagine each run/level pair is a single VLC code
[18:36:14]  <michaelni> so for example the run level pairs (0,3), (5,7) would result in 3,0,0,0,0,0,7
[18:37:04]  <michaelni> and that list of level values is than stored into a 8x8 array in a zigzag (or some other) order
[18:37:13]  <michaelni> at the 0,0 position is the DC value
[18:37:50]  <michaelni> then, if enabled the top row or left colunm gets the AC predicted coeffs added
[18:38:40]  <michaelni> then all are multiplied with the quantizatinon factor (not exact so but the rounding details are irrelevant for us here)
[18:38:47]  <michaelni> and then it goes in the IDCT
[18:39:03]  <michaelni> now i skiped over how the end of the up to 63 coeffs is detected
[18:39:30]  <michaelni> each VLC value doesnt just store the run and level but also if its the last VLC for a block
[18:39:57]  <michaelni> above is from memory, i hope ive not forgotten something :)