The functions that have been optimized for AVX are primarily floating point. I believe many of the functions used by the UIC JPEG encoder are integer based. For a complete list of the IPP functions that have been hand-tuned for AVX, please see this article:
Note that the UIC JPEG encoder also takes advantage of multi-threading, which can provide substantial acceleration, in addition to SIMD. So you may simply be seeing a larger percentage of acceleration due to the multi-threading when compared to the SIMD. The combination of the two will generally give the best results, which is what you're seeing with the UIC implementation.