fpng has two sets of precomputed Deflate literal Huffman tables which are used in the default fast (single pass) mode. They were computed from 6k test images. You can improve fpng's ratio for specific content by computing a new set of Huffman tables and recompiling.
To do so, set FPNG_TRAIN_HUFFMAN_TABLES to 1 in fpng.h. Then rebuild fpng_test. Now, prepare a "listing" text file containing the full .PNG filenames to use to compute new Huffman tables. There must be 1 filename per line.
Then run "fpng_test -t @listing.txt". Once it's done, it'll output the C++ code containing new Huffman table array definitions which you can copy & paste into fpng.cpp.
Note these Huffman tables can be used to compress any image, not just images in your test set. However they'll perform best on content that resembles your training data.