Grittibanzli is a tool to compress a deflate stream to a smaller file, which can be decoded to the original deflate stream again. That is, it compresses not only the data inside the deflate stream, but also the deflate-related information such as LZ77 symbols and Huffman trees, to reproduce a gzip, png, ... file exactly.

Usually compressing a compressed file does not work well, and Grittibanzli aims to improve this situation for compressing deflate files.

Grittibanzli splits a deflate container into 2 or 3 streams: -uncompressed data: the original data -deflate choices needed to reproduce the exact deflate stream. It uses prediction to encode it well no matter what deflate compressor (gzip -1 to -9, zopfli, ...) was used. The prediction works best for gzip -6 to -9. -when applicable: container metadata (such as gzip header)

These different byte streams then need to be compressed to produce a new file smaller than the original deflate container. Suggested compression algorithms:

-uncompressed data: Brotli quality 11 -deflate choices: PPMd, or Brotli quality 10 or 11 -metadata: Brotli quality 11

Well compressed choices will be roughly 0.5-1% for gzip -6 to -9, 9% for zopfli, 21% for gzip -1. As long as this percentage is smaller than the gain from Brotli-compressing the rest, the deflate container can be losslessly reduced.

Performing the compression is not included in this API. Grittibanzli's main result is the deflate choices bytestream, which is large but has a lot of zeroes so that you can compress it well.

This is not an official Google product.