One of the targets is meant to minimize decoding failure regression tests. It relies on comparing the output of the local decoder against the version on Github's default branch. One of the goals is to remove copyrighted content from a reproduction case as well.

cargo +nightly fuzz tmin regression known_reported_reproduction.jpeg

The second fuzzing target will help you generate an image for reftest. It relies on mozjpeg to keep the image correct while ensuring that a) the previous decoder failed and b) the current decoder succeeds and agrees with the output of mozjpeg. This is typically a tedious process so you'll need a lot of runs but the lib is really fast as everything happens in process. Unfortunately, it is also quite finicky as the color treatment is not the same with some deviation and seemingly mozjpeg is quite strict in what it accepts. You may need to manually edit the file structure and rely on this only for filling in some colors and index bytes that work.

To work, create a new folder, copy a known minimized regression test case in and let it go to work. You might want to supply some other jpeg samples for convenience.

tmpdir="$(mktemp)"
cp -t "$tmpdir" minimized_reproduction.jpg
cargo +nightly fuzz run fail_tmin "$tmpdir"