Lossless GIF recompression via exhaustive search

https://blog.arusekk.pl/posts/lossless-gif-recompression/

by ZacnyLos • 29 days ago

65 15 29 days ago

blog.arusekk.pl

kernelbugs• 29 days ago

Cool effort! I appreciated the background into the why, as initially I thought this would have to do with optimizing animated GIFs and didn't realize there was a usecase for single frame GIFs anymore.

chrismorgan• 29 days ago

didn't realize there was a usecase for single frame GIFs anymore

If the purpose is supporting NCSA Mosaic… I’m content to say that there isn’t. Definitely not “anymore”.

gsquaredxc• 29 days ago

Here I should note that Python is not the best choice for CPU-bound software. I want to take the opportunity to learn Zig.

For optimizing CPU-bound operations in Python, there’s some low hanging fruit with numba. I would recommend this as a 5-minute solution to you limiting your algorithm because regular Python is too slow. I regularly tell people that if their Python program is slow enough to take several minutes, you could probably learn numba before it finishes.

davidkwast• 29 days ago

Nice advice

sltkr• 29 days ago

PyPy is an option too.

EliasWatson• 29 days ago

For exhaustive searches, a library like or-tools or clingo can speed things up a lot. But first you have to figure out how to express your problem declaratively.

Stitch4223• 29 days ago

Love the goal of supporting ancient browsers. What I’m missing in the article about images is images. For example the achieved results and some table of the amount of space saved using the compression methods described. Nonetheless interesting read.

richard_todd• 29 days ago

I agree, it would be way more interesting to see a table of original GIF sizes versus the improved size. I'm left wondering how bad the plain greedy version is on a typical image.

jjcm• 29 days ago

Neat. I tested it and here were my results:

ffmpeg conversion: 1,515 bytes, 0.05s

zgif conversion: 1,479 bytes, 90.1s

I used cursor to port it to rust as well, and it got the conversion time down to 20s. Still likely not worth it as even with the rust port it's a 400x increase in processing time (that scales exponentially) for a ~2% decrease in size.

BigTTYGothGF2• 29 days ago

We need to be careful here, because finding the actual best compression for an arbitrary format can be equivalent to solving the halting problem.

I don't see why that would be relevant here. The goal is to find the best compression of a given source among all possible DEFLATE streams, not to find the best compression of a given source among all possible ways to compress data. There are only so many DEFLATE streams shorter than the original source, you "simply" try them all, and then you've halted. That said:

Now, Zopfli is the software that performs an exhaustive search across all possible syntactically valid streams

Maybe I don't understand LZ77 as well as I thought, but surely the search space is way too big to exhaust?

dj_axl• 29 days ago

Too big to exhaust... the concluding sentences of the article contain the details. "Hybrid search with pruning" is the current state of the codebase.

BigTTYGothGF2• 29 days ago

That's where OP ended up, I was disputing the description of zopfli.

ghssds• 29 days ago

We need to be careful here, because finding the actual best compression for an arbitrary format can be equivalent to solving the halting problem.

That sentence was pasted unmodified from the LLM output.

teddyh• 29 days ago

For PNG, there’s ”pngcrush”: <https://pmt.sourceforge.io/pngcrush/>

duskwuff• 29 days ago

pngcrush is pretty old. A newer, more effective option is oxipng: https://github.com/oxipng/oxipng

Lossless GIF recompression via exhaustive search

Related Stories