Posted by AshleysBrain 11 hours ago
For that tool you need to output an uncompressed target and use the reference command line tools, preferably libjxl and cjxl (which is still the reference, while jxl-rs remains the experimental), to create the JXL. You will have better compression with better quality, minus the glitches.
The main two contenders were .avif and .webp. For a few reasons, I selected .avif and in hindsight I think it was the better, choice. Both would be objectively better than jpg or png.
With JPEG XL ... hmmm. AVIF is not perfect, in particular when the compression rate is very high I noticed that some photos lose a lot of intrinsic quality that is not instantly obvious; I noticed this when I took various pictures over the years from outdoors. Still, AVIF beats jpg and png just about on every metric when compression is required. With JPEG XL I guess I have to re-evaluate, but right now I am still sticking to avif. The two factors that will be important for me is compression ratio and quality. I am ok with a bit of loss of quality, if the compression is better, but I am not sure JPEG XL beats AVIF here clearly either, so I am not sure what to do with JPEG XL.
https://news.ycombinator.com/item?id=49690554
The author, who has worked on said AV1 encoders, has expressed doubt that JPEG XL encoders could improve much with only reasonable amounts of effort, but that was also only their personal (if educated) guess. With JPEG XL finally seeing adoption on the web, we will hopefully see work on the reference encoder resume in earnest soon.
So for now, I’d say you’re fine with AVIF. Unless you’re doing lossless, because AVIF sucks at lossless. (And JPEG XL has the unique feature of supporting lossless JPEG transcoding, too.)
How? Line-based loading is a very simple way to exploit 2D spatial coherence. A 1D format like gzip loses this advantage for minimal gain in simplicity.
And lossless jpeg-xl exceeds your idea
Your llm has JPEG xl programs in the training data. You copied from GitHub.
One to embody power, the other one to crave it?
I'm confused about the connection.
What counts as a “safety mechanism”?
https://chromium.googlesource.com/chromium/src/+/main/docs/s...
If you use a safe language like Rust, you can afford untrusted input and no sandboxing.
If you use an unsafe language such as C++, you must chose between trusted inputs or sandboxing.
I was curious as to where rfgplk's implementation lay within the Venn diagram.
> you can afford untrusted input and no sandboxing.
I would not trust a rust program un-sandboxed! There are security bugs in rust itself.