Strange blog post. Given that it's from a developer of proprietary and paid commercial encoders, the inclusion of their upcoming "aperture-alpha" encoder on the charts that are purportedly about JPEG XL makes this feel like a marketing piece of some fashion. Certainly, they advertise their Iris-WebP encoder throughout the post, with an implication that if you haven't paid to the author for it, you're not using WebP correctly, I guess?
Moreover, there are some methodological oddities. For example, JPEG XL decoding is tested in single-threaded mode when it's optimized for multi-threaded decoding (where AVIF has artifacts from tiling). Also, the tested command is "jxl_cli --speedtest", which means that the benchmarked time includes the warm-up procedure of the "--speedtest" flag, rather than the pure decoding time.
Another example of this is progressive rendering, where the comparison moment is chosen at a local maximum that favors AVIF. JPEG XL materializes a full frame much sooner than AVIF, and starting at, say, the 96,256 B mark, it looks better than AVIF until the full image is loaded.
I expected one of these bad-faith readings, so I can address that Aperture is mentioned once and Iris-WebP is only shown in numbers, because I have access to these encoders and thus they cannot be ignored. The only encoders I heavily advertise here are the incredible open-source AV1 encoders, that I contributed to for free and I think people should use. Also, not sure where you infer that point about WebP; libwebp is a fine encoder.
AVIF does not have artifacts from tiling any more than JPEG XL has artifacts from being JPEG XL; if you read the details post at the bottom, you'd see there's a 0.5-1.0% BD-rate regression with tiles, which is effectively a rounding error.
For progressive, JXL shows a blurry mess for the majority of its decode, while AVIF shows a crisp image that clearly shows what is in the image. Go ahead and try the demo yourself! AVIF also supports more than one layer, but I used Team JXL's image on purpose to show that even there, AVIF looks better for 90% of the decode time. You need to watch what I'm showing you instead of adopting the most bad-faith reading because some things are mentioned.
> For progressive, JXL shows a blurry mess for the majority of its decode, while AVIF shows a crisp image that clearly shows what is in the image. Go ahead and try the demo yourself! AVIF also supports more than one layer, but I used Team JXL's image on purpose to show that even there, AVIF looks better for 90% of the decode time.
I'm not sure I agree? IMO after about 30% most images look better on JXL than on AVIF, the exceptions being the pigeon, which looks better on JXL after 43% (still less than half) and the sunflower, which looks better on 66% (but you can clearly see what's on the image a lot earlier). Also, blurry convey better the idea of loading and the AVIF version may have some weird artifacts / look weirder (although that's subjective), for example the Quechua woman's eyes are very distorted on the progressive AVIF and, on the Poke bowl, some of the seeds on top of one of the top radish pieces are kind of missing / look like a shadow (while other seeds of the same size appear). In contrast the JXL version is usually blurrier and less saturated at the beginning but is more "uniform/reliable" (distorts all of the "objects" more or less the same), and later it looks finished but actually isn't (which may be a problem on its own).
Worth mentioning: their demo only uses two passes for progressive AVIF -- it's just their choice. You can have up to two more intermediate passes, so at 30% you can have something much closer to the original image.
Even still, the demo proves that AVIF can deliver a usable image with up to 3x as fewer bytes as JXL!
> their demo only uses two passes for progressive AVIF -- it's just their choice
Also worth mentioning: AFAIK "their choice" here is just the default, i.e. what `avifenc --progressive` outputs, so maybe if the default is suboptimal, it could be improved?
> Even still, the demo proves that AVIF can deliver a usable image with up to 3x as fewer bytes as JXL!
Usable for what? As a clearly loading image / placeholder, I personally like the JXL version better (as already indicated in my previous comment). As a final image I think both of them are unusable, and if that's the intention I think it would be better to encode both aiming for very low quality, maybe also reducing the resolution, without progressive loading, then compare. My understanding is that AVIF is usually better at very low bitrates, so it would probably be better, but I don't think this example "proves" that.
Well, the default in avifenc can always be changed. Do keep in mind there's no "one size fits all" implementation, as customers desire different loading tradeoffs. You might be surprised, but during testing (outside HN), we've seen people actually prefer "2 layer" loading.
I'm surprised HN likes progressive loading to be more granular, and use that to push back. I'm wondering if there are generational differences at play? Maybe it's the difference of being used to the blurhash vs. old-school JPEG loading experience.
Anyway, the more expressive mode in avifenc is `--layered` (yes, I know the name is weird).
> Usable for what?
Well, focusing on the "Poke bowl" example: with AVIF you can clearly tell apart each ingredient at 8kB. You can derive enough semantic understanding just from this base layer. Having decent edge preservation helps significantly here.
On the other hand, JXL is just too blurry at 8kB to make sense of the image -- only the egg and carrot are recognizable, maaybe the cucumber? IMO JXL needs the pass at ~28kB to make everything salient, including sprouts and beet. Yes, I know there are subjective effects at play and I'm sure we'll disagree on exact image thresholds, but recognizing objects within an image is so important in real-life use cases.
> Well, the default in avifenc can always be changed.
Sure, that was kind of my point, but I think it's important to point out it's the default. I had read somewhere else in the comments that they just "happened" to use 2 passes (that I now realized was also written by you), that I interpreted as them maybe making a bad decision in the comparison page, but I think it's very reasonable to use the default. I also usually prefer to use the default unless I have a good reason not to do so.
And I think that `--layered` is fine as a name, and I had realized that as the docs were right below the docs for `--progressive`, which do mention it encodes a "layered image", although it could use an example on how to use it effectively.
> I'm surprised HN likes progressive loading to be more granular, and use that to push back.
Not necessarily? Using more passes would probably help in specific the metric in the comment I was responding to, by making AVIF possibly look better on a larger fraction of the decoding time. I mostly interpreted your comment as saying that it would have been better if they used more passes, and that's on me.
The reason I was comparing the breakpoints in the decode time was because I was responding to an allegation that "even there, AVIF looks better for 90% of the decode time", and I decided to point out that's not not the case and chose some pretty clear breakpoints that are very hard to argue about. Very subjectively the actual breakpoints could be a lot earlier, for the Poke the AVIF only looks better (in the sense that you can figure out what's there better) in a 6% range (from 2 to 8%) or maybe 9% (from 2% to 11%) of the decode time. There are many (relatively big) seeds that are just not there on the AVIF side, but you can distinguish on the JXL side, even if they're blurry. Percentages used in reference to the JXL side.
> I'm wondering if there are generational differences at play? Maybe it's the difference of being used to the blurhash vs. old-school JPEG loading experience.
Hmmm I don't know, maybe. TBF with today's usual internet speeds progressive loading isn't as important as it was as more steps make less sense if you have a good connection.
> Well, focusing on the "Poke bowl" example: with AVIF you can clearly tell apart each ingredient at 8kB. You can derive enough semantic understanding just from this base layer. Having decent edge preservation helps significantly here.
Sure, I see what you mean, you can distinguish more image element in the AVIF example, that's true. I don't like the lack of consistency and some of the artifacts to the point where I'd prefer not using progressive encoding, so that's not what I'd call "usable", but that seems like a matter of taste. It surely does look closer to a "final image" than the JXL at that point.
As a note, I think what I'd prefer would be 2 passes, with the 1st pass being around how JXL looks at around 11% / 39 kB. You can get a rough idea of what's the image about and the elements on it, with roughly the right colors, but it's still clearly loading, so there's no confusion, and without many "artifacts" like the progressive AVIF in the example (no idea if AVIF is able to make progressive encoding in a way that's more "blurry"). The caveat is that more steps are better when it takes too long (more than a few seconds) so the user knows it's not stuck.
> Sure, that was kind of my point, but I think it's important to point out it's the default. I had read somewhere else in the comments that they just "happened" to use 2 passes (that I now realized was also written by you), that I interpreted as them maybe making a bad decision in the comparison page, but I think it's very reasonable to use the default.
Yeah, I can see that. By "happened" I meant that the default is good and serves a common use case, but also it can't be expected for a 2-pass AVIF to remotely match JXL's finely-incremental experience. The tricky thing is coming up with a good-enough compromise -- one extreme wants their first pass be more like a blurhash (quality 0, 1/8 scaling), while the other wants a medium quality image (quality 30-40, full scaling), and everybody else is in between.
For reference, `avifenc --progressive` is currently quality 10, 1/2 scaling.
> Hmmm I don't know, maybe. TBF with today's usual internet speeds progressive loading isn't as important as it was as more steps make less sense if you have a good connection.
Interestingly enough, I frequently get reminded of spotty internet connections -- turns out you just need to take the subway hah. This is why I'm so passionate about progressive image loading in general.
> As a note, I think what I'd prefer would be 2 passes, with the 1st pass being around how JXL looks at around 11% / 39 kB.
That sounds like you want your first pass be half scaling, around quality 25:
`image1.png` and `image2.png` can be the same source image. You can blur `image1.png` or even better: add a tiny "loading" icon to indicate the image is still downloading.
Oh, i left the wrong quotation whereas i intended to reply to another message. Throughout the blogpost you're quality matching and comparing encoders based on objective metrics whereas it'd be more telling to get the crowd subjective comparisons. I think it's pretty evident that most codecs benchmaxx to the point of objective metrics being useless.
Ah, no worries! I can't speak for Iris and Aperture, but both "tune IQ" modes in SVT-AV1 and libaom had extensive human evaluations to make sure they weren't accidentally being benchmaxxed at the expense of subjective quality.
- Upload the problematic image to the AOM analyzer
- Press 'L' to show tiles view
- How many tiles (yellow rectangles) do you count?
- Are artifacts actually at those tile borders?
I was even more surprised & disappointed to see this dismissed as an ad for my company. Think about how fantastic JPEG XL would be for my company: an image codec with tons of fans and an extremely weak reference encoder? Sign me up. I'm writing this as someone who wants a better Internet.
If you're sensitive to tiling artifacts, you must deeply dislike the way JPEG XL images look. Every block of a JPEG XL image is effectively a small tile, because there is no deblocking loop filter (I talk about this in the article).
JPEG XL has two filters that are pretty much equivalents of what AVIF has, while AVIF has three; plus, tile boundaries are handled by the DLF. Your point doesn't really stand on two legs if it means to come to JPEG XL's defense here.
By "small", it's meant in a relative sense. JXL's coded blocks can be as big as 64x64, and block boundaries can create visible seams. Example: https://juliobbv.com/pics/photo.jxl
These seams are especially noticeable in the background, and align with coded block edges. You can tell the encoder is trying to conceal them as best as it can, but this cannot be properly mitigated without a proper deblocking filter.
For comparison, AV1 tiles normatively go through the deblocking filter.
I participated in the design of those filters, so no, I do not deeply dislike the way they look.
This gaslighting is not convincing. No matter how many filters AVIF has, I distinctly remember tile artifacts being particularly disturbing. More so than individual blocks, whose size and border effects differ; tiles are a straight line through the entire image.
JPEG XL does not have this problem because the design and codestream enables parallel decoding (thanks to per-group offsets encoded in the 'TOC'), hence does not require separate tiles.
Hi there! I'm Julio (co-developer of libaom and SVT-AV1's tune IQ). Here there are some points worth mentioning, because I'm catching a whiff of bad faith with your comment that honestly needs to be called out:
- The inclusion of his two proprietary encoders (Aperture and Iris) just serves to further support the argument that JXL encoder devs have work to do to perform at the frontier, while also proving you only need a person or two to do so. The two FOSS AV1 encoders in the compo (libaom and SVT-AV1) are enough to prove this. Given that blog posts often double up as a way to show-case projects, I think it's fair game to show off a bit. Also, keep in mind Gianni is just 21 and starting his career -- reporting such strong efficiency results across several image formats (AVIF, WebP, Aperture) is impressive and worthy of celebration by the community!
- Tiles in AV1 go through the deblocking filter, so there won't be any seams after decoding. In fact, AVIF encoding solutions (like libavif) enable tiling by default. If there were seams, people would've noticed those artifacts and yelled at the libavif maintainers.
- *Because* JXL doesn't have a deblocking filter, you could argue that JXL effectively decodes to numerous "mini-tiles" -- each one equaling the size of a coded block. And indeed, you WILL see those boundary artifacts when quality isn't high enough for EPF, Gaborish and/or LF smoothing to mitigate satisfactorily. This is what Gianni's post covers.
- AVIF scales very well under multithreaded decoding scenarios, thanks to the excellent work of the dav1d devs. The main conclusion wouldn't have changed -- AVIF is significantly faster to decode than JXL.
- In progressive decoding, a very valuable feature is "bytes to first usable image". That's what the comparison is focusing on -- it's not a cherry-picked point at all. By usable: you can tell the pass isn't a "blurhash", but you can actually discern each element in the picture with reasonable detail. You can play with the JXL demo yourself -- JXL roughly needs 3x as many bytes to get to where AVIF is in quality, and JXL is still a bit more blurry in general. This applies to every image in the demo, not just the poke bowl.
- The folks who coded the JXL demo happened to use two passes for progressive AVIF, but you can use up to four -- including adding an even lower-quality "blurhash" pass, and/or a medium quality pass. Yes, it's desirable to control the number of passes and quality at the encode stage.
There are two Graphite companies. The time series DB for metrics (not this) and the stacked diff code review platform (this). Looking at other comments under the post, they seem to have executed a hard AI pivot recently.
4o on ChatGPT.com vs. Opus in an IDE is like cooking food without kitchen tools vs. using them. 4o is neither a coding-optimized model nor a reasoning model in general.
Moreover, there are some methodological oddities. For example, JPEG XL decoding is tested in single-threaded mode when it's optimized for multi-threaded decoding (where AVIF has artifacts from tiling). Also, the tested command is "jxl_cli --speedtest", which means that the benchmarked time includes the warm-up procedure of the "--speedtest" flag, rather than the pure decoding time.
Another example of this is progressive rendering, where the comparison moment is chosen at a local maximum that favors AVIF. JPEG XL materializes a full frame much sooner than AVIF, and starting at, say, the 96,256 B mark, it looks better than AVIF until the full image is loaded.
reply