• DrNeurohax@kbin.social
    link
    fedilink
    arrow-up
    2
    ·
    1 year ago

    Oh, I’ve just been toying around with Stable Diffusion and some general ML tidbits. I was just thinking from a practical point of view. From what I read, it sounds like the files are smaller at the same quality, require the same or less processor load (maybe), are tuned for parallel I/O, can be encoded and decoded faster (and there being less difference in performance between the two), and supports progressive loading. I’m kinda waiting for the catch, but haven’t seen any major downsides, besides less optimal performance for very low resolution images.

    I don’t know how they ingest the image data, but I would assume they’d be constantly building sets, rather than keeping lots of subsets, if just for the space savings of de-duplication.

    (I kinda ramble below, but you’ll get the idea.)

    Mixing and matching the speed/efficiency and storage improvement could mean a whole bunch of improvements. I/O is always an annoyance in any large set analysis. With JPEG XL, there’s less storage needed (duh), more images in RAM at once, faster transfer to and from disc, fewer cycles wasted on waiting for I/O in general, the ability to store more intermediate datasets and more descriptive models, easier to archive the raw photo sets (which might be a big deal with all the legal issues popping up), etc. You want to cram a lot of data into memory, since the GPU will be performing lots of operations in parallel. Accessing the I/O bus must be one of the larger time sinks and CPU load becomes a concern just for moving data around.

    I also wonder if the support for progressive loading might be useful for more efficient, low resolution variants of high resolution models. Just store one set of high res images and load them in progressive steps to make smaller data sets. Like, say you have a bunch of 8k images, but you only want to make a website banner based on the model from those 8k res images. I wonder if it’s possible to use the the progressive loading support to halt reading in the images at 1k. Lower resolution = less model data = smaller datasets to store or transfer. Basically skipping the downsampling.

    Any time I see a big feature jump, like better file size, I assume the trade off in another feature negates at least half the benefit. It’s pretty rare, from what I’ve seen, to have improvements on all fronts.

    • Yote.zip@pawb.social
      link
      fedilink
      English
      arrow-up
      2
      ·
      1 year ago

      JXL has been ready for practical use for a while now - the only place where JXL support is still missing is browsers (due to Google’s politically-motivated removal from chromium). I’m not sure if anyone has tried using JXL with ML, but it’s certainly ready to be tested right now. IMO JXL has been ready since their libJXL 0.7.0 release, which happened September 2022 last year. They’re still working towards a 1.0 but every image-related application has built-in support for JXL already and it can more or less be considered ready.

      haven’t seen any major downsides, besides less optimal performance for very low resolution images

      Just to note here, to be precise AVIF starts (barely) winning at low fidelity ranges, not low resolution. Meaning if you want a blurry mess that looks like this, AVIF will compress slightly better (that’s an actual AVIF converted to PNG by the way).

      At the risk of sounding like sour grapes, this compression advantage doesn’t truly matter. This level of compression is almost never used, and even if it was, even drastic relative filesize savings would ultimately amount to bytes/kilobytes in the grand scheme of all images you’re serving. It’s more impactful to compress large images simply because they are larger. Smaller images are already small and efficiency deltas in a 1kB vs 1.1kB image are meaningless compared to a 600kB vs 800kB image.

      I also wonder if the support for progressive loading might be useful for more efficient, low resolution variants of high resolution models. Just store one set of high res images and load them in progressive steps to make smaller data sets. Like, say you have a bunch of 8k images, but you only want to make a website banner based on the model from those 8k res images. I wonder if it’s possible to use the the progressive loading support to halt reading in the images at 1k

      I’m not fully confident on this aspect but I’m pretty sure that JXL supports more than just traditional progressive decoding - you can actually pull “complete” images out of the bitstream from arbitrary ranges. Meaning you could efficiently store a full range of quality options in just one image, then serve them on the fly.

      Any time I see a big feature jump, like better file size, I assume the trade off in another feature negates at least half the benefit. It’s pretty rare, from what I’ve seen, to have improvements on all fronts.

      JXL is self-described “alien technology from the future”, and it was made by a “dream team” of image engineers who have had a hand in just about every image codec and compression technique from our past. It also benefits from being a real image codec, whereas every recent image format that has gained widespread adoption has been derived from a video codec (WebP, AVIF, HEIC).

      The only truly useful thing it doesn’t perform best-in-class at is animation encoding (losing to AVIF because it’s based on the amazing AV1 video codec), and I would honestly recommend just serving AV1 videos instead, and skipping image formats entirely.

      A neutral aspect of JXL is that it does worse in single-core decode speed compared to JPEG (which is disgustingly fast), but JXL can be parallelized whereas JPEG cannot. This is ultimately an advantage for JXL for general usecase where users have at least 4 cores available, but for large-scale distributed processing I imagine this property of JPEG may still have an edge use-case?

      If you’re curious about the technical aspects of JXL, I recommend reading their official slidedeck. The nitty-gritty details start at page 59, but the whole thing is a good read.

      • DrNeurohax@kbin.social
        link
        fedilink
        arrow-up
        2
        ·
        1 year ago

        At first glance, I probably thought JXL was another attempt at JPEG2000 by a few bitter devs, so I had ignored it.

        Yeah, my examples/description was more intended to be conceptual for folks that may not have dealt with the nitty gritty. Just mental exercises. I’ve only done a small bit of image analysis, so I have a general understanding of what’s possible, but I’m sure there are folks here (like you) that can waaay outclass me on details.

        These intermediate-to-deep dives are very interesting. Not usually my cup of tea, but this does seem big. Thanks for the info.

    • gregoryw3@lemmy.ml
      link
      fedilink
      English
      arrow-up
      0
      ·
      1 year ago

      Interesting I hadn’t thought about how the reduced image size could allow for more data throughput overall. Also great to hear that’s it’s similar or lower processing required. Although I’m not sure what tuned for parallel I/O means? Do they split the data into subgroups so multiple threads can process it at the same time?

      • DrNeurohax@kbin.social
        link
        fedilink
        arrow-up
        1
        ·
        1 year ago

        (fair warning - I go a little overboard on the examples. Sorry for the length.)

        No idea on the details, but apparently it’s more efficient for multithreaded reading/writing.

        I guess that you could have a few threads reading the file data at once into memory. While one CPU core reads the first 50% of the file, and second can be reading in the second 50% (though I’m sure it’s not actually like that, but as a general example). Image compression usually works some form of averaging over an area, so figuring out ways to chop the area up, such that those patches can load cleanly without data from the adjoining patches is probably tricky.

        I found this semi-visual explanation with a quick google. The image in 3.4 is kinda what I’m talking about. In the end you need equally sized pixels, but during compression, you’re kinda stretching out the values and/or mapping of values to pixels.

        Not an actual example, but highlights some of the problems when trying to do simultaneous operations…

        Instead of pixels 1, 2, 3, 4 being colors 1.1, 1.2, 1.3, 1.4, you apply a function that assigns the colors 1.1, 1.25, 1.25, 1.4. You now only need to store the values 1.1, 1.25, 1.4 (along with location). A 25% reduction in color data. If you wanted to cut that sequence in half for 2 CPUs with separate memory blocks to read at once, you lose some of that optimization. Now CPU1 and CPU2 need color 1.25, so it’s duplicated. Not a big deal in this example, but these bundles of values can span many pixels and intersect with other bundles (like color channels - blue can be most efficiently read in 3 pixels wide chunks, green 2 pixel wide chunks, and red 10 pixel wide chunks). Now where do you chop those pixels up for the two CPUs? Well, we can use our “average 2 middle values in 4 pixel blocks” approach, but we’re leaving a lot of performance on the table with empty or useless values. So, we can treat each of those basic color values as independent layers.

        But, now that we don’t care how they line up, how do we display a partially downloaded image? The easiest way is to not show anything until the full image is loaded. Nothing nothing nothing Tada!

        Or we can say we’ll wait at the end of every horizontal line for the values to fill in, display that line, then start processing the next. This is the old waiting for the picture to slowly load in 1 line at a time cliche. Makes sense from a human interpretation perspective.

        But, what if we take 2D chunks and progressively fill in sub-chunks? If every pixel is a different color, it doesn’t help, but what about a landscape photo?

        First values in the file: Top half is blue, bottom green. 2 operations and you can display that. The next values divide the halves in half each. If it’s a perfect blue sky (ignoring the horizon line), you’re done and the user can see the result immediately. The bottom half will have its values refined as more data is read, and after a few cycles the user will be able to see that there’s a (currently pixelated) stream right up the middle and some brownish plant on the right, etc. That’s the image loading in blurry and appearing to focus in cliche.

        All that is to say, if we can do that 2D chunk method for an 8k image, maybe we don’t need to wait until the 8k resolution is loaded if we need smaller images for a set. Maybe we can stop reading the file once we have a 1024x1024 pixel grid. We can have 1 high res image of a stoplight, but treat is as any resolution less than the native high res, thanks to the progressive loading.

        So, like I said, this is a general example of the types of conditions and compromises. In reality, almost no one deals with the files on this level. A few smart folks write libraries to handle the basic functions and everyone else just calls those libraries in their paint, or whatever, program.

        Oh, that was long. Um, sorry? haha. Hope that made sense!

        • gregoryw3@lemmy.ml
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          Yeah a bit confusing at times but overall very cool and informative, thanks so much!