Hey Thiemo, On 04.12.2017 10:43, Thiemo Kreuz wrote:> I consider myself an image file format nerd, so thanks a lot for > sharing this! FLIF was new to me.Don't mind it! :)
> I would like to share two important notes: > > 1. Unfortunately the flif.info website does not say a word about the > CPU resources their current implementation burns when converting a, > let's say, PNG to FLIF.Well, currently it's single core and not optimized at all. You should know CABAC encoding from x264/x265 so it's slow, but not dead slow. The website doesn't mention it because it highly depends on the subject as well as the setting on encoding named effort. Currently, effort is default 60, I tried 100 a lot, but there's nearly nothing to gain. So I assume we always want to use the good default. :) PNG Encoding of the current featured picture of today[1] at a medium image size for the web, 1.280×868 Pixel, opened in Gimp, converted to sRGB and exported as maximum compression without any checkbox done in Gimp ... takes Gimp 3 seconds to write it to the Harddisk. Transcoding this PNG to FLIF takes on my machine (i3-2330M@2.20GHz; 12 GB RAM): real 0m7,405s user 0m7,320s sys 0m0,053s decoding the file again to PNG on the same machine (with FLIF) real 0m1,077s user 0m1,067s sys 0m0,007s For comparison, we save 708 KByte in comparison to PNG in this case, and the PNG exported by FLIF is just 14 KByte bigger than the one created by Gimp. > It's pretty important to realize that CPU > resources are even more valuable than storage space and network > bandwidth. Sure, It's not really possible to come up with an exact > threshold. But if, let's say, converting a PNG to FLIF saves 100 KiB, > but takes a minute, this minute will never pay off. So my Point was more: Get rid of this bloody Download a JPG, do some Stuff & Upload a 3 Times locally saved JPG again, calling it an improvement. > If you follow the discussions on Wikimedia Commons, you will find this > argument quite often: Downloading PNGs, optimizing them, and uploading > them again is practically never worth it. Think of all the resources > that are burned to do this: CPU time, download and upload, database > storage and time, disk storage for the new revision, and not to forget > the user doing all this. Yep, but enabling FLIF for new files would do nothing to the old Data at all, I don't want to start a discussion about converting petabytes of Data, but all new revisions should be done in FLIF, if you ask me. > Sure, your suggestion avoids a lot of this. But the CPUs involved will > experience heavy load, both on the server as well as clients that need > to recode FLIF files via a JavaScript library first. FLIF is very fast to decode via JavaScript, and in general, as the example shown above, it takes just 1 second to decode and encode a medium size image as PNG with just one thread on a pretty outdated notebook with an unoptimized decoder and encoder. :) Try adding a FLIF to a website and test out if the website load anywhat slower with the FLIF ... at the small image sizes you get on articles, the performance impact is neglectable and comparable to loading a font file to the browser. > 2. Lossy file formats like JPEG should never be converted to lossless > formats. This will actually decrease quality (over time). The problem > is that the image data will still contain the exact same JPEG > artifacts, but the fact that it was a JPEG (and how it was encoded) is > lost. Yes, but those images should never be saved as JPG in the first place. Even under Android RAW-Photography is going to be a thing. FLIF just started to get rudimentary RAW-capabilities, which means you can just convert the special RAW-File to a FLIF and upload it with any loss in quality. > No tool specialized in squeezing the maximum quality out of > lossy JPEGs can work anymore. And there are a lot of super-awesome > tools like this. Not only tools like JPEGCROP and such that can cut > and even combine JPEGs without actually recoding them. Well, if you really want to start a discussion about a lossless rotation of JPG files, let me guess how many uploads are rotated losslessly... 0.0002%? > There are also "JPEG repair" filters that know how to reverse the JPEG encoding > algorithm and can squeeze out a tiny little bit of extra information > regular JPEG decoders can't see. Great! Just recommend this for new uploads to be done if the source material is JPG, let the ppl convert it with this to PNG and then to FLIF or ask the FLIF maintainers if they want to add this as an import-filter, for FLIF itself! :) But the your argument was: "Think of all the resources that are burned to do this: CPU time, download and upload, database storage and time, disk storage for the new revision, and not to forget the user doing all this." Which perfectly fit's here too. On a 20 MPixel picture small JPG Artefacts are no issue at all, but users which Download the JPG and upload it again, after doing some needed work to it, like cropping or color enhancement, this is a problem. Each version get more artifacts and we constantly get a loss of quality which each revision... I don't think we should have accepted JPGs in the first place. :) > This said, I'm all for adding FLIF to the list of allowed file formats! Wonderful, hope I get some feedback from you on the things I pointed out! :) If you want to let a note on the GIMP/eog enhancement request, feel free to do so: https://bugzilla.gnome.org/buglist.cgi?quicksearch=flif [1] File:Umschreibung by Olafur Eliasson, Munich, December 2016 -04.jpg Best regards Ruben _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
