[{"data":1,"prerenderedAt":459},["ShallowReactive",2],{"project-23652753":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":35,"aspects":36,"strip":47,"post_text":49,"what":50,"said":74,"maker_in_thread":178,"shape":187,"grid":192,"attempts":420,"same_week":421,"maker_shows":33,"maker_hidden":451,"reader_version":452,"card":453},"23652753","HiFiC","High-Fidelity Generative Image Compression – demo and paper","title","Show HN: HiFiC – High-Fidelity Generative Image Compression – demo and paper","https:\u002F\u002Fhific.github.io","hific.github.io","atorodius",1593183827,2020,281,89,149,3028,5,"legend","ai","generative image compression","other",[24,25,26,27],"image compression","generative adversarial network","learned compression","neural codec","debated",0.025,{"positive":31,"neutral":32,"mixed":33,"critical":34},18,27,1,16,62,{"idea":37,"performance":40,"execution":42,"design":44,"docs":46},{"positive":38,"critical":39},10,3,{"positive":39,"critical":41},6,{"positive":43,"critical":18},2,{"positive":33,"critical":45},0,{"positive":45,"critical":43},[18,18,39,39,33,39,33,18,48,18,39,33,39,18,39,39,39,33,39,39,33,45,33,18,39,33,48,45,18,33,45,48,33,18,33,18,18,39,39,39],4,null,{"problem":51,"audience":49,"pricing":49,"motivation":62,"time_to_build":49,"stack":66,"openness":71,"stage":72,"team_size":73},{"value":52,"source":53,"evidence":54},"Compressing images to very small sizes while keeping them looking high fidelity","maker",[55,58],{"id":56,"by":11,"text":57},"23653490","Hi, author here! This is the demo of a research project demonstrating the potential of using GANs for compression. For this to be used in practice, more research is required to…",{"id":59,"by":60,"text":61},"23653952","gtoderici","That was exactly the goal of the project! Basically if the size doesn't allow to have detail, we need to \"hallucinate it\". This of course is not necessary if there's enough…",{"value":63,"source":53,"evidence":64},"Research demonstrating the potential of GANs for image compression",[65],{"id":56,"by":11,"text":57},[67],{"value":68,"source":53,"evidence":69},"TensorFlow",[70],{"id":56,"by":11,"text":57},"unknown","prototype","team",{"praise":75,"objections":94,"requests":116,"questions":143,"alternatives":158},[76,88],{"value":77,"source":78,"evidence":79},"Results beat JPEG and BPG by a wide margin at low bitrates","commenter",[80,84],{"id":81,"by":82,"text":83},"23654040","ricardobeat","This is amazing. Not only it completely blows JPEG and BPG out of the water in efficiency, but it makes some images look better than the original! It seems to reduce local…",{"id":85,"by":86,"text":87},"23653715","ksec",">Currently, this runs at approximately 0.7 megapixel\u002Fsec on a GPU with unoptimized TensorFlow. Forgive me, is that in encoding or decoding? And size of library required to decode…",{"value":89,"source":53,"evidence":90},"Noise and skin texture preserved well",[91],{"id":92,"by":11,"text":93},"23654251","Thanks! Noise is actually preserved really well, and is one of the strengths of using a GAN. Check out this visual example from the \"All Evaluation Images\" link:…",[95,109],{"value":96,"source":78,"evidence":97},"Fine details like small text and faces are regenerated rather than real",[98,102,106],{"id":99,"by":100,"text":101},"23653857","fiddlerwoaroof","Yeah, it seems like the HiFiC version has the same large-scale structure, but the details are sort of \"regenerated\": e.g. it's as if the image format says \"there's some hair…",{"id":103,"by":104,"text":105},"23654236","jasonjayr","This is impressive; but isn't this also susceptible to the type of bug that Xerox copiers got hit with many years ago?…",{"id":107,"by":11,"text":108},"23654161","Thanks for the kind words! We hope that the algorithm we presented is a step in the right direction, and we acknowledge that there's more work to be done! Just like with any…",{"value":110,"source":78,"evidence":111},"Model size (726MB) and slow speed make practical use hard",[112,115],{"id":113,"by":11,"text":114},"23653605","Thanks for the kind words! > What size library would be required to decode these type of images? The model is 726MB. Keep in mind that this is a research project - further…",{"id":56,"by":11,"text":57},[117,128],{"value":118,"source":78,"evidence":119},"Compare against AVIF and JPEG XL",[120,124],{"id":121,"by":122,"text":123},"23656903","keyi","Have you compared the result with latest image compression formats, such as AVIF[0]? [0]: https:\u002F\u002Faomediacodec.github.io\u002Fav1-avif\u002F",{"id":125,"by":126,"text":127},"23659865","bscphil","I would add JPEG-XL in addition if you're looking for suggestions for other codecs to compare to. It's very competitive with AV1 and beats it, in my opinion, at higher bitrates.…",{"value":129,"source":78,"evidence":130},"Extend to video with temporal consistency",[131,135,139],{"id":132,"by":133,"text":134},"23654126","crazygringo","I've been waiting for something to implement this concept for so long, and I'm so happy to finally get a chance to explore how it works in practice! It's really fascinating to…",{"id":136,"by":137,"text":138},"23656390","isoprophlex","Super exciting! Fantastic work and cool demo! I hope this can be 'stabilized' to allow video compression. If you run eg. style transfer networks on video, the artifacts in the…",{"id":140,"by":141,"text":142},"23655332","lokl","Has this been extended to video (beyond compressing each frame separately)?",[144,151],{"value":145,"source":78,"evidence":146},"How the adversarial loss works",[147],{"id":148,"by":149,"text":150},"23655406","nbardy","> The adversarial loss \"learns\" what is a compressed image and tries to make the decoder go away from such outputs. Could you expand on this point",{"value":152,"source":78,"evidence":153},"Training data and compute needed to replicate",[154],{"id":155,"by":156,"text":157},"23660206","littlestymaar","Amazing work! It's something we knew would come at some point, but I didn't expect it to be that good so early! Since some of the authors are in this thread answering questions,…",[159,166,170,174],{"value":160,"source":78,"evidence":161},"JPEG",[162],{"id":163,"by":164,"text":165},"23654407","stereo","The obvious question when looking at the comparison is, what kind of jpeg did they pick, and was it fair? It turns out that mozjpeg, which is pretty much state of the art, can…",{"value":167,"source":78,"evidence":168},"BPG \u002F HEVC",[169],{"id":85,"by":86,"text":87},{"value":171,"source":78,"evidence":172},"JPEG XL",[173],{"id":125,"by":126,"text":127},{"value":175,"source":78,"evidence":176},"AVIF",[177],{"id":121,"by":122,"text":123},{"comments":31,"answered":38,"key":179},[180,181,184],{"id":56,"text":57},{"id":182,"text":183},"23655937","The idea is that any distortion loss imposes specific artifacts. For example, MSE tends to blur outputs, CNN-feature-based losses (VGG, LPIPS) tend to produce gridding patterns…",{"id":185,"text":186},"23655355","Hi- no the images on the demo page are not part of the training set, they are only used to evaluate the method. Arbitrary images of natural scenes will look similar at these…",{"commenters":188,"max_depth":189,"maker_replies":31,"first_comment_mins":190,"lifespan_hours":191},47,7,38,291,[193,195,199,202,205,208,211,214,218,220,223,226,229,232,236,239,242,244,245,247,251,255,258,260,261,262,264,268,270,273,276,279,280,282,283,286,288,291,292,294,297,299,302,304,307,310,314,317,319,321,323,326,328,330,333,336,337,341,343,346,348,350,351,355,357,358,361,363,366,369,371,374,376,378,382,385,388,390,392,394,397,400,402,405,409,411,415,417,418],[132,45,18,194,133,45],"project\u002Fpositive\u002Fidea",[196,45,39,197,198,45],"23654209","project\u002Fneutral\u002Fidea","dividuum",[200,45,18,194,201,45],"23654265","swiley",[203,45,18,204,86,45],"23654483","project\u002Fpositive\u002Fperformance",[206,45,39,207,133,45],"23654678","project\u002Fneutral\u002Fperformance",[209,45,45,210,86,45],"23655878","commenter\u002Fneutral\u002Fnone",[212,45,18,194,213,45],"23657221","IshKebab",[215,45,33,216,217,45],"23657831","project\u002Fcritical\u002Fidea","phs318u",[219,45,33,216,133,45],"23657991",[221,45,33,216,222,45],"23658568","sitkack",[224,45,33,225,217,45],"23659944","project\u002Fcritical\u002Fexecution",[227,45,33,225,228,45],"23658476","viraptor",[230,45,18,194,231,45],"23658757","rrrrrrrrrrrryan",[233,45,45,234,235,45],"23659025","other\u002Fneutral\u002Fnone","dehrmann",[237,45,33,225,238,45],"23659126","gfody",[240,45,18,194,241,45],"23660030","bamfito",[56,33,45,243,11,33],"maker\u002Fneutral\u002Fperformance",[85,33,18,204,86,45],[246,33,39,243,60,45],"23653774",[248,33,39,249,250,45],"23772261","project\u002Fneutral\u002Fnone","JyrkiAlakuijala",[252,33,18,253,254,45],"23653745","project\u002Fpositive\u002Fexecution","petermcneeley",[256,33,39,257,60,45],"23653792","maker\u002Fneutral\u002Fdesign",[148,33,39,259,149,45],"maker\u002Fneutral\u002Fnone",[182,33,45,257,11,33],[99,33,33,225,100,45],[59,33,39,263,60,45],"maker\u002Fneutral\u002Fidea",[265,33,39,266,267,45],"23655272","project\u002Fneutral\u002Fexecution","UnbugMe",[185,33,45,269,11,33],"maker\u002Fneutral\u002Fexecution",[271,33,39,272,267,45],"23655518","maker\u002Fneutral\u002Fdocs",[274,33,18,194,275,45],"23656022","cs702",[277,33,45,278,60,45],"23656813","commenter\u002Fneutral\u002Fidea",[121,33,39,207,122,45],[281,33,39,243,60,45],"23657004",[125,33,39,207,126,45],[284,33,39,266,285,45],"23657086","magicalhippo",[287,33,45,269,11,33],"23660328",[289,33,39,207,290,45],"23668171","belval",[103,43,33,225,104,45],[293,43,45,269,11,33],"23654489",[295,43,39,197,296,45],"23654548","mmastrac",[298,43,39,197,60,45],"23654570",[300,43,39,197,301,45],"23654744","AbrahamParangi",[303,43,45,263,11,33],"23654864",[305,43,39,207,306,45],"23655220","the8472",[308,43,45,309,11,33],"23655251","project\u002Fcritical\u002Fdesign",[311,39,33,312,313,45],"23655820","project\u002Fcritical\u002Fdocs","nullc",[315,39,45,210,316,45],"23655906","ebg13",[318,39,45,210,313,45],"23655990",[320,39,45,272,11,33],"23656046",[322,39,33,312,313,45],"23656421",[324,39,45,325,316,45],"23656544","commenter\u002Fneutral\u002Fdocs",[327,39,45,272,11,33],"23656762",[329,39,18,194,313,45],"23659105",[331,48,39,207,332,45],"23653226","deweller",[334,48,39,207,335,45],"23653532","ache7",[113,48,45,243,11,33],[338,48,33,339,340,45],"23656067","project\u002Fcritical\u002Fperformance","pornel",[342,48,39,243,60,45],"23657328",[344,18,18,194,345,45],"23654134","0-_-0",[347,18,39,263,60,45],"23654220",[349,18,45,210,345,45],"23654414",[92,18,45,204,11,33],[352,18,18,353,354,45],"23657781","project\u002Fpositive\u002Fnone","Darkphibre",[81,41,39,356,82,45],"project\u002Fmixed\u002Fperformance",[107,41,45,207,11,33],[359,41,33,339,360,45],"23655944","pbowyer",[362,41,45,210,82,45],"23658491",[364,41,45,365,360,45],"23660394","commenter\u002Fcritical\u002Fperformance",[367,41,18,204,368,45],"23658648","CamperBob2",[370,41,33,339,360,45],"23660391",[372,189,18,373,296,45],"23653383","maker\u002Fpositive\u002Fdesign",[163,375,33,339,164,45],8,[377,375,45,207,11,33],"23654572",[379,380,18,194,381,45],"23655537",9,"roywiggins",[383,38,39,197,384,45],"23658222","mchusma",[386,387,33,339,316,45],"23656933",11,[389,387,39,207,60,45],"23658068",[136,391,18,194,137,45],12,[393,391,45,207,11,33],"23656597",[395,391,39,197,396,45],"23657117","mikepurvis",[155,398,18,399,156,45],13,"maker\u002Fpositive\u002Fexecution",[401,398,45,249,11,33],"23660299",[403,398,18,404,156,45],"23660759","maker\u002Fpositive\u002Fnone",[406,407,33,339,408,45],"23659210",14,"microcolonel",[410,407,45,207,11,33],"23660317",[412,413,39,207,414,45],"23656156",15,"fizixer",[416,413,45,207,11,33],"23656649",[140,34,39,197,141,45],[419,34,45,249,11,33],"23656026",[],[422,427,432,437,442,447],{"id":423,"name":424,"points":425,"category":426},"23658721","unim.press",638,"productivity",{"id":428,"name":429,"points":430,"category":431},"23669650","Paper HN",627,"consumer_social",{"id":433,"name":434,"points":435,"category":436},"23661808","Rainbow",389,"hardware",{"id":438,"name":439,"points":440,"category":441},"23643096","Aviary.sh",276,"infrastructure",{"id":443,"name":444,"points":445,"category":446},"23700486","Coding Career Handbook",245,"learning_science",{"id":448,"name":449,"points":450,"category":426},"23668507","Promnesia",234,false,"v2",{"id":454,"name":5,"line":6,"ai":451,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":455,"top":18,"kw":456,"maker":11,"domain":10,"tier":19,"attempts":45,"strip":457,"url":9,"mix":458,"praise":77,"objection":96,"sub":21},23652753,0.03,[24,25,26,27],"5533131545313533313310153140510415155333",[31,32,33,34],1794036592762]