[{"data":1,"prerenderedAt":476},["ShallowReactive",2],{"project-45536694":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":45,"post_text":46,"what":47,"said":63,"maker_in_thread":150,"shape":162,"grid":166,"attempts":438,"same_week":439,"maker_shows":33,"maker_hidden":468,"reader_version":469,"card":470},"45536694","Discrete Distribution Networks","I invented a new generative model and got accepted to ICLR","title","Show HN: I invented a new generative model and got accepted to ICLR","https:\u002F\u002Fdiscrete-distribution-networks.github.io\u002F","discrete-distribution-networks.github.io","diyer22",1760086914,2025,656,91,44,3837,2,"legend","ai","generative models","other",[24,25,26,27],"generative model","discrete distribution","zero-shot conditional generation","tree latent","curious",0.221,{"positive":31,"neutral":32,"mixed":33,"critical":34},16,24,1,7,48,{"idea":37,"performance":39,"design":42,"execution":43,"docs":44},{"positive":38,"critical":33},5,{"positive":40,"critical":41},0,3,{"positive":41,"critical":40},{"positive":33,"critical":33},{"positive":40,"critical":18},[41,38,33,41,33,41,41,40,38,41,38,41,41,41,38,40,38,41,38,40,41,38,41,41,40,41,38,40,41,38,33,38,38,41,33,38,41,40,41,33],"I invented Discrete Distribution Networks, a novel generative model with simple principles and unique properties, and the paper has been accepted to ICLR2025!\n\nModeling data distribution is challenging; DDN adopts a simple yet fundamentally different approach compared to mainstream generative models (Diffusion, GAN, VAE, autoregressive model):\n\n1. The model generates multiple outputs simultaneously in a single forward pass, rather than just one output. 2. It uses these multiple outputs to approximate the target distribution of the training data. 3. These outputs together represent a discrete d",{"problem":48,"audience":54,"pricing":54,"motivation":55,"time_to_build":54,"stack":59,"openness":60,"stage":61,"team_size":62},{"value":49,"source":50,"evidence":51},"Modeling data distributions with a single-pass alternative to diffusion, GAN, VAE and autoregressive models","maker",[52],{"id":53},"post",null,{"value":56,"source":50,"evidence":57},"Invented a new generative model accepted to ICLR 2025",[58],{"id":53},[],"open_source","launched","solo",{"praise":64,"objections":81,"requests":99,"questions":115,"alternatives":131},[65],{"value":66,"source":67,"evidence":68},"Novel approach, a genuinely new generative method","commenter",[69,73,77],{"id":70,"by":71,"text":72},"45538635","moconnor","Super cool, I spent a lot of time playing with representation learning back in the day and the grids of MNIST digits took me right back :) A genuinely interesting and novel…",{"id":74,"by":75,"text":76},"45541453","nothrowaways","Impressive, congrats.",{"id":78,"by":79,"text":80},"45542314","elchananHaas","First, I think this is really cool. Its great to see novel generative architectures. Here are my thoughts on the statistics behind this. First, let D be the data sample. Start…",[82,88],{"value":83,"source":50,"evidence":84},"Results are preliminary and far from state of the art",[85],{"id":86,"by":11,"text":87},"45537799","The first version of DDN was developed in less than three months, almost entirely by one person. Consequently, the experiments were preliminary and the results far from SoTA. The…",{"value":89,"source":67,"evidence":90},"Relation to normalizing flows and VQ-VAE questioned",[91,95],{"id":92,"by":93,"text":94},"45546648","cooljoseph","This sounds somewhat like a normalizing flow from a discrete space to a continuous space. I think there's a way you can rewrite your DDN layer as a normalizing flow which avoids…",{"id":96,"by":97,"text":98},"45537539","serf","isn't this kind of like an 80% vq-vae?",[100,108],{"value":101,"source":67,"evidence":102},"Scaling to larger models and non-image domains",[103,104],{"id":70,"by":71,"text":72},{"id":105,"by":106,"text":107},"45537538","p1esk","How does it compare to state of the art models? Does it scale?",{"value":109,"source":67,"evidence":110},"Reproduction details such as hardware and training curves",[111],{"id":112,"by":113,"text":114},"45540024","highd","Do you have any details on the experiment procedures? E.g. hardware, training time, loss curves? It is difficult to confidently reproduce research without at least some of these…",[116,120],{"value":117,"source":67,"evidence":118},"How does it compare to state of the art models",[119],{"id":105,"by":106,"text":107},{"value":121,"source":67,"evidence":122},"Can it be used for object detection or text-to-speech",[123,127],{"id":124,"by":125,"text":126},"45538286","FitchApps","Can you train this model to detect objects (e.g detect a fish in the picture)?",{"id":128,"by":129,"text":130},"45541622","cellis","Could this be used to train a text -> audio model? I'm thinking of an architecture that uses RVQ. Would RVQ still be necessary?",[132,139,146],{"value":133,"source":67,"evidence":134},"Diffusion models",[135,136],{"id":96,"by":97,"text":98},{"id":137,"by":11,"text":138},"45538449","I believe DDN is exceptionally well-suited to the “generative models for discriminative tasks” paradigm for object detection. Much like DiffusionDet, which applies diffusion…",{"value":140,"source":67,"evidence":141},"VQ-VAE",[142,143],{"id":96,"by":97,"text":98},{"id":144,"by":11,"text":145},"45538108","No, DDN and VQ-VAE are clearly different. Similarities: - Both map data to a discrete latent space. Differences: - VQ-VAE needs an external prior over code indices (e.g. PixelCNN…",{"value":147,"source":67,"evidence":148},"Normalizing flows",[149],{"id":92,"by":93,"text":94},{"comments":32,"answered":151,"key":152},18,[153,156,159],{"id":154,"text":155},"45552406","Thanks for the idea, but DDN and flow can’t be flipped into each other that easily. 1. DDN doesn’t need to be invertible. 2. Its latent is discrete, not continuous. 3. As far as…",{"id":157,"text":158},"45541861","I agree with @ActivePattern and thank you for your help in answering. Supplement for @f_devd: During training, the K outputs share the stem feature from the NN blocks, so…",{"id":160,"text":161},"45541636","In DDN, 1×1 convolutions are used only in the output layers of the Discrete Distribution Layer (DDL). The NN blocks between DDLs, which supply the fundamental computational power…",{"commenters":163,"max_depth":38,"maker_replies":32,"first_comment_mins":164,"lifespan_hours":165},60,92,152,[167,169,171,174,178,182,186,190,193,196,198,200,204,209,210,213,215,218,222,227,230,234,236,240,243,246,248,249,252,256,258,262,264,269,272,276,279,282,286,289,291,294,296,299,302,304,306,307,308,311,313,317,319,321,323,325,329,331,332,335,337,341,344,347,351,354,356,358,361,366,368,372,376,378,382,386,389,392,394,397,401,404,407,411,415,418,421,425,430,433,436],[92,40,41,168,93,40],"project\u002Fneutral\u002Fidea",[154,40,40,170,11,33],"commenter\u002Fneutral\u002Fnone",[172,40,41,168,173,40],"45602050","godelski",[175,33,38,176,177,40],"45539172","maker\u002Fpositive\u002Fnone","michaeldoron",[179,18,33,180,181,40],"45537537","project\u002Fcritical\u002Fperformance","f_devd",[183,18,41,184,185,40],"45537638","project\u002Fneutral\u002Fperformance","yorwba",[187,18,38,188,189,40],"45539642","maker\u002Fpositive\u002Fidea","kevmo314",[191,18,33,180,192,40],"45541272","crondee",[194,18,40,170,195,40],"45540788","ActivePattern",[157,18,40,197,11,33],"commenter\u002Fneutral\u002Fperformance",[199,18,33,180,181,40],"45549800",[201,41,40,202,203,40],"45546417","other\u002Fneutral\u002Fnone","mysterEFrank",[205,206,41,207,208,40],"45541532",4,"project\u002Fneutral\u002Fdesign","qazxcvbnm",[160,206,40,207,11,33],[211,206,41,207,212,40],"45542985","randomNumber7",[214,206,40,207,11,33],"45546663",[216,206,40,202,217,40],"45543079","gwern",[219,38,40,220,221,40],"45540558","commenter\u002Fneutral\u002Fidea","intalentive",[223,224,38,225,226,40],"45538666",6,"project\u002Fpositive\u002Fdesign","GaggiX",[228,224,38,225,229,40],"45538991","BrokenCogs",[231,34,41,232,233,40],"45544754","project\u002Fneutral\u002Fexecution","hatthew",[235,34,40,232,11,33],"45547003",[237,238,40,202,239,40],"45539611",8,"CuriouslyC",[70,241,38,242,71,40],9,"project\u002Fpositive\u002Fidea",[244,241,40,245,11,33],"45538880","maker\u002Fneutral\u002Fnone",[124,247,41,168,125,40],10,[137,247,40,168,11,33],[250,247,41,168,251,40],"45541017","porridgeraisin",[253,254,41,168,255,40],"45540832",11,"0xdeadbeefbabe",[257,254,40,168,11,33],"45541994",[259,260,41,207,261,40],"45549011",12,"frumiousirc",[263,260,40,207,11,33],"45550807",[265,266,38,267,268,40],"45539179",13,"project\u002Fpositive\u002Fnone","Lerc",[270,271,40,245,11,33],"45607601",14,[273,274,40,202,275,40],"45543745",15,"aseg",[277,274,40,202,278,40],"45544728","3abiton",[280,31,38,176,281,40],"45540940","cs702",[283,284,41,168,285,40],"45537268",17,"VoidWhisperer",[287,284,41,168,288,40],"45537309","booli",[290,284,40,168,11,33],"45537623",[292,284,38,242,293,40],"45539561","vintermann",[295,284,40,242,11,33],"45539656",[297,284,41,168,298,40],"45543052","cubefox",[300,284,40,202,301,40],"45548807","lukan",[303,284,40,202,298,40],"45549890",[305,284,40,170,11,33],"45550877",[105,151,41,184,106,40],[86,151,40,184,11,33],[309,151,38,242,310,40],"45544198","aDyslecticCrow",[312,151,40,242,11,33],"45552460",[314,315,41,184,316,40],"45542648",19,"throwaway314155",[318,315,40,184,11,33],"45546962",[320,315,40,170,316,40],"45552870",[128,322,41,168,129,40],20,[324,322,40,168,11,33],"45557187",[326,327,40,202,328,40],"45543135",21,"mellosouls",[96,330,41,168,97,40],22,[144,330,40,168,11,33],[112,333,41,334,113,40],23,"project\u002Fneutral\u002Fdocs",[336,333,40,334,11,33],"45540320",[338,333,38,339,340,40],"45541650","maker\u002Fpositive\u002Fexecution","Zacharias030",[342,32,40,220,343,40],"45554663","ProjectArcturis",[345,32,40,346,11,33],"45555645","maker\u002Fneutral\u002Fidea",[348,349,41,207,350,40],"45546511",25,"cttet",[352,349,40,353,11,33],"45546559","maker\u002Fneutral\u002Fdesign",[355,349,41,353,350,40],"45546963",[357,349,40,353,11,33],"45547068",[359,349,38,360,350,40],"45547078","maker\u002Fpositive\u002Fdesign",[362,363,33,364,365,40],"45538898",26,"project\u002Fcritical\u002Fexecution","gurtinator",[367,363,40,232,11,33],"45539133",[369,370,38,176,371,40],"45540498",27,"kaiokendev",[373,374,41,168,375,40],"45543529",28,"wordglyph",[74,377,38,176,75,40],29,[379,380,41,168,381,40],"45550122",30,"v9v",[383,384,40,202,385,40],"45538403",31,"Der_Einzige",[387,384,40,202,388,40],"45538434","albertzeyer",[78,390,41,391,79,40],32,"project\u002Fmixed\u002Fidea",[393,390,41,207,79,40],"45543166",[395,390,33,396,79,40],"45543472","project\u002Fcritical\u002Fidea",[398,399,38,176,400,40],"45538324",33,"nvr219",[402,399,38,242,403,40],"45539020","curtistyr",[405,399,40,406,11,33],"45539594","maker\u002Fneutral\u002Fdocs",[408,409,41,406,410,40],"45543191",34,"Invictus0",[412,409,40,413,414,40],"45543226","commenter\u002Fneutral\u002Fdocs","Uehreka",[416,409,40,413,417,40],"45546939","mindcrime",[419,409,40,413,420,40],"45547293","antegamisou",[422,423,38,176,424,40],"45547292",35,"jama211",[426,427,33,428,429,40],"45544834",36,"project\u002Fcritical\u002Fdocs","E-Reverance",[431,427,40,413,432,40],"45545333","idiotsecant",[434,427,40,413,435,40],"45546408","kcexn",[437,427,33,428,429,40],"45547100",[],[440,445,450,455,459,464],{"id":441,"name":442,"points":443,"category":444},"45567770","SQLite Online",465,"data",{"id":446,"name":447,"points":448,"category":449},"45526890","Lavandula",420,"dev_tools",{"id":451,"name":452,"points":453,"category":454},"45529393","keyer",411,"hardware",{"id":456,"name":457,"points":458,"category":449},"45590949","Halloy",381,{"id":460,"name":461,"points":462,"category":463},"45558611","Ambi",314,"consumer_social",{"id":465,"name":466,"points":467,"category":449},"45553995","Rift",212,false,"v2",{"id":471,"name":5,"line":6,"ai":468,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":472,"top":18,"kw":473,"maker":11,"domain":10,"tier":19,"attempts":40,"strip":474,"url":9,"mix":475,"praise":66,"objection":83,"sub":21},45536694,0.22,[24,25,26,27],"3513133053533350535035330350351553153031",[31,32,33,34],1794036683952]