ConsistencyModelsYangSong1PrafullaDhariwal1MarkChen1IlyaSutskever1AbstractDiffusionmodelshavemadesignificantbreak-throughsinimage,audio,andvideogeneration,buttheydependonaniterativegenerationprocessthatcausesslowsamplingspeedandcapstheirpotentialforreal-timeapplications.Toovercomethislimitation,weproposeconsistencymodels,anewfamilyofgenerativemodelsthatachievehighsamplequalitywithoutadversarialtraining.Theysupportfastone-stepgenerationbydesign,whilestillallowingforfew-stepsamplingtotradecomputeforsamplequality.Theyalsosupportzero-shotdataediting,likeimageinpainting,col-orization,andsuper-resolution,withoutrequir-ingexplicittrainingonthesetasks.Consistencymodelscanbetrainedeitherasawaytodistillpre-traineddiffusionmodels,orasstandalonegen-erativemodels.Throughextensiveexperiments,wedemonstratethattheyoutperformexistingdis-tillationtechniquesfordiffusionmodelsinone-andfew-stepgeneration.Forexample,weachievethenewstate-of-the-artFIDof3.55onCIFAR-10and6.20onImageNet64ˆ64forone-stepgeneration.Whentrainedasstandalonegenera-tivemodels,consistencymodelsalsooutperformsingle-step,non-adversarialgenerativemodelsonstandardbenchmarkslikeCIFAR-10,ImageNet64ˆ64andLSUN256ˆ256.1.IntroductionDiffusionmodels(Sohl-Dicksteinetal.,2015;Song&Er-mon,2019;2020;Hoetal.,2020;Songetal.,2021),alsoknownasscore-basedgenerativemodels,haveachievedunprecedentedsuccessacrossmultiplefields,includingim-agegeneration(Dhariwal&Nichol,2021;Nicholetal.,2021;Rameshetal.,2022;Sahariaetal.,2022;Rombachetal.,2022),audiosynthesis(Kongetal.,2020;Chenetal.,2021;Popovetal.,2021),andvideogeneration(Hoetal.,2022b;a).UnlikeGenerativeAdversarialNetworks(GANs,1OpenAI,SanFrancisco,CA94110,USA.Correspondenceto:YangSong.Preprint.Workinprogress.Figure1:GivenaProbabilityFlow(PF)ODEthatsmoothlyconvertsdatatonoise,welearntomapanypoint(e.g.,xt,xt1,andxT)ontheODEtrajectorytoitsorigin(e.g.,x0)forgenerativemodeling.Modelsofthesemappingsarecalledconsistencymodels,astheiroutputsaretrainedtobeconsistentforp...
