原创力文档- 1、本文档共19页，可阅读全部内容。
- 2、原创力文档（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
- 3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。
ImprovingImageGenerationwithBetterCaptions
JamesBetker??
jbetker@
GabrielGoh??
ggoh@
LiJing??
lijing@
TimBrooks?
JianfengWang?LinjieLi?LongOuyang?JuntangZhuang?JoyceLee?YufeiGuo?WesamManassra? PrafullaDhariwal? CaseyChu? YunxinJiao?
AdityaRamesh??
aramesh@
Abstract
Weshowthatpromptfollowingabilitiesoftext-to-imagemodelscanbesub-stantiallyimprovedbytrainingonhighlydescriptivegeneratedimagecaptions.Existingtext-to-imagemodelsstruggletofollowdetailedimagedescriptionsandoftenignorewordsorconfusethemeaningofprompts.Wehypothesizethatthisissuestemsfromnoisyandinaccurateimagecaptionsinthetrainingdataset.Weaddressthisbytrainingabespokeimagecaptioneranduseittorecaptionthetrainingdataset.Wethentrainseveraltext-to-imagemodelsandfindthattrainingonthesesyntheticcaptionsreliablyimprovespromptfollowingability.Finally,weusethesefindingstobuildDALL-E3:anewtext-to-imagegenerationsystem,andbenchmarkitsperformanceonanevaluationdesignedtomeasurepromptfollowing,coherence,andaesthetics,findingthatitcomparesfavorablytocompetitors.Wepublishsamplesandcodefortheseevaluationssothatfutureresearchcancontinueoptimizingthisimportantaspectoftext-to-imagesystems.
Introduction
Recentadvancesingenerativemodelinghaveallowedtext-to-imagegenerativemodelstoachievedrasticperformanceimprovements.Inparticular,tacklingtheproblemwithsampling-basedap-proachessuchasautoregressivegenerativemodeling[27,2,1,20,30]orusingdiffusionprocesses[25,6,11,12,19,22]haveallowedustodecomposetheproblemofimagegenerationintosmall,discretestepswhicharemoretractableforneuralnetworkstolearn.
Inparallel,researchershavefoundwaystobuildimagegeneratorsoutofstacksofself-attentionlayers[15,3,4].Decouplingimagegenerationfromtheimplicitspatial
文档评论（0）