{"id":3947,"date":"2024-06-21T18:22:20","date_gmt":"2024-06-21T10:22:20","guid":{"rendered":"https:\/\/nullthought.net\/?p=3947"},"modified":"2025-02-23T11:09:47","modified_gmt":"2025-02-23T03:09:47","slug":"cvpr2024%e6%9c%80%e4%bd%b3%e8%ae%ba%e6%96%87%ef%bc%9arich-human-feedback-for-text-to-image-generation","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=3947","title":{"rendered":"CVPR2024\u6700\u4f73\u8bba\u6587\uff1aRich Human Feedback for Text-to-Image Generation"},"content":{"rendered":"\n

\u8ba1\u7b97\u673a\u56fe\u50cf\u548c\u6a21\u5f0f\u8bc6\u522b2024\u5e74\u5ea6\u5927\u4f1a\uff08Computer Vision and Pattern Recognition 2024 conference, CVPR2024\uff09\u4e0a\uff0c\u8bba\u6587\u201cRich Human Feedback for Text-to-Image Generation<\/a>\u201d\u83b7\u8bc4\u6700\u4f73\u8bba\u6587\u3002\u8bba\u6587\u4f5c\u8005\u5305\u62ec Youwei Liang\uff08\u52a0\u5229\u798f\u5c3c\u4e9a\u5927\u5b66\u5723\u5730\u4e9a\u54e5\u5206\u6821\uff09\u3001Junfeng He\uff08Google \u7814\u7a76\u9662\uff09\u3001Gang Li\uff08Google \u7814\u7a76\u9662\uff09\u3001Peizhao Li\uff08\u5e03\u5170\u4ee3\u65af\u5927\u5b66\uff09\u3001Arseniy Klimovskiy\u3001Nicholas Carolan\u3001Jiao Sun\u3001Jordi Pont-Tuset\u3001Sarah Young\u3001Feng Yang\u3001Junjie Ke\u3001Krishnamurthy Dj Dvijotham\u3001Katherine M. Collins\uff08\u5251\u6865\u5927\u5b66\uff09\u3001Yiwen Luo\u3001Yang Li\u3001Kai J Kohlhoff\u3001Deepak Ramachandran \u548c Vidhya Navalpakkam\uff08\u4ee5\u4e0a\u5747\u96b6\u5c5e\u4e8e Google \u7814\u7a76\u9662\uff09\u3002\ud83d\udc4d\ud83d\udc4d<\/p>\n\n\n\n

\"\"
Architecture of Rich Human Feedback Model
Rich Human Feedback for Text-to-Image Generation<\/a><\/figcaption><\/figure>\n\n\n\n

\u8bba\u6587\u5185\u5bb9\u6982\u8ff0\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\u8bba\u6587\u6458\u8981\uff1a\u5c3d\u7ba1\u6700\u8fd1\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\uff08T2I\uff09\u6a21\u578b\u5982 Stable Diffusion \u548c Imagen \u5728\u6839\u636e\u6587\u672c\u63cf\u8ff0\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65b9\u9762\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\uff0c\u4f46\u751f\u6210\u7684\u56fe\u50cf\u4ecd\u5b58\u5728\u4f2a\u5f71\/\u4e0d\u5408\u7406\u6027\u3001\u4e0e\u6587\u672c\u63cf\u8ff0\u4e0d\u5bf9\u9f50\u548c\u7f8e\u5b66\u8d28\u91cf\u4f4e\u7b49\u95ee\u9898\u3002\u53d7\u5f3a\u5316\u5b66\u4e60\u4e0e\u4eba\u7c7b\u53cd\u9988\uff08RLHF\uff09\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u6210\u529f\u542f\u53d1\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u901a\u8fc7\u6536\u96c6\u4e30\u5bcc\u7684\u4eba\u7c7b\u53cd\u9988\uff08RHF\uff09\u6765\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u3002RHF \u5305\u62ec\u5bf9\u751f\u6210\u56fe\u50cf\u4e2d\u4e0d\u5408\u7406\u6216\u4e0d\u5bf9\u9f50\u7684\u533a\u57df\u8fdb\u884c\u8be6\u7ec6\u6ce8\u91ca\uff0c\u5e76\u6536\u96c6\u4e86 18,000 \u5f20\u56fe\u50cf\uff08RichHF-18K\uff09\u7684\u6b64\u7c7b\u53cd\u9988\u6570\u636e\u3002\u8bad\u7ec3\u4e86\u4e00\u4e2a\u591a\u6a21\u6001 Transformer \u6a21\u578b\u6765\u81ea\u52a8\u9884\u6d4b\u8fd9\u4e9b\u4e30\u5bcc\u7684\u53cd\u9988\uff0c\u53ef\u4ee5\u7528\u4e8e\u6539\u8fdb T2I \u751f\u6210\u6a21\u578b\u3002\u8fd9\u4e9b\u6539\u8fdb\u53ef\u4ee5\u63a8\u5e7f\u5230\u4e0d\u540c\u7684\u6a21\u578b\uff0c\u8868\u660e\u53cd\u9988\u673a\u5236\u5177\u6709\u9c81\u68d2\u6027\u3002<\/p>\n\n\n\n

\u4e00\u3001\u5f15\u8a00<\/h4>\n\n\n\n

\u6587\u672c\u751f\u6210\u56fe\u50cf\u6a21\u578b\u5728\u827a\u672f\u3001\u8bbe\u8ba1\u548c\u5e7f\u544a\u7b49\u521b\u610f\u9886\u57df\u53d8\u5f97\u8d8a\u6765\u8d8a\u91cd\u8981\u3002\u5c3d\u7ba1\u53d6\u5f97\u4e86\u8fdb\u5c55\uff0c\u751f\u6210\u7684\u56fe\u50cf\u4ecd\u5b58\u5728\u4f2a\u5f71\u3001\u6587\u672c\u4e0d\u5bf9\u9f50\u548c\u7f8e\u5b66\u8d28\u91cf\u4f4e\u7684\u95ee\u9898\u3002\u73b0\u6709\u7684\u81ea\u52a8\u8bc4\u4f30\u6307\u6807\u5f80\u5f80\u65e0\u6cd5\u6355\u6349\u8fd9\u4e9b\u7ec6\u5fae\u5dee\u522b\u3002\u672c\u6587\u5f15\u5165\u4e86\u4e00\u4e2a\u6570\u636e\u96c6\uff08RichHF-18K\uff09\u548c\u4e00\u4e2a\u591a\u6a21\u6001 Transformer \u6a21\u578b\uff0c\u63d0\u4f9b\u6bd4\u5355\u4e00\u8bc4\u5206\u6307\u6807\uff08\u5982 IS\u3001FID \u548c CLIP-Score\uff09\u66f4\u7ec6\u81f4\u548c\u53ef\u89e3\u91ca\u7684\u8bc4\u4f30\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u76f8\u5173\u5de5\u4f5c<\/h4>\n\n\n\n

\u672c\u6587\u5c06\u5176\u8d21\u732e\u653e\u5728\u6587\u672c\u751f\u6210\u56fe\u50cf\u548c\u8bc4\u4f30\u7684\u5e7f\u6cdb\u80cc\u666f\u4e0b\uff0c\u5f3a\u8c03\u4e86\u4ece GANs \u548c VAEs \u5230\u6269\u6563\u6a21\u578b\u7684\u53d1\u5c55\u3002\u6279\u8bc4\u4e86\u73b0\u6709\u8bc4\u4f30\u65b9\u6cd5\u7684\u4e8c\u5143\u6027\u8d28\u548c\u7f3a\u4e4f\u8be6\u7ec6\u53cd\u9988\uff0c\u63d0\u51fa\u4e86\u66f4\u4e30\u5bcc\u7684\u53cd\u9988\u673a\u5236\u4f5c\u4e3a\u663e\u8457\u6539\u8fdb\u3002<\/p>\n\n\n\n

\u6570\u636e\u6536\u96c6<\/h4>\n\n\n\n

RichHF-18K \u6570\u636e\u96c6\u5305\u62ec\uff1a<\/p>\n\n\n\n

    \n
  1. \u70b9\u6ce8\u91ca<\/strong>\uff1a\u6807\u8bb0\u56fe\u50cf\u4e2d\u5b58\u5728\u4f2a\u5f71\u6216\u4e0d\u5bf9\u9f50\u7684\u533a\u57df\u3002<\/li>\n\n\n\n
  2. \u5173\u952e\u8bcd\u6ce8\u91ca<\/strong>\uff1a\u8bc6\u522b\u6587\u672c\u63d0\u793a\u4e2d\u88ab\u8bef\u89e3\u6216\u7f3a\u5931\u7684\u8bcd\u8bed\u3002<\/li>\n\n\n\n
  3. \u8bc4\u5206<\/strong>\uff1a\u56db\u79cd\u7ec6\u5316\u8bc4\u5206\uff0c\u5305\u62ec\u5408\u7406\u6027\u3001\u56fe\u6587\u5bf9\u9f50\u3001\u7f8e\u5b66\u548c\u6574\u4f53\u8bc4\u5206\u3002<\/li>\n<\/ol>\n\n\n\n

    \u6ce8\u91ca\u8fc7\u7a0b\u6d89\u53ca\u6bcf\u4e2a\u56fe\u50cf-\u6587\u672c\u5bf9\u7531\u4e09\u540d\u6ce8\u91ca\u8005\u6807\u6ce8\uff0c\u4ee5\u786e\u4fdd\u53ef\u9760\u6027\uff0c\u5e76\u8bbe\u8ba1\u4e86\u8be6\u7ec6\u7684\u7f51\u7edc\u7528\u6237\u754c\u9762\uff08UI\uff09\u4ee5\u63d0\u9ad8\u6570\u636e\u6536\u96c6\u6548\u7387\u548c\u5168\u9762\u6027\u3002<\/p>\n\n\n\n

    \u4e09\u3001\u6a21\u578b\u67b6\u6784<\/h4>\n\n\n\n

    \u63d0\u51fa\u7684\u6a21\u578b\u4f7f\u7528\u4e86\u57fa\u4e8e ViT \u548c T5X \u6a21\u578b\u7684\u89c6\u89c9-\u8bed\u8a00\u67b6\u6784\uff0c\u5b9e\u73b0\u4e86\u53cc\u5411\u4fe1\u606f\u4f20\u64ad\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u9884\u6d4b\u4e0d\u5408\u7406\u548c\u4e0d\u5bf9\u9f50\u7684\u533a\u57df\u3001\u9519\u914d\u7684\u5173\u952e\u8bcd\u4ee5\u53ca\u7ec6\u5316\u8bc4\u5206\uff0c\u4ece\u800c\u63d0\u4f9b\u66f4\u8be6\u7ec6\u548c\u53ef\u89e3\u91ca\u7684\u56fe\u50cf\u8d28\u91cf\u8bc4\u4f30\u3002<\/p>\n\n\n\n

    \u56db\u3001\u5b9e\u9a8c\u4e0e\u7ed3\u679c<\/h4>\n\n\n\n

    \u6a21\u578b\u5728 16K \u4e2a\u8bad\u7ec3\u6837\u672c\u4e0a\u8bad\u7ec3\uff0c\u5e76\u5728 RichHF-18K \u6570\u636e\u96c6\u7684 1K \u9a8c\u8bc1\u548c 1K \u6d4b\u8bd5\u6837\u672c\u4e0a\u8fdb\u884c\u8bc4\u4f30\u3002\u8bc4\u4f30\u6307\u6807\u5305\u62ec\u76ae\u5c14\u900a\u7ebf\u6027\u76f8\u5173\u7cfb\u6570\uff08PLCC\uff09\u548c\u65af\u76ae\u5c14\u66fc\u79e9\u76f8\u5173\u7cfb\u6570\uff08SRCC\uff09\u7528\u4e8e\u8bc4\u5206\u9884\u6d4b\uff0c\u4ee5\u53ca\u5404\u79cd\u663e\u8457\u6027\u6307\u6807\u7528\u4e8e\u70ed\u56fe\u9884\u6d4b\u3002\u7ed3\u679c\u663e\u793a\uff0c\u4e0e ResNet-50 \u548c CLIP \u7b49\u57fa\u7ebf\u76f8\u6bd4\uff0c\u6a21\u578b\u5728\u9884\u6d4b\u8bc4\u5206\u548c\u70ed\u56fe\u65b9\u9762\u663e\u8457\u6539\u8fdb\u3002<\/p>\n\n\n\n

    \u4e94\u3001\u5e94\u7528<\/h4>\n\n\n\n
      \n
    1. \u6539\u8fdb\u56fe\u50cf\u751f\u6210<\/strong>\uff1a\u4f7f\u7528\u9884\u6d4b\u7684\u70ed\u56fe\u5bf9\u95ee\u9898\u533a\u57df\u8fdb\u884c\u4fee\u8865\uff0c\u5e76\u4f7f\u7528\u8bc4\u5206\u5fae\u8c03\u6a21\u578b\uff08\u5982 Muse\uff09\u3002<\/li>\n\n\n\n
    2. \u8de8\u6a21\u578b\u6cdb\u5316<\/strong>\uff1a\u6539\u8fdb\u80fd\u5f88\u597d\u5730\u63a8\u5e7f\u5230\u4e0d\u540c\u7684\u6a21\u578b\uff0c\u8868\u660e\u53cd\u9988\u673a\u5236\u5177\u6709\u9c81\u68d2\u6027\u3002<\/li>\n<\/ol>\n\n\n\n

      \u516d\u3001\u7ed3\u8bba\u4e0e\u5c40\u9650\u6027<\/h4>\n\n\n\n

      \u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u4e2a\u65b0\u9896\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u4e30\u5bcc\u4eba\u7c7b\u53cd\u9988\u6570\u636e\u96c6\u548c\u6a21\u578b\uff0c\u5c55\u793a\u4e86\u5728\u56fe\u50cf\u8d28\u91cf\u65b9\u9762\u7684\u663e\u8457\u6539\u8fdb\u3002\u5c40\u9650\u6027\u5305\u62ec\u5bf9\u9f50\u6ce8\u91ca\u7684\u6311\u6218\u548c\u6570\u636e\u96c6\u4e3b\u4f53\u7684\u591a\u6837\u6027\u3002\u672a\u6765\u5de5\u4f5c\u5c06\u63a2\u7d22\u66f4\u591a\u6837\u5316\u7684\u6570\u636e\u96c6\u548c\u4e30\u5bcc\u4eba\u7c7b\u53cd\u9988\u7684\u989d\u5916\u5e94\u7528\u3002<\/p>\n\n\n\n

      \u4e03\u3001\u4e3b\u8981\u8d21\u732e<\/h4>\n\n\n\n
        \n
      1. RichHF-18K \u6570\u636e\u96c6<\/strong>\uff1a\u4e00\u4e2a\u5305\u542b 18K \u5f20\u56fe\u50cf\u7684\u8be6\u7ec6\u4eba\u7c7b\u53cd\u9988\u7efc\u5408\u6570\u636e\u96c6\u3002<\/li>\n\n\n\n
      2. \u591a\u6a21\u6001 Transformer \u6a21\u578b<\/strong>\uff1a\u4e00\u4e2a\u9884\u6d4b\u4e30\u5bcc\u4eba\u7c7b\u53cd\u9988\u7684\u6a21\u578b\uff0c\u63d0\u9ad8\u56fe\u50cf\u751f\u6210\u8d28\u91cf\u3002<\/li>\n\n\n\n
      3. \u5b9e\u9645\u5e94\u7528<\/strong>\uff1a\u5c55\u793a\u4e86\u5229\u7528\u4e30\u5bcc\u53cd\u9988\u589e\u5f3a\u751f\u6210\u6a21\u578b\u7684\u5b9e\u7528\u6027\u3002<\/li>\n<\/ol>\n\n\n\n

        \u516b\u3001\u4f26\u7406\u8003\u8651<\/h4>\n\n\n\n

        \u6570\u636e\u6536\u96c6\u8fc7\u7a0b\u5df2\u83b7\u5f97\u673a\u6784\u5ba1\u67e5\u59d4\u5458\u4f1a\uff08IRB\uff09\u7684\u6279\u51c6\uff0c\u786e\u4fdd\u5728\u6ce8\u91ca\u548c\u4f7f\u7528\u6570\u636e\u96c6\u65b9\u9762\u7684\u4f26\u7406\u6807\u51c6\u3002<\/p>\n\n\n\n

        \u8fd9\u4e00\u8be6\u7ec6\u5206\u6790\u7a81\u51fa\u4e86\u5229\u7528\u4e30\u5bcc\u4eba\u7c7b\u53cd\u9988\u6765\u89e3\u51b3\u6587\u672c\u751f\u6210\u56fe\u50cf\u4e2d\u6301\u4e45\u95ee\u9898\u7684\u521b\u65b0\u65b9\u6cd5\uff0c\u6807\u5fd7\u7740\u8be5\u9886\u57df\u7684\u91cd\u8981\u8fdb\u6b65\u3002<\/p>\n\n\n\n

        \n\n\n\n

        P.S., RichHF-18K data set on GitHub<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

        \u8ba1\u7b97\u673a\u56fe\u50cf\u548c\u6a21\u5f0f\u8bc6\u522b2024\u5e74\u5ea6\u5927\u4f1a\uff08Computer Vision and Pattern Recognit […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8,36],"tags":[39,63,95,90,80,66],"class_list":{"0":"post-3947","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-tech","7":"category-36","8":"tag-ai","10":"tag-transformer","11":"tag-multimodal","12":"tag-cv","13":"tag-google"},"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"Tech<\/a> \u79d1\u5b66<\/a>","rttpg_excerpt":"\u8ba1\u7b97\u673a\u56fe\u50cf\u548c\u6a21\u5f0f\u8bc6\u522b2024\u5e74\u5ea6\u5927\u4f1a\uff08Computer Vision and Pattern Recognit…","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/3947","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3947"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/3947\/revisions"}],"predecessor-version":[{"id":3949,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/3947\/revisions\/3949"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3947"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3947"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3947"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}