{"id":9558,"date":"2026-06-23T02:37:04","date_gmt":"2026-06-22T18:37:04","guid":{"rendered":"https:\/\/infernews.com\/blog\/stylisticbias\/"},"modified":"2026-06-23T02:38:29","modified_gmt":"2026-06-22T18:38:29","slug":"stylisticbias","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/stylisticbias\/","title":{"rendered":"StylisticBias \u62c6\u89e3 MLLMs \u8996\u89ba\u504f\u898b"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-d31f6234af02.jpg\" alt=\"StylisticBias pipeline overview\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u5c11 Multimodal Large Language Models\uff08MLLMs\uff09\u504f\u898b\u7814\u7a76\uff0c\u901a\u5e38\u62ff\u4e0d\u540c\u4eba\u7269\u6216\u7fa4\u7d44\u4e92\u76f8\u6bd4\u8f03\uff1b\u554f\u984c\u662f\u5916\u8c8c\u5dee\u7570\u8207\u8eab\u4efd\u5dee\u7570\u6703\u7e8f\u5728\u4e00\u8d77\uff0c\u6700\u5f8c\u5f88\u96e3\u5224\u65b7\u6a21\u578b\u7a76\u7adf\u662f\u53d7\u5e74\u9f61\u3001\u8863\u8457\u3001\u8eab\u5f62\u5f71\u97ff\uff0c\u9084\u662f\u53ea\u662f\u63db\u4e86\u53e6\u4e00\u500b\u4eba\u3002StylisticBias \u63d0\u51fa\u7684\u505a\u6cd5\u5f88\u660e\u78ba\uff1a\u5148\u751f\u6210 500 \u5f35 photorealistic base faces\uff0c\u518d\u70ba\u6bcf\u5f35\u81c9\u5efa\u7acb\u7d04 50 \u500b single-attribute variations\uff0c\u4ee4\u8cc7\u6599\u96c6\u7d2f\u7a4d\u5230\u7d04 25K images\uff0c\u7528\u300c\u56fa\u5b9a\u8eab\u4efd\u3001\u53ea\u6539\u4e00\u500b\u8996\u89ba\u5c6c\u6027\u300d\u7684\u65b9\u5f0f\u91cf\u5ea6 social bias\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u5c6c\u65bc\u4e00\u500b <strong>Dataset \u6578\u64da\u96c6 \/ benchmark<\/strong> \u9805\u76ee\uff0c\u5be6\u969b\u89e3\u6c7a\u7684\u662f\u300c\u600e\u6a23\u66f4\u7d30\u7dfb\u5730\u6e2c\u8a66 MLLMs \u6703\u56e0\u54ea\u4e9b\u5916\u89c0\u7dda\u7d22\u800c\u6539\u8b8a\u5c0d\u4eba\u7684\u793e\u6703\u5224\u65b7\u300d\u3002\u8cc7\u6599\u6d41\u7a0b\u4e5f\u5beb\u5f97\u6e05\u695a\uff1a<code>output\/images\/<\/code> \u653e base faces \u8207 metadata\uff0c<code>output\/banana\/<\/code> \u653e\u8b8a\u9ad4\uff0c<code>output\/judgements\/<\/code> \u6536\u96c6\u539f\u59cb\u6a21\u578b\u56de\u61c9\uff0c<code>output\/evaluation\/<\/code> \u5247\u6574\u7406\u7d71\u8a08\u3001\u8868\u683c\u8207\u5716\u8868\uff1b\u5373\u4f7f\u4e0d\u81ea\u884c\u91cd\u8dd1\u751f\u6210\u6d41\u7a0b\uff0c\u53ea\u770b\u9019\u5e7e\u5c64\u8f38\u51fa\uff0c\u4e5f\u8db3\u4ee5\u7406\u89e3\u6574\u500b\u8a55\u6e2c\u908f\u8f2f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u548c\u4e00\u822c fairness benchmark \u76f8\u6bd4\uff0c\u9019\u500b\u9805\u76ee\u6700\u503c\u5f97\u7559\u610f\u7684\u662f\u5b83\u4e0d\u662f\u53ea\u554f\u300c\u6a21\u578b\u6709\u6c92\u6709\u504f\u898b\u300d\uff0c\u800c\u662f\u8ffd\u5230\u300c\u54ea\u4e00\u985e\u8996\u89ba\u63d0\u793a\u6700\u6703\u63a8\u52d5\u504f\u898b\u300d\u3002\u4f5c\u8005\u8a55\u6e2c six MLLMs\u300125 \u500b binary social judgment scenarios\uff0c\u6307\u51fa age \u8207 body type \u4e3b\u5c0e identity-level effects\uff0c\u800c fashion style \u8207\u5176\u4ed6 visual cues \u5e36\u4f86\u6700\u5927\u7684 attribute-level shifts\uff1b\u53e6\u5916\u5927\u7d04 15 \u500b attributes \u5df2\u4f54\u8fd1 80% \u7e3d\u8b8a\u7570\uff0c\u4ee3\u8868\u504f\u898b\u4e26\u975e\u5e73\u5747\u6563\u843d\uff0c\u800c\u662f\u96c6\u4e2d\u5728\u5c11\u6578\u53ef\u8fa8\u8a8d\u7dda\u7d22\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u56fa\u5b9a\u540c\u4e00\u5f35\u81c9\uff0c\u53ea\u6539\u4e00\u500b\u5c6c\u6027\uff0c\u8f03\u6613\u5206\u958b appearance effects \u8207 identity differences<\/li>\n\n\n\n<li>\u898f\u6a21\u7d04 25K images\uff0c\u9069\u5408\u505a\u8f03\u7d30\u7c92\u5ea6\u7684 bias analysis<\/li>\n\n\n\n<li>\u7d50\u679c\u986f\u793a age\u3001body type\u3001fashion style \u662f\u9ad8\u654f\u611f\u56e0\u7d20<\/li>\n\n\n\n<li>judgement \u5c0d appearance \u8a9e\u610f\u8f03\u8cbc\u8fd1\u7684\u5834\u666f\u6700\u654f\u611f\uff0c\u5c24\u5176 socioeconomic \u8207 style-related \u5224\u65b7<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u9805\u76ee\u6700\u9069\u5408\u8a55\u4f30\u591a\u6a21\u614b\u7522\u54c1\u98a8\u96aa\u7684\u5718\u968a\u3001\u7814\u7a76 AI fairness \u7684\u5b78\u8005\uff0c\u4ee5\u53ca\u8981\u6bd4\u8f03\u4e0d\u540c vision-language model \u884c\u70ba\u7684\u4eba\u3002\u76f8\u95dc\u6a21\u578b\u8cc7\u8a0a\u5728\u73fe\u6709\u6750\u6599\u672a\u5b8c\u6574\u5217\u51fa\u516d\u500b\u540d\u7a31\uff0c\u4f46\u9805\u76ee\u660e\u78ba\u570d\u7e5e MLLMs\uff0c\u4e26\u5728\u751f\u6210\u968e\u6bb5\u63d0\u5230 Google Vertex AI Imagen 4\uff0c\u4ee5\u53ca variation builder \u4f7f\u7528 Nano Banana approach\uff1b\u82e5\u4f60\u95dc\u5fc3\u6a21\u578b\u90e8\u7f72\u524d\u7684\u504f\u898b\u6aa2\u67e5\uff0c\u9019\u500b benchmark \u6bd4\u55ae\u7d14\u770b\u6574\u9ad4\u6e96\u78ba\u7387\u66f4\u6709\u5206\u6790\u50f9\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/timo-cavelius\/StylisticBias\" rel=\"noopener noreferrer\">https:\/\/github.com\/timo-cavelius\/StylisticBias<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9805\u76ee\u4e3b\u9801\uff1a<\/strong> <a href=\"https:\/\/huggingface.co\/datasets\/shaghayegh\/stylistic-bias-dataset\" rel=\"noopener noreferrer\">https:\/\/huggingface.co\/datasets\/shaghayegh\/stylistic-bias-dataset<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.20527\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.20527<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9019\u4e0d\u662f\u4e00\u822c\u4eba\u81c9\u8cc7\u6599\u96c6\uff0c\u800c\u662f\u5c08\u9580\u91cf\u5ea6 MLLMs \u793e\u6703\u5224\u65b7\u504f\u5dee\u7684\u63a7\u5236\u578b benchmark\u3002\u5b83\u628a\u8eab\u4efd\u56fa\u5b9a\uff0c\u53ea\u6539\u4e00\u500b\u5916\u89c0\u7dda\u7d22\uff0c\u8f03\u6613\u770b\u6e05\u6a21\u578b\u504f\u898b\u5f9e\u54ea\u88e1\u4f86\u3002<\/p>\n","protected":false},"author":8,"featured_media":9557,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,178,140,180,30,119,170,149,188,197,199],"tags":[],"class_list":["post-9558","post","type-post","status-publish","format-standard","hentry","category-133","category-google","category-gemini","category-nanobanana","category-image","category-119","category-170","category-149","category-meta","category-framework","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9558","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=9558"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9558\/revisions"}],"predecessor-version":[{"id":9560,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9558\/revisions\/9560"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9557"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9558"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9558"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9558"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}