{"id":10676,"date":"2026-08-03T18:17:27","date_gmt":"2026-08-03T10:17:27","guid":{"rendered":"https:\/\/infernews.com\/blog\/tech-report-context-scaling-scaling-properties-of-text-conditioning-in-visual-ge\/"},"modified":"2026-08-03T18:20:24","modified_gmt":"2026-08-03T10:20:24","slug":"tech-report-context-scaling-scaling-properties-of-text-conditioning-in-visual-ge","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/tech-report-context-scaling-scaling-properties-of-text-conditioning-in-visual-ge\/","title":{"rendered":"Context Scaling \u628a\u6587\u751f\u5716\u63d0\u793a\u8a5e\u5e36\u5230\u7d50\u69cb\u5316\u968e\u6bb5"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/figure2_short.jpg\" alt=\"Text prompt scaling law overview\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u6587\u751f\u5716\u5361\u4f4f\u7684\u4f4d\u7f6e\uff0c\u5f88\u591a\u6642\u4e0d\u662f\u6a21\u578b\u4e0d\u5920\u5927\uff0c\u800c\u662f\u63d0\u793a\u8a5e\u4ea4\u4ee3\u5f97\u5514\u5920\u53ef\u8a08\u7b97\u3002Context Scaling \u805a\u7126\u5f71\u50cf\u751f\u6210\u4e2d\u7684\u6587\u5b57\u689d\u4ef6 scaling law\uff0c\u5c6c\u65bc\u4e00\u500b\u7d50\u5408\u7814\u7a76\u3001\u6a21\u578b\u8207\u5de5\u4f5c\u6d41\u7a0b\u8a2d\u8a08\u7684\u9805\u76ee\uff0c\u8655\u7406\u7684\u662f\u63d0\u793a\u8a5e\u600e\u6a23\u66f4\u6709\u6548\u63cf\u8ff0\u69cb\u5716\u3001\u5c6c\u6027\u540c\u7a7a\u9593\u95dc\u4fc2\uff0c\u4ee4 diffusion model \u66f4\u5bb9\u6613\u5b78\u5230\u3001\u4ea6\u66f4\u5bb9\u6613\u6309\u8981\u6c42\u751f\u6210\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u6700\u503c\u5f97\u7559\u610f\u7684\u5224\u65b7\uff0c\u662f caption \u9577\u5ea6\u672c\u8eab\u8ddf\u6548\u679c\u95dc\u4fc2\u5514\u7b97\u7dca\u5bc6\uff0c\u53cd\u800c structured language \u7684\u8cc7\u8a0a\u91cf\u66f4\u91cd\u8981\u3002\u5718\u968a\u7528\u5169\u500b\u6307\u6a19\u53bb\u91cf\u5ea6\u9019\u4ef6\u4e8b\uff1awhite-box likelihood metric \u7684 GPG\uff0c\u540c black-box attribute metric \u7684 ED\uff1b\u53d7\u63a7\u8a13\u7df4\u7d50\u679c\u986f\u793a\uff0cconverged diffusion loss \u6703\u96a8 GPG \u8fd1\u7dda\u6027\u4e0b\u964d\uff0c\u4ea6\u6703\u8ddf ED \u5448 power law \u95dc\u4fc2\u3002<\/p>\n\n\n<figure class=\"wp-block-embed-youtube wp-block-embed is-type-video is-provider-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"lyte-wrapper\" title=\"Context Scaling: Scaling Properties of Text Conditioning in Visual Generation\" style=\"width:853px;max-width:100%;margin:5px auto;\"><div class=\"lyMe\" id=\"WYL_pNUQAMGKVvc\" itemprop=\"video\" itemscope itemtype=\"https:\/\/schema.org\/VideoObject\"><div><meta itemprop=\"thumbnailUrl\" content=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2FpNUQAMGKVvc%2Fhqdefault.jpg\" \/><meta itemprop=\"embedURL\" content=\"https:\/\/www.youtube.com\/embed\/pNUQAMGKVvc\" \/><meta itemprop=\"duration\" content=\"PT5M9S\" \/><meta itemprop=\"uploadDate\" content=\"2026-08-03T02:00:19Z\" \/><\/div><div id=\"lyte_pNUQAMGKVvc\" data-src=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2FpNUQAMGKVvc%2Fhqdefault.jpg\" class=\"pL\"><div class=\"tC\"><div class=\"tT\" itemprop=\"name\">Context Scaling: Scaling Properties of Text Conditioning in Visual Generation<\/div><\/div><div class=\"play\"><\/div><div class=\"ctrl\"><div class=\"Lctrl\"><\/div><div class=\"Rctrl\"><\/div><\/div><\/div><noscript><a href=\"https:\/\/youtu.be\/pNUQAMGKVvc\" rel=\"nofollow\"><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2FpNUQAMGKVvc%2F0.jpg\" alt=\"Context Scaling: Scaling Properties of Text Conditioning in Visual Generation\" width=\"853\" height=\"460\" \/><br \/>Watch this video on YouTube<\/a><\/noscript><meta itemprop=\"description\" content=\"Scaling Properties of Text Conditioning in Visual Generation Zilong Chen, Chaorui Deng, Kunchang Li, Hongyi Yuan, Haoqi Fan ByteDance Seed Chapters 00:00 The fourth axis 00:25 Why longer captions don&#039;t help 00:54 Measuring information \u2014 GPG and ED 01:35 The scaling properties 02:11 Diffusability x Promptability 02:30 The structured prompt schema 02:54 Scaling and training the prompter 03:45 Results 04:01 Agentic inference-time scaling 04:25 Qualitative results 04:58 Summary\"><\/div><\/div><div class=\"lL\" style=\"max-width:100%;width:853px;margin:5px auto;\"><\/div><figcaption><\/figcaption><\/figure>\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u65b9\u5411\u4e0d\u53ea\u505c\u7559\u5728\u5206\u6790\u3002\u9805\u76ee\u628a Structured prompts\uff08SP\uff09\u505a\u6210\u5e36 semantic \u8207 geometric fields \u7684 JSON schema\uff0c\u518d\u914d\u5408 trainable LLM prompter + captioner\uff0c\u5c07\u7528\u6236\u8981\u6c42\u6216\u8005\u8f38\u5165\u5716\u7247\u8f49\u6210\u66f4\u6709\u7d50\u69cb\u7684\u63cf\u8ff0\uff1bzero-shot structured editing \u4ea6\u56e0\u6b64\u8b8a\u5f97\u53ef\u884c\uff0c\u56e0\u70ba\u6bcf\u500b\u53ef\u7de8\u8f2f\u56e0\u7d20\u90fd\u88ab\u62c6\u6210\u547d\u540d\u6b04\u4f4d\uff0c\u6539\u5176\u4e2d\u4e00\u90e8\u5206\u6642\uff0c\u5176\u4ed6\u69cb\u5716\u5143\u7d20\u8f03\u5bb9\u6613\u4fdd\u7559\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u91cd\u9ede\u4e0d\u5728\u63d0\u793a\u8a5e\u6709\u5e7e\u9577\uff0c\u800c\u5728\u756b\u9762\u8cc7\u8a0a\u6709\u5e7e\u53ef\u5c0d\u9f4a<\/li>\n\n\n\n<li>GPG \u8207 ED \u70ba\u63d0\u793a\u8a5e\u8cc7\u8a0a\u91cf\u63d0\u4f9b\u5169\u7a2e\u91cf\u5316\u65b9\u6cd5<\/li>\n\n\n\n<li>Structured prompts\uff08SP\uff09\u76f4\u63a5\u88dc\u5f37 compositional\u3001reasoning\u3001world-knowledge \u985e\u751f\u6210<\/li>\n\n\n\n<li>prompter \u7d93 supervised fine-tuning\u3001cold-start\u3001verifier-gated on-policy distillation \u8a13\u7df4<\/li>\n\n\n\n<li>\u5df2\u516c\u958b Code\u3001Models \u8207 Demo\uff0c\u7406\u89e3\u8def\u7dda\u53ef\u5148\u770b project page\uff0c\u518d\u5230 Hugging Face collection<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ddf\u540c\u985e\u505a\u6cd5\u76f8\u6bd4\uff0c\u5b83\u7684\u53d6\u6368\u76f8\u7576\u6e05\u695a\uff1a\u4e0d\u662f\u55ae\u9760\u66f4\u5927\u6a21\u578b\u6216\u66f4\u9577\u81ea\u7136\u8a9e\u8a00\u63cf\u8ff0\u53bb\u78b0\u904b\u6c23\uff0c\u800c\u662f\u5148\u63d0\u5347 captions \u5c0d\u5f71\u50cf\u7684\u53ef\u76e3\u7763\u6027\uff0c\u518d\u8a13\u7df4 prompter \u53bb\u7a69\u5b9a\u7522\u751f\u9019\u985e\u7d50\u69cb\u3002\u5b98\u65b9\u8aaa\u6cd5\u6307\u51fa\uff0c\u7cfb\u7d71\u5728\u591a\u500b compositional\u3001reasoning \u8207 world-knowledge benchmark \u4e0a\u8d85\u8d8a\u5168\u90e8\u5df2\u8a55\u6e2c open-weight models\uff0c\u4e26\u5728\u591a\u6578\u8a55\u6e2c\u8ffd\u5e73\u6216\u8d85\u904e\u6700\u5f37 closed-weight models\uff1b\u4e0d\u904e\u516c\u958b\u8cc7\u8a0a\u4ecd\u4ee5\u7814\u7a76\u7d50\u679c\u70ba\u4e3b\uff0c\u90e8\u7f72\u7d30\u7bc0\u8207\u5b8c\u6574\u5b89\u88dd\u6d41\u7a0b\u672a\u7b97\u591a\uff0c\u73fe\u968e\u6bb5\u8f03\u9069\u5408\u505a\u65b9\u6cd5\u7814\u7a76\u3001\u63d0\u793a\u8a5e\u5de5\u7a0b\u3001\u5f71\u50cf\u7de8\u8f2f\u6d41\u7a0b\u8a2d\u8a08\uff0c\u4ee5\u53ca\u8a55\u4f30\u4e0b\u4e00\u4ee3\u6587\u751f\u5716\u4ecb\u9762\u7684\u5718\u968a\u53c3\u8003\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/heheyas.github.io\/context-scaling\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/heheyas\/context-scaling\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/collections\/heheyas\/context-scaling\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5beb\u5f97\u66f4\u9577\u672a\u5fc5\u66f4\u6e96\uff0c\u95dc\u9375\u53cd\u800c\u662f\u63d0\u793a\u8a5e\u6709\u5e7e\u591a\u53ef\u5c0d\u61c9\u756b\u9762\u7684\u7d50\u69cb\u8cc7\u8a0a\u3002Context Scaling \u628a\u9019\u4ef6\u4e8b\u91cf\u5316\uff0c\u9023\u5e36\u6539\u5beb\u6587\u751f\u5716\u8a13\u7df4\u8207\u7de8\u8f2f\u6d41\u7a0b\u3002<\/p>\n","protected":false},"author":8,"featured_media":10675,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,176,129,157,76,127,199],"tags":[],"class_list":["post-10676","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-176","category-txt2img","category-157","category-76","category-127","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10676","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=10676"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10676\/revisions"}],"predecessor-version":[{"id":10678,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10676\/revisions\/10678"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10675"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10676"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10676"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10676"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}