{"id":2889,"date":"2024-10-18T16:23:13","date_gmt":"2024-10-18T08:23:13","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=2889"},"modified":"2024-10-18T16:25:01","modified_gmt":"2024-10-18T08:25:01","slug":"%e6%af%94flux%e6%9b%b4%e5%bc%ba%e5%a4%a7%e7%9a%84%e6%96%87%e7%94%9f%e5%9b%be%e6%a8%a1%e5%9e%8b%e6%9d%a5%e4%ba%86%ef%bc%81%e7%a7%98%e8%af%80%e6%98%af%e9%9b%86%e7%99%be%e5%ae%b6%e4%b9%8b","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2024\/10\/18\/%e6%af%94flux%e6%9b%b4%e5%bc%ba%e5%a4%a7%e7%9a%84%e6%96%87%e7%94%9f%e5%9b%be%e6%a8%a1%e5%9e%8b%e6%9d%a5%e4%ba%86%ef%bc%81%e7%a7%98%e8%af%80%e6%98%af%e9%9b%86%e7%99%be%e5%ae%b6%e4%b9%8b\/","title":{"rendered":"\u6bd4Flux\u66f4\u5f3a\u5927\u7684\u6587\u751f\u56fe\u6a21\u578b\u6765\u4e86\uff01\u79d8\u8bc0\u662f\u201c\u96c6\u767e\u5bb6\u4e4b\u957f\u201d\u00a0"},"content":{"rendered":"\t\t
\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

\u6253\u9020\u66f4\u5f3a\u5927\u6587\u751f\u56fe\u6a21\u578b\u00a0\u65b0\u601d\u8def<\/strong>\u6709\u2014\u2014<\/p>

\u9762\u5bf9Flux\u3001stable diffusion\u3001Omost\u7b49\u7206\u706b\u6a21\u578b\uff0c\u6709\u4eba\u5f00\u59cb\u4e3b\u6253\u00a0\u201c\u96c6\u5404\u5bb6\u6240\u957f\u201d\u3002<\/strong><\/span><\/p>

\u5177\u4f53\u6765\u8bf4\uff0c\u6e05\u5317\u3001\u725b\u6d25\u3001\u666e\u6797\u65af\u987f\u7b49\u673a\u6784\u7684\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u2f00\u4e2a\u00a0\u5168\u65b0\u6587\u751f\u56fe\u6846\u67b6IterComp<\/strong>\u3002<\/p>

\u5b83\u80fd\u63d0\u53d6\u4e0d\u540c\u6a21\u578b\u5728\u5404\u81ea\u9886\u57df\u7684\u4f18\u52bf\uff0c\u540c\u65f6\u4e0d\u5f15\u5165\u989d\u5916\u7684\u590d\u6742\u6307\u4ee4\u6216\u589e\u52a0\u8ba1\u7b97\u5f00\u9500\u3002<\/p>

<\/p>

\u8bba\u6587\u4e00\u7ecf\u53d1\u5e03\uff0c\u5373\u5728\ud835\udd4f\u00a0\uff08\u524d\u63a8\u7279\uff09<\/span>\u83b7\u5f97AI\u8bba\u6587\u9886\u57df\u5927V\u8f6c\u53d1\uff0c\u5438\u5f15\u5927\u91cf\u5173\u6ce8\u3002<\/p>

<\/p>

\u90a3\u4e48\uff0c\u7814\u7a76\u4eba\u5458\u5177\u4f53\u662f\u5982\u4f55\u5b9e\u73b0\u7684\u5462\uff1f<\/p>

\u5168\u65b0\u6587\u751f\u56fe\u6846\u67b6\uff1aIterComp<\/p>

\u2f832022\u5e74\u4ee5\u6765\uff0c\u57fa\u4e8ediffusion\u7684\u6587\u751f\u56fe\u6a21\u578b\u53d6\u5f97\u4e86\u5feb\u901f\u53d1\u5c55\uff0c\u5c24\u5176\u5728\u590d\u6742\u7ec4\u5408\u751f\u6210\uff08complex\/compositional generation\uff09\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\u3002<\/p>

\u4f8b\u5982\uff0c\u4eca\u5e748\u6708\u53d1\u5e03\u7684Flux\u5c55\u73b0\u51fa\u4e86\u5341\u5206\u9707\u64bc\u7684\u590d\u6742\u573a\u666f\u751f\u6210\u80fd\u529b\u4e0e\u7f8e\u5b66\u8d28\u91cf\uff1b<\/p>

RPG\u901a\u8fc7MLLM\u7684\u8f85\u52a9\uff0c\u5c06\u590d\u6742\u7684\u7ec4\u5408\u751f\u6210\u4efb\u52a1\u5206\u89e3\u4e3a\u7b80\u5355\u5b50\u4efb\u52a1\uff1b<\/p>

InstanceDiffusion\u901a\u8fc7\u5e03\u5c40\u63a7\u5236\uff08layoutbased\uff09\uff0c\u5b9e\u73b0\u4e86\u4e0e\u5e03\u5c40\u5206\u5e03\u9ad8\u5ea6\u4e00\u81f4\u7684\u7cbe\u786e\u56fe\u50cf\u751f\u6210\u3002<\/p>

<\/p>

\u7136\u800c\uff0c\u8fd9\u4e9b\u6a21\u578b\u7684\u4f18\u52bf\u00a0\u4ec5\u9650\u4e8e\u67d0\u4e9b\u7279\u5b9a\u7684\u7ec4\u5408\u751f\u6210\u4efb\u52a1\uff0c\u4e14\u5b58\u5728\u8bf8\u591a\u5c40\u9650<\/strong>\u3002<\/p>

\u57fa\u4e8e\u6587\u672c\u7684\u751f\u6210\u65b9\u6cd5<\/strong>\uff08\u5982SDXL\u3001SD3\u3001Flux\u7b49\uff09\uff0c\u7531\u4e8e\u5176\u8bad\u7ec3\u6570\u636e\u548c\u6846\u67b6\u7684\u4f18\u52bf\uff0c\u5728\u7269\u4f53\u4e0e\u573a\u666f\u7684\u6a21\u62df\u4e0a\u8868\u73b0\u51fa\u8272\uff0c\u7279\u522b\u5728\u7f8e\u5b66\u8d28\u91cf\u3001\u5c5e\u6027\u7ed1\u5b9a\u548c\u975e\u7a7a\u95f4\u5173\u7cfb\uff08non-spatial relationships\uff09\u65b9\u9762\u5177\u6709\u663e\u8457\u4f18\u52bf\u3002<\/p>

\u7136\u800c\uff0c\u5f53\u6d89\u53ca\u591a\u4e2a\u7269\u4f53\uff0c\u4e14\u5b58\u5728\u590d\u6742\u7684\u7a7a\u95f4\u5173\u7cfb\u65f6\uff0c\u8fd9\u7c7b\u6a21\u578b\u5f80\u5f80\u8868\u73b0\u51fa\u660e\u663e\u4e0d\u8db3\u3002<\/p>

\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM-based\uff09<\/strong>\u7684\u751f\u6210\u65b9\u6cd5\uff0c\u5982RPG\u548cOmost\uff0c\u901a\u5e38\u9700\u8981\u5bf9LLM\u8fdb\u2f8f\u989d\u5916\u8bad\u7ec3\u6216\u8bbe\u8ba1\u590d\u6742\u7684\u63d0\u793a\u3002<\/p>

\u7136\u800c\uff0c\u5bf9\u4e8eLLM\u5904\u7406\u540e\u7684\u590d\u6742\u6307\u4ee4\uff0cdiffusion backbone\u5e76\u4e0d\u5177\u5907\u7cbe\u786e\u751f\u6210\u7684\u80fd\u529b\u3002<\/p>

\u57fa\u4e8e\u5e03\u5c40\uff08layourbased\uff09<\/strong>\u7684\u751f\u6210\u65b9\u6cd5\uff0c\u5982Instancediffusion\u548cLMD+\uff0c\u867d\u7136\u63d0\u4f9b\u4e86\u7cbe\u786e\u7684\u63a7\u5236\uff0c\u4f46\u5728\u56fe\u50cf\u7684\u7f8e\u5b66\u8d28\u91cf\u548c\u771f\u5b9e\u6027\u4e0a\u5b58\u5728\u660e\u663e\u4e0b\u964d\uff0c\u5e76\u4e14\u9700\u8981\u4eba\u4e3a\u8bbe\u8ba1\u5e03\u5c40\u3002<\/p>

\u56e0\u6b64\uff0c\u4e00\u4e2a\u503c\u5f97\u6df1\u2f0a\u63a2\u8ba8\u7684\u95ee\u9898\u662f\uff1a<\/p>

\u80fd\u5426\u8bbe\u8ba1\u51fa\u4e00\u4e2a\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u5728\u4e0a\u8ff0\u591a\u4e2a\u65b9\u9762\u90fd\u8868\u73b0\u51fa\u2f8a\uff0c\u540c\u65f6\u4e0d\u5f15\u5165\u989d\u5916\u7684\u590d\u6742\u6307\u4ee4\u6216\u589e\u52a0\u8ba1\u7b97\u5f00\u9500\uff1f<\/strong><\/p><\/blockquote>

\u57fa\u4e8e\u6b64\uff0c\u7814\u7a76\u4eba\u5458\u63d0\u51fa\u4e00\u4e2a\u5168\u65b0\u7684\u6587\u751f\u56fe\u6846\u67b6\uff1a\u00a0IterComp<\/strong>\u3002<\/p>

<\/p>

\u8981\u5145\u5206\u89e3\u51b3\u8fd9\u2f00\u95ee\u9898\uff0c\u7814\u7a76\u9762\u4e34\u00a0\u4e24\u5927\u96be\u70b9<\/strong>\uff1a<\/p>

1\u3001<\/strong><\/span>\u5982\u4f55\u63d0\u53d6\u4e0d\u540c\u6a21\u578b\u5728\u5404\u81ea\u9886\u57df\u7684\u4f18\u52bf\uff0c\u5e76\u5f15\u5bfc\u6a21\u578b\u8fdb\u884c\u6709\u6548\u5b66\u4e60\uff1f<\/strong><\/p>

\u9488\u5bf9\u8be5\u95ee\u9898\uff0c\u7814\u7a76\u4eba\u5458\u9996\u5148\u6784\u5efa\u4e86\u6a21\u578b\u5e93\uff08model gallery)\uff0c\u5176\u5305\u542b\u591a\u79cd\u5728\u4e0d\u540c\u65b9\u9762\u8868\u73b0\u51fa\u8272\u7684\u6a21\u578b\uff0c\u6bcf\u4e2a\u6a21\u578b\u90fd\u5177\u5907\u7279\u5b9a\u7684\u7ec4\u5408\u751f\u6210\u504f\u597d\uff08composition-aware model preference\uff09\u3002<\/p>

\u7814\u7a76\u4eba\u5458\u5c1d\u8bd5\u901a\u8fc7\u00a0\u6269\u6563\u6a21\u578b\u5bf9\u9f50\u65b9\u6cd5<\/strong>\uff0c\u5c06base diffusion model\u4e0e\u6a21\u578b\u5e93\u4e2d\u7684\u7ec4\u5408\u751f\u6210\u504f\u597d\u8fdb\u884c\u5bf9\u9f50\u3002<\/p>

\u56e2\u961f\u805a\u7126\u4e8ecompositional generation\u7684\u4e09\u4e2a\u5173\u952e\u65b9\u9762\uff1a<\/p>

  • \u5c5e\u6027\u7ed1\u5b9a\uff08attribute binding\uff09<\/li>
  • \u7a7a\u95f4\u5173\u7cfb\uff08spatial relationship\uff09<\/li>
  • \u2fae\u7a7a\u95f4\u5173\u7cfb\uff08non-spatial relationship\uff09<\/li><\/ul>

    \u4e3a\u6b64\uff0c\u7814\u7a76\u4eba\u5458\u6536\u96c6\u4e86\u4e0d\u540c\u6a21\u578b\u5728\u8fd9\u4e9b\u65b9\u9762\u7684\u504f\u597d\u6570\u636e\uff0c\u5e76\u901a\u8fc7\u4eba\u5de5\u6392\u5e8f\uff0c\u6784\u5efa\u4e86\u2f00\u4e2a\u9762\u5411\u7ec4\u5408\u751f\u6210\u7684\u00a0\u6a21\u578b\u504f\u597d\u6570\u636e\u96c6<\/strong>\uff08composition-aware model preference dataset\uff09\u3002<\/p>

    <\/p>

    \u9488\u5bf9\u8fd9\u4e09\u4e2a\u65b9\u9762\uff0c\u56e2\u961f\u5206\u522b\u8bad\u7ec3\u4e09\u4e2acomposition-aware reward models\uff0c\u5e76\u5bf9base\u6a21\u578b\u8fdb\u884c\u00a0\u591a\u5956\u52b1\u53cd\u9988\u4f18\u5316<\/strong>\u3002<\/p>

    2\u3001\u7ec4\u5408\u751f\u6210\u4efb\u52a1\u5f88\u96be\u4f18\u5316\uff0c\u5982\u4f55\u66f4\u5145\u5206\u5730\u5b66\u4e60\u5230\u4e0d\u540c\u6a21\u578b\u7684\u7ec4\u5408\u504f\u597d\uff1f<\/strong><\/p>

    \u7814\u7a76\u4eba\u5458\u5728diffusion\u9886\u57df\u521b\u65b0\u5730\u5f15\u5165\u00a0\u8fed\u4ee3\u5f0f\u5b66\u4e60\u6846\u67b6<\/strong>\uff08iterative feedback learning\uff09\uff0c\u5b9e\u73b0reward models\u4e0ebase diffusion model\u4e4b\u95f4\u201c\u5de6\u811a\u8e29\u53f3\u811a\u767b\u5929\u201d\u3002<\/p>

    \u5177\u4f53\u6765\u8bf4\uff0c\u5728\u4e0a\u8ff0\u7b2c\u4e00\u8f6e\u4f18\u5316\u540e\uff0c\u56e2\u961f\u5c06optimized base diffusion model\u4ee5\u53ca\u989d\u5916\u9009\u62e9\u7684\u5176\u4ed6\u6a21\u578b\uff08\u4f8b\u5982Omost\u7b49\uff09\u6dfb\u52a0\u8fdbmodel gallery\u3002<\/p>

    \u5bf9\u65b0\u589e\u6a21\u578b\u8fdb\u884c\u504f\u597d\u91c7\u6837\uff0c\u4e0e\u521d\u59cbmodel gallery\u4e2d\u7684\u56fe\u50cf\u6309\u7167\u9884\u8bad\u7ec3\u7684reward model\u6784\u5efa\u56fe\u50cf\u5bf9\u3002<\/p>

    \u8fd9\u4e9b\u56fe\u50cf\u5bf9\u88ab\u7528\u4e8e\u8fdb\u2f00\u6b65\u4f18\u5316\u5956\u52b1\u6a21\u578b\uff0c\u968f\u540e\u7ee7\u7eed\u7528\u66f4\u65b0\u7684\u5956\u52b1\u6a21\u578b\u4f18\u5316\u57fa\u7840\u6269\u6563\u6a21\u578b\u3002<\/p>

    \u5177\u4f53\u7684\u6d41\u7a0b\u5982\u4e0b\u4f2a\u4ee3\u7801\u6240\u793a\uff1a<\/p>

    <\/p>

    \u5b9e\u9a8c<\/p>

    \u5728\u00a0\u5b9a\u6027\u5b9e\u9a8c<\/strong>\u4e2d\uff0c\u4e0e\u5176\u4ed6\u4e09\u7c7bcompositional generation\u65b9\u6cd5\u76f8\u6bd4\uff0cIterComp\u53d6\u5f97\u4e86\u663e\u8457\u7684\u7ec4\u5408\u8d28\u91cf\u7684\u63d0\u5347\uff0c\u5e76\u4e14\u4e0d\u4f1a\u5f15\u5165\u989d\u5916\u7684\u8ba1\u7b97\u91cf\u3002<\/p>

    <\/p>

    \u4ece\u5b9a\u91cf\u7ed3\u679c\u4e2d\u53ef\u4ee5\u770b\u51fa\uff0cIterComp\u5728T2I-CompBench\u4e0a\u53d6\u5f97\u4e86\u5404\u65b9\u9762\u7684\u9886\u5148\u3002<\/p>

    \u53e6\u5916\uff0c\u9488\u5bf9\u56fe\u50cf\u771f\u5b9e\u6027\u4ee5\u53ca\u7f8e\u5b66\u8d28\u91cf\uff0cIterComp\u4e5f\u8868\u73b0\u51fa\u8272\u3002<\/p>

    <\/p>

    \u7136\u800cIterComp\u7684\u5e94\u7528\u6f5c\u529b\u4e0d\u9650\u4e8e\u6b64\uff0c\u5176\u53ef\u4ee5\u4f5c\u4e3a\u5f3a\u5927\u7684backbone\u663e\u8457\u63d0\u5347Omost, RPG\u7b49\u6a21\u578b\u7684\u751f\u6210\u80fd\u529b\u3002<\/p>

    <\/p>

    \u66f4\u591a\u7ec6\u8282\u6b22\u8fce\u67e5\u9605\u539f\u8bba\u6587\u3002<\/p>

    \u8bba\u6587\u5730\u5740\uff1a<\/span><\/p>

    https:\/\/arxiv.org\/abs\/2410.07171<\/span><\/p>

    \u4ee3\u7801\u5730\u5740\uff1a<\/span><\/p>

    https:\/\/github.com\/YangLing0818\/IterComp<\/span><\/p>

    \u6a21\u578b\u5730\u5740\uff1a<\/span><\/p>

    https:\/\/huggingface.co\/comin\/IterComp<\/span><\/p>

    Civitai:<\/span><\/p>

    \u2014\u00a0\u5b8c<\/strong>\u2014<\/span><\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"

    IterComp\u56e2\u961f \u6295\u7a3f<\/p>\n","protected":false},"author":1,"featured_media":2891,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2889","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/2889","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=2889"}],"version-history":[{"count":8,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/2889\/revisions"}],"predecessor-version":[{"id":2898,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/2889\/revisions\/2898"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/2891"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=2889"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=2889"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=2889"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}