{"id":15728,"date":"2024-07-06T14:31:00","date_gmt":"2024-07-06T06:31:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=15728"},"modified":"2024-07-05T14:31:34","modified_gmt":"2024-07-05T06:31:34","slug":"dit-transformer","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/07\/06\/dit-transformer\/","title":{"rendered":"\u6269\u6563\u6a21\u578bDiT\uff1a\u7eafTransformer\u67b6\u6784\u66ff\u6362\u4f20\u7edf\u7684U-Net\u4e3b\u5e72"},"content":{"rendered":"\n

DiT\u539f\u6587: <\/a>https:\/\/arxiv.org\/abs\/2212.09748<\/a> Code:https:\/\/github.com\/facebookresearch\/DiT<\/a> Huggingface:https:\/\/huggingface.co\/spaces\/wpeebles\/DiT<\/a><\/strong><\/p>\n\n\n\n

DiT\u6269\u6563\u6a21\u578b\uff0c\u51fa\u81ea\u8c22\u8d5b\u5b81\u4e0eSora\u7814\u53d1\u4e4b\u4e00\u5a01\u5ec9\u00b7\u76ae\u6ce2\u5c14\u65af\u5408\u8457\u7684\u4e00\u7bc7\u8bba\u6587\u300aScalable diffusion models with transformers\u300b\uff0c\u5b83\u4e0d\u4ec5\u5c06 Transformer \u6210\u529f\u5e94\u7528\u5230\u4e86\u6269\u6563\u6a21\u578b\u4e0a\uff0c\u8fd8\u6df1\u5165\u63a2\u7a76\u4e86transformer\u67b6\u6784\u5728\u6269\u6563\u6a21\u578b\u4e0a\u7684scalability\u80fd\u529b\u3002<\/p>\n\n\n\n

\u6838\u5fc3\u601d\u60f3\uff1a\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\uff08\u53ef\u6269\u5c55\uff09\u6269\u6563\u6a21\u578b\u67b6\u6784\uff0c\u79f0\u4e3a DiT\uff0c\u8be5\u67b6\u6784\u4f7f\u7528 Transformer \u66ff\u6362\u4e86\u4f20\u7edf\u7684 U-Net \u4e3b\u5e72\u3002<\/strong><\/p>\n\n\n\n

\u6269\u6563\u6a21\u578b\u5927\u90e8\u5206\u662f\u91c7\u7528UNet\u67b6\u6784<\/strong>\u6765\u8fdb\u884c\u5efa\u6a21\uff0cUNet\u53ef\u4ee5\u5b9e\u73b0\u8f93\u51fa\u548c\u8f93\u5165\u4e00\u6837\u7ef4\u5ea6\uff0c\u6240\u4ee5\u5929\u7136\u9002\u5408\u6269\u6563\u6a21\u578b\u3002\u6269\u6563\u6a21\u578b\u4f7f\u7528\u7684UNet\u9664\u4e86\u5305\u542b\u57fa\u4e8e\u6b8b\u5dee\u7684\u5377\u79ef\u6a21\u5757\uff0c\u540c\u65f6\u4e5f\u5f80\u5f80\u91c7\u7528self-attention\u3002\u81ea\u4eceViT\u4e4b\u540e\uff0ctransformer\u67b6\u6784\u5df2\u7ecf\u5927\u91cf\u5e94\u7528\u5728\u56fe\u50cf\u4efb\u52a1\u4e0a\uff0c\u968f\u7740\u6269\u6563\u6a21\u578b\u7684\u6d41\u884c\uff0c\u4e5f\u5df2\u7ecf\u6709\u5de5\u4f5c\u5c1d\u8bd5\u91c7\u7528transformer\u67b6\u6784\u6765\u5bf9\u6269\u6563\u6a21\u578b\u5efa\u6a21\uff0c\u8fd9\u7bc7\u6587\u7ae0\u6211\u4eec\u5c06\u4ecb\u7ecdMeta\u7684\u5de5\u4f5cDiT<\/strong>\uff1aScalable Diffusion Models with Transformers<\/strong>\uff0c\u5b83\u662f\u5b8c\u5168\u57fa\u4e8etransformer\u67b6\u6784\u7684\u6269\u6563\u6a21\u578b<\/strong>\uff0c\u8fd9\u4e2a\u5de5\u4f5c\u4e0d\u4ec5\u5c06transformer\u6210\u529f\u5e94\u7528\u5728\u6269\u6563\u6a21\u578b\uff0c\u8fd8\u63a2\u7a76\u4e86transformer\u67b6\u6784\u5728\u6269\u6563\u6a21\u578b\u4e0a\u7684scalability\u80fd\u529b<\/strong>\uff0c\u5176\u4e2d\u6700\u5927\u7684\u6a21\u578bDiT-XL\/2\u5728ImageNet 256×256\u7684\u7c7b\u522b\u6761\u4ef6\u751f\u6210\u4e0a\u8fbe\u5230\u4e86SOTA\uff08FID\u4e3a2.27\uff09<\/strong>\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

DiT\u57fa\u672c\u6cbf\u7528\u4e86ViT\u7684\u8bbe\u8ba1\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u9996\u5148\u91c7\u7528\u4e00\u4e2apatch embedding\u6765\u5c06\u8f93\u5165\u8fdb\u884cpatch\u5316<\/strong>\uff0c\u5373\u5f97\u5230\u4e00\u7cfb\u5217\u7684tokens\u3002\u5176\u4e2dpatch size\u5c5e\u4e8e\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u5b83\u76f4\u63a5\u51b3\u5b9a\u4e86tokens\u7684\u6570\u91cf\uff0c\u8fd9\u4f1a\u5f71\u54cd\u6a21\u578b\u7684\u8ba1\u7b97\u91cf\u3002DiT\u7684patch size\u5171\u9009\u62e9\u4e86\u4e09\u79cd\u8bbe\u7f6e\uff1a\ud835\udc5d=2,4,8\u3002\u6ce8\u610ftoken\u5316\u4e4b\u540e\uff0c\u8fd9\u91cc\u8fd8\u8981\u52a0\u4e0apositional embeddings\uff0c\u8fd9\u91cc\u91c7\u7528\u975e\u5b66\u4e60\u7684sin-cosine\u4f4d\u7f6e\u7f16\u7801\u3002<\/p>\n\n\n\n

\u5c06\u8f93\u5165token\u5316\u4e4b\u540e\uff0c\u5c31\u53ef\u4ee5\u50cfViT\u90a3\u6837\u63a5transformer blocks\u4e86\u3002\u4f46\u662f\u5bf9\u4e8e\u6269\u6563\u6a21\u578b\u6765\u8bf4\uff0c\u5f80\u5f80\u8fd8\u9700\u8981\u5728\u7f51\u7edc\u4e2d\u5d4c\u5165\u989d\u5916\u7684\u6761\u4ef6\u4fe1\u606f\uff0c\u8fd9\u91cc\u7684\u6761\u4ef6\u5305\u62ectimesteps\u4ee5\u53ca\u7c7b\u522b\u6807\u7b7e\uff08\u5982\u679c\u662f\u6587\u751f\u56fe\u5c31\u662f\u6587\u672c\uff0c\u4f46\u662fDiT\u8fd9\u91cc\u5e76\u6ca1\u6709\u6d89\u53ca\uff09\u3002\u8981\u8bf4\u660e\u7684\u4e00\u70b9\u662f\uff0c\u65e0\u8bba\u662ftimesteps\u8fd8\u662f\u7c7b\u522b\u6807\u7b7e\uff0c\u90fd\u53ef\u4ee5\u91c7\u7528\u4e00\u4e2aembedding\u6765\u8fdb\u884c\u7f16\u7801\u3002DiT\u5171\u8bbe\u8ba1\u4e86\u56db\u79cd\u65b9\u6848\u6765\u5b9e\u73b0\u4e24\u4e2a\u989d\u5916embeddings\u7684\u5d4c\u5165\uff0c\u5177\u4f53\u5982\u4e0b\uff1a<\/p>\n\n\n\n

  1. In-context conditioning<\/strong>\uff1a\u5c06\u4e24\u4e2aembeddings\u770b\u6210\u4e24\u4e2atokens\u5408\u5e76\u5728\u8f93\u5165\u7684tokens\u4e2d\uff0c\u8fd9\u79cd\u5904\u7406\u65b9\u5f0f\u6709\u70b9\u7c7b\u4f3cViT\u4e2d\u7684cls token\uff0c\u5b9e\u73b0\u8d77\u6765\u6bd4\u8f83\u7b80\u5355\uff0c\u4e5f\u4e0d\u57fa\u672c\u4e0a\u4e0d\u989d\u5916\u5f15\u5165\u8ba1\u7b97\u91cf\u3002<\/li>
  2. Cross-attention block<\/strong>\uff1a\u5c06\u4e24\u4e2aembeddings\u62fc\u63a5\u6210\u4e00\u4e2a\u6570\u91cf\u4e3a2\u7684\u5e8f\u5217\uff0c\u7136\u540e\u5728transformer block\u4e2d\u63d2\u5165\u4e00\u4e2across attention\uff0c\u6761\u4ef6embeddings\u4f5c\u4e3across attention\u7684key\u548cvalue\uff1b\u8fd9\u79cd\u65b9\u5f0f\u4e5f\u662f\u76ee\u524d\u6587\u751f\u56fe\u6a21\u578b\u6240\u91c7\u7528\u7684\u65b9\u5f0f\uff0c\u5b83\u9700\u8981\u989d\u5916\u5f15\u516515%\u7684Gflops\u3002<\/li>
  3. Adaptive layer norm (adaLN) block<\/strong>\uff1a\u91c7\u7528adaLN\uff0c\u8fd9\u91cc\u662f\u5c06time embedding\u548cclass embedding\u76f8\u52a0\uff0c\u7136\u540e\u6765\u56de\u5f52scale\u548cshift\u4e24\u4e2a\u53c2\u6570\uff0c\u8fd9\u79cd\u65b9\u5f0f\u4e5f\u57fa\u672c\u4e0d\u589e\u52a0\u8ba1\u7b97\u91cf\u3002<\/li>
  4. adaLN-Zero block<\/strong>\uff1a\u91c7\u7528zero\u521d\u59cb\u5316\u7684adaLN\uff0c\u8fd9\u91cc\u662f\u5c06adaLN\u7684linear\u5c42\u53c2\u6570\u521d\u59cb\u5316\u4e3azero\uff0c\u8fd9\u6837\u7f51\u7edc\u521d\u59cb\u5316\u65f6transformer block\u7684\u6b8b\u5dee\u6a21\u5757\u5c31\u662f\u4e00\u4e2aidentity\u51fd\u6570\uff1b\u53e6\u5916\u4e00\u70b9\u662f\uff0c\u8fd9\u91cc\u9664\u4e86\u5728LN\u4e4b\u540e\u56de\u5f52scale\u548cshift\uff0c\u8fd8\u5728\u6bcf\u4e2a\u6b8b\u5dee\u6a21\u5757\u7ed3\u675f\u4e4b\u524d\u56de\u5f52\u4e00\u4e2ascale\uff0c\u5982\u4e0a\u56fe\u6240\u793a\u3002<\/li><\/ol>\n\n\n\n

    \u8bba\u6587\u5bf9\u56db\u79cd\u65b9\u6848\u8fdb\u884c\u4e86\u5bf9\u6bd4\u8bd5\u9a8c\uff0c\u53d1\u73b0\u91c7\u7528adaLN-Zero<\/strong>\u6548\u679c\u662f\u6700\u597d\u7684\uff0c\u6240\u4ee5DiT\u9ed8\u8ba4\u90fd\u91c7\u7528\u8fd9\u79cd\u65b9\u5f0f\u6765\u5d4c\u5165\u6761\u4ef6embeddings\u3002<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    \u867d\u7136DiT\u53d1\u73b0adaLN-Zero<\/strong>\u6548\u679c\u662f\u6700\u597d\u7684\uff0c\u4f46\u662f\u8fd9\u79cd\u65b9\u5f0f\u53ea\u9002\u5408\u8fd9\u79cd\u53ea\u6709\u7c7b\u522b\u4fe1\u606f\u7684\u7b80\u5355\u6761\u4ef6\u5d4c\u5165\uff0c\u56e0\u4e3a\u53ea\u9700\u8981\u5f15\u5165\u4e00\u4e2aclass embedding\uff1b\u4f46\u662f\u5bf9\u4e8e\u6587\u751f\u56fe\u6765\u8bf4\uff0c\u5176\u6761\u4ef6\u5f80\u5f80\u662f\u5e8f\u5217\u7684text embeddings\uff0c\u91c7\u7528cross-attention\u65b9\u6848\u53ef\u80fd\u662f\u66f4\u5408\u9002\u7684\u3002 \u7531\u4e8e\u5bf9\u8f93\u5165\u8fdb\u884c\u4e86token\u5316\uff0c\u6240\u4ee5\u5728\u7f51\u7edc\u7684\u6700\u540e\u8fd8\u9700\u8981\u4e00\u4e2adecoder\u6765\u6062\u590d\u8f93\u5165\u7684\u539f\u59cb\u7ef4\u5ea6\uff0cDiT\u91c7\u7528\u4e00\u4e2a\u7b80\u5355\u7684linear\u5c42\u6765\u5b9e\u73b0\uff0c\u76f4\u63a5\u5c06\u6bcf\u4e2atoken\u6620\u5c04\u4e3a\ud835\udc5d\u00d7\ud835\udc5d\u00d72\ud835\udc36\u7684tensor\uff0c\u7136\u540e\u518d\u8fdb\u884creshape\u6765\u5f97\u5230\u548c\u539f\u59cb\u8f93\u5165\u7a7a\u95f4\u7ef4\u5ea6\u4e00\u6837\u7684\u8f93\u51fa\uff0c\u4f46\u662f\u7279\u5f81\u7ef4\u5ea6\u5927\u5c0f\u662f\u539f\u6765\u76842\u500d\uff0c\u5206\u522b\u7528\u6765\u9884\u6d4b\u566a\u97f3\u548c\u65b9\u5dee\u3002<\/p>\n\n\n\n

    \u6ce8\u610f\u8fd9\u91cc\u5148\u8fdb\u884cLayerNorm\uff0c\u540c\u65f6\u4e5f\u5f15\u5165\u4e86zero adaLN\uff0c\u5e76\u4e14decoder\u7684linear\u5c42\u4e5f\u91c7\u7528zero\u521d\u59cb\u5316\u3002 \u4eff\u7167ViT\uff0cDiT\u4e5f\u8bbe\u8ba1\u4e864\u79cd\u4e0d\u540c\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u5206\u522b\u662fDiT-S\u3001DiT-B\u3001DiT-L\u548cDiT-XL\uff0c\u5176\u4e2d\u6700\u5927\u7684\u6a21\u578bDiT-XL\u53c2\u6570\u91cf\u4e3a675M\uff0c\u8ba1\u7b97\u91cfGflops\u4e3a29.1\uff08256×256\u56fe\u50cf\uff0cpatch size=4\u65f6\uff09\u3002<\/p>\n\n\n\n

    \u5728\u5177\u4f53\u6027\u80fd\u4e0a\uff0c\u6700\u5927\u7684\u6a21\u578bDiT-XL\/2\u91c7\u7528classifier free guidance\u53ef\u4ee5\u5728class-conditional image generation on ImageNet 256\u00d7256\u4efb\u52a1\u4e0a\u5b9e\u73b0\u5f53\u65f6\u7684sota\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u867d\u7136DiT\u770b\u8d77\u6765\u4e0d\u9519\uff0c\u4f46\u662f\u53ea\u5728ImageNet\u4e0a\u751f\u6210\u505a\u4e86\u5b9e\u9a8c\uff0c\u5e76\u6ca1\u6709\u6269\u5c55\u5230\u5927\u89c4\u6a21\u7684\u6587\u751f\u56fe\u6a21\u578b\u3002\u800c\u4e14\u5728DiT\u4e4b\u524d\uff0c\u5176\u5b9e\u4e5f\u6709\u57fa\u4e8etransformer\u67b6\u6784\u7684\u6269\u6563\u6a21\u578b\u7814\u7a76\u5de5\u4f5c\uff0c\u6bd4\u5982U-ViT\uff0c\u76ee\u524d\u4e5f\u5df2\u7ecf\u6709\u5c06transformer\u5e94\u7528\u5728\u5927\u89c4\u6a21\u6587\u751f\u56fe\uff08\u57fa\u4e8e\u6269\u6563\u6a21\u578b\uff09\u7684\u5de5\u4f5c\uff0c\u6bd4\u5982UniDiffuser\uff0c\u4f46\u662f\u5176\u5b9e\u90fd\u6ca1\u6709\u53d7\u5230\u592a\u5927\u7684\u5173\u6ce8\u3002\u76ee\u524d\u4e3b\u6d41\u7684\u6587\u751f\u56fe\u6a21\u578b\u8fd8\u662f\u91c7\u7528\u57fa\u4e8eUNet\uff0cUNet\u672c\u8eab\u4e5f\u6df7\u5408\u4e86\u5377\u79ef\u548cattention\uff0c\u5b83\u7684\u4f18\u52bf\u4e00\u65b9\u9762\u662f\u9ad8\u6548\uff0c\u53e6\u5916\u4e00\u65b9\u9762\u662f\u4e0d\u9700\u8981\u4f4d\u7f6e\u7f16\u7801\u6bd4\u8f83\u5bb9\u6613\u5b9e\u73b0\u53d8\u5c3a\u5ea6\u7684\u751f\u6210\uff0c\u8fd9\u4e9b\u5bf9\u5177\u4f53\u843d\u5730\u5e94\u7528\u90fd\u662f\u6bd4\u8f83\u91cd\u8981\u7684\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    DiT\u539f\u6587: https:\/\/arxiv.org\/abs\/2212.09748 Code:https:\/\/gi … \u7ee7\u7eed\u9605\u8bfb\u6269\u6563\u6a21\u578bDiT\uff1a\u7eafTransformer\u67b6\u6784\u66ff\u6362\u4f20\u7edf\u7684U-Net\u4e3b\u5e72<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[29,4],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15728"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=15728"}],"version-history":[{"count":11,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15728\/revisions"}],"predecessor-version":[{"id":15742,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15728\/revisions\/15742"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=15728"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=15728"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=15728"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}