{"id":1878,"date":"2026-02-25T10:11:00","date_gmt":"2026-02-25T02:11:00","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/1878"},"modified":"2026-02-25T10:11:00","modified_gmt":"2026-02-25T02:11:00","slug":"%e4%bb%8e0%e5%88%b01%e8%b7%91%e9%80%9asora%e5%90%8c%e6%ac%bedit%e8%a7%86%e9%a2%91%e7%94%9f%e6%88%90%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e4%b8%80%e9%94%ae%e9%95%9c%e5%83%8f%e6%b5%b7%e9%87%8f","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/1878","title":{"rendered":"\u4ece0\u52301\u8dd1\u901aSora\u540c\u6b3eDiT\u89c6\u9891\u751f\u6210\uff1a\u661f\u5b87\u667a\u7b97\u4e00\u952e\u955c\u50cf+\u6d77\u91cf\u6570\u636e\u793c\u5305"},"content":{"rendered":"
\"\u4ece0\u52301\u8dd1\u901aSora\u540c\u6b3eDiT\u89c6\u9891\u751f\u6210\uff1a\u661f\u5b87\u667a\u7b97\u4e00\u952e\u955c\u50cf+\u6d77\u91cf\u6570\u636e\u793c\u5305\"<\/figure>\n
\n

\u201cSora \u5c1a\u672a\u5168\u9762\u5f00\u653e\uff0c\u4f46 DiT \u67b6\u6784\u5df2\u7ecf\u8ba9\u5168\u7403\u5b9e\u9a8c\u5ba4\u5377\u51fa\u4e86 48 \u5c0f\u65f6\u4e00\u8fed\u4ee3\u7684\u2018\u89c6\u9891\u751f\u6210\u5185\u5377\u6f6e\u2019\u3002\u201d
\n\u2014\u2014The Information \u6700\u65b0\u62a5\u9053<\/p>\n<\/blockquote>\n

01 \u70ed\u70b9\uff1aDiT \u70b9\u71c3 AI \u89c6\u9891\u8d5b\u9053<\/h2>\n

\u8fc7\u53bb\u4e09\u4e2a\u6708\uff0c\u4ece Stability AI \u7684 Stable Video Diffusion \u5230\u521a\u9732\u8138\u7684 OpenAI Sora\uff0cDiffusion Transformer\uff08DiT\uff09\u67b6\u6784\u6210\u4e3a\u201c\u751f\u6210\u5f0f\u89c6\u9891\u201d\u552f\u4e00\u5173\u952e\u8bcd\u3002\u76f8\u6bd4\u4f20\u7edf U-Net\uff0cDiT \u628a\u65f6\u7a7a Patch \u5b8c\u5168\u4ea4\u7ed9 Transformer \u81ea\u6ce8\u610f\u529b\uff0c\u4e00\u53e5\u8bdd\u5c31\u80fd\u8ba9 2 \u79d2\u7247\u6bb5\u91cc\u7684\u4eba\u7269\u7728\u773c\u3001\u955c\u5934\u5e73\u79fb\u3001\u5149\u5f71\u81ea\u7136\u8fc7\u6e21\u3002\u4f46\u70ed\u95f9\u80cc\u540e\uff0c\u884c\u4e1a\u5a92\u4f53\u5708\u65e9\u5c31\u6d41\u4f20\u4e00\u53e5\u9ed1\u8bdd\uff1a\u201c512 \u5f20 A100 + \u4e00\u5468\u4e0d\u505c\u8dd1 + \u6e05\u6d17\u597d\u7684 TB \u7ea7\u9ad8\u6e05\u6570\u636e\uff0c\u624d\u6709\u8d44\u683c\u4e0a\u684c\u3002\u201d<\/p>\n

02 \u95e8\u69db\uff1a\u7b97\u529b\u3001\u6570\u636e\u3001\u5de5\u7a0b\uff0c\u4e09\u5ea7\u5927\u5c71<\/h2>\n
    \n
  • \u7b97\u529b<\/strong>\uff1a\u5355\u5361 A100 \u8bad\u7ec3 256\u00d7256 2 \u79d2\u7247\u6bb5\uff0cFP16 \u6df7\u5408\u7cbe\u5ea6\u4e5f\u8981 18 \u5c0f\u65f6\uff1b\u60f3\u8dd1 16 \u5e27\/\u79d2\u3001512\u00d7384 \u5206\u8fa8\u7387\uff0c\u5343\u5361\u8d77\u6b65\u3002 <\/li>\n
  • \u6570\u636e<\/strong>\uff1a\u516c\u5f00\u6570\u636e\u96c6 WebVid-10M \u5206\u8fa8\u7387\u666e\u904d 360p\uff0c\u6c34\u5370\u3001\u8f6c\u573a\u3001\u5b57\u5e55\u6df7\u6742\uff0c\u6e05\u6d17\u540e\u53ef\u7528\u7387\u4e0d\u8db3 30%\u3002 <\/li>\n
  • \u5de5\u7a0b<\/strong>\uff1aDiT \u5bf9\u5e27\u95f4\u4e00\u81f4\u6027\u6781\u5ea6\u654f\u611f\uff0c\u9700\u8981\u91cd\u65b0\u5199 dataloader\u3001\u6539\u65f6\u5e8f position embedding\u3001\u8c03 gradient checkpointing\uff0c\u5149\u73af\u5883\u914d\u7f6e\u5c31\u80fd\u529d\u9000 80% \u7684\u5c0f\u56e2\u961f\u3002<\/li>\n<\/ul>\n

    03 \u661f\u5b87\u667a\u7b97\u65b9\u6848\uff1a\u4e00\u952e\u955c\u50cf + 1.2TB HD \u6570\u636e\u793c\u5305<\/h2>\n

    \u661f\u5b87\u667a\u7b97\u628a\u201c\u4e09\u5ea7\u5927\u5c71\u201d\u6253\u5305\u6210\u4e00\u6761\u547d\u4ee4\u3002
    \n1. \u955c\u50cf\uff1a\u5b98\u65b9\u5185\u7f6e DiT-training-24.04<\/code> \u955c\u50cf\uff0cPyTorch 2.1 + CUDA 12.1 + xFormers 0.0.23 \u5168\u90e8\u914d\u597d\uff0cNCCL \u62d3\u6251\u9488\u5bf9 4090\/A100 \u6df7\u5408\u7ec4\u7f51\u8c03\u4f18\uff0c\u591a\u5361\u5e76\u884c\u6548\u7387 93%\u3002
    \n2. \u6570\u636e\uff1a\u5e73\u53f0\u516c\u5171\u8d44\u6e90\u5e93\u4e00\u6b21\u6027\u6302\u8f7d 1.2TB \u5f00\u6e90 HD \u89c6\u9891-\u6587\u672c\u5bf9\uff081080p\u300124fps\u3001\u5e26\u539f\u59cb\u5b57\u5e55\u6587\u4ef6\uff09\uff0c\u5df2\u8dd1\u901a\u53bb\u91cd\u3001\u955c\u5934\u5207\u5206\u3001\u7f8e\u5b66\u6253\u5206 3 \u9053\u6e05\u6d17\u6d41\u7a0b\uff0c\u53ef\u76f4\u63a5\u590d\u5236\u5230\u5b9e\u4f8b\u5185\u8bad\u7ec3\u3002
    \n3. \u542f\u52a8\uff1a\u5728\u63a7\u5236\u53f0\u9009\u62e9\u201cDiT \u89c6\u9891\u751f\u6210\u201d\u6a21\u677f \u2192 8\u00d7A100 80G \u5b9e\u4f8b \u2192 \u70b9\u51fb\u201c\u4e00\u952e\u542f\u52a8\u201d\uff0c5 \u5206\u949f\u540e\u65e5\u5fd7\u51fa\u73b0 Training epoch 0\/100<\/code>\uff0c\u5373\u4ee3\u8868\u73af\u5883\u5c31\u7eea\u3002
    \n4. \u5b58\u50a8\uff1a\u8bad\u7ec3\u4e2d\u95f4 checkpoint \u81ea\u52a8\u5199\u5165\u4e91\u786c\u76d8\uff0c\u652f\u6301\u70ed\u63d2\u62d4\u5230 16 \u5361\u5b9e\u4f8b\u7ee7\u7eed scale up\uff1b\u65e5\u5fd7\u4e0e TensorBoard \u5b9e\u65f6\u540c\u6b65\u5230\u4e91\u5b58\u50a8\uff0c\u6d4f\u89c8\u5668\u53ef\u76f4\u63a5\u67e5\u770b loss \u66f2\u7ebf\u3002<\/p>\n

    04 \u6210\u672c\uff1a1024 \u5361\u00b7\u65f6 \u2248 900 \u5143<\/h2>\n

    \u4ee5 512\u00d7256 \u5206\u8fa8\u7387\u30012 \u79d2 16 \u5e27\u7247\u6bb5\u4e3a\u4f8b\uff0c8\u00d7A100 \u5e76\u884c\u8bad\u7ec3 128 \u6b65\u5373\u53ef\u6536\u655b\u3002\u661f\u5b87\u667a\u7b97\u91c7\u7528\u201c\u6309\u5206\u949f\u8ba1\u8d39 + \u95f2\u65f6 7 \u6298\u201d\u7b56\u7565\uff1a
    \n– 8 \u5361 A100 \u5355\u4ef7 1.5 \u5143\/\u5361\/\u65f6\uff0c1024 \u5361\u00b7\u65f6\u5408\u8ba1 1536 \u5143\uff1b
    \n– \u95f2\u65f6\uff080:00-8:00\uff09\u81ea\u52a8\u89e6\u53d1\u6298\u6263\uff0c\u5b9e\u4ed8 900 \u5143\u51fa\u5934\u3002
    \n\u5bf9\u6bd4\u4e91\u5382\u5546\u6309\u9700 3.2 \u5143\/\u5361\/\u65f6\u7684\u6807\u51c6\u4ef7\uff0c\u76f4\u63a5\u780d 60%\u3002\u5982\u679c\u53ea\u60f3\u5148\u9a8c\u7b97\uff0c\u6ce8\u518c\u5c31\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u53ef\u767d\u5ad6 80 \u5361\u00b7\u65f6\uff0c\u8db3\u591f\u628a demo \u8dd1\u901a\u3002<\/p>\n

    05 Gradio Demo\uff1a\u6d4f\u89c8\u5668\u91cc\u201c\u4e00\u952e\u51fa\u7247\u201d<\/h2>\n

    \u8bad\u7ec3\u5b8c\u628a sample.mp4<\/code> \u62d6\u8fdb\u5e73\u53f0\u81ea\u5e26\u7684 Gradio \u6a21\u677f\uff0c3 \u5206\u949f\u5c31\u80fd\u642d\u4e00\u4e2a H5 \u9875\u9762\u3002\u8f93\u5165\u4e00\u53e5\u201c\u65e0\u4eba\u673a\u89c6\u89d2\u4fef\u77b0\u96ea\u540e\u4eac\u90fd\u201d\uff0c\u540e\u7aef\u81ea\u52a8\u8c03\u7528\u5df2\u8f6c\u6362\u7684 DiT-diffusers \u683c\u5f0f\u6743\u91cd\uff0c2 \u79d2 512\u00d7256 \u89c6\u9891 15 \u79d2\u751f\u6210\u5b8c\u6bd5\uff0c\u652f\u6301\u8fb9\u64ad\u8fb9\u4e0b\u8f7d\u3002Demo \u955c\u50cf\u5df2\u88c5 FFmpeg + Streamlit\uff0c\u516c\u7f51 URL \u4e00\u952e\u53ef\u8f6c\u53d1\uff0c\u62ff\u53bb\u505a\u4ea7\u54c1\u8def\u6f14\u3001\u878d\u8d44 demo \u90fd\u591f\u7528\u3002<\/p>\n

    06 \u65b0\u624b\u6307\u5357\uff1a30 \u5206\u949f\u4ece\u6ce8\u518c\u5230\u51fa\u7247<\/h2>\n
      \n
    1. \u6ce8\u518c\uff1a\u5b98\u7f51\u624b\u673a\u53f7\u9a8c\u8bc1\uff0c\u7acb\u5f97 10 \u5143\u4f53\u9a8c\u91d1\u3002 <\/li>\n
    2. \u9009\u5b9e\u4f8b\uff1aGPU \u5e02\u573a \u2192 8\u00d7A100 \u2192 \u955c\u50cf\u9009\u62e9 \u201cDiT-training-24.04\u201d\u3002 <\/li>\n
    3. \u62f7\u6570\u636e\uff1a\u5b9e\u4f8b\u5185\u6267\u884c cp -r \/public\/DiT-HD-1.2T .\/data<\/code>\u3002 <\/li>\n
    4. \u5f00\u8bad\uff1atorchrun --nproc_per_node=8 train.py --config configs\/dit_512x256.yaml<\/code>\u3002 <\/li>\n
    5. \u63a8\u7406\uff1a\u8bad\u7ec3\u65e5\u5fd7\u51fa\u73b0 saved checkpoint at step 128<\/code> \u540e\uff0c\u8fd0\u884c python gradio_app.py --ckpt .\/checkpoints\/dit_512x256.bin<\/code>\u3002 <\/li>\n
    6. \u5206\u4eab\uff1a\u628a Gradio \u516c\u7f51\u94fe\u63a5\u7529\u5230\u7fa4\u91cc\uff0c\u6536\u83b7\u201c\u54c7\u201d\u58f0\u4e00\u7247\u3002<\/li>\n<\/ol>\n

      07 \u5199\u5728\u6700\u540e<\/h2>\n

      \u5f53\u89c6\u9891\u751f\u6210\u8fdb\u5165\u201cTransformer \u65f6\u4ee3\u201d\uff0c\u6a21\u578b\u521b\u65b0\u53ea\u5360\u5230 20% \u7684\u80dc\u7387\uff0c\u5269\u4e0b 80% \u62fc\u7684\u662f\u5de5\u7a0b\u843d\u5730\u4e0e\u8d44\u6e90\u8c03\u914d\u3002\u661f\u5b87\u667a\u7b97\u628a\u7b97\u529b\u3001\u6570\u636e\u3001\u955c\u50cf\u3001 Demo \u505a\u6210\u4e00\u6761\u201c\u6d41\u6c34\u7ebf\u201d\uff0c\u8ba9\u7814\u7a76\u5458\u56de\u5230\u7b97\u6cd5\u672c\u8eab\uff0c\u8ba9\u521b\u4e1a\u8005\u7528 900 \u5143\u5c31\u80fd\u9a8c\u8bc1 PMF\u3002
      \n\u73b0\u5728\u6ce8\u518c\uff0c10 \u5143\u4f53\u9a8c\u91d1\u5df2\u5165\u8d26\uff0c\u4e0b\u4e00\u6761\u5237\u5c4f\u7684 AI \u89c6\u9891\uff0c\u4e5f\u8bb8\u5c31\u51fa\u81ea\u4f60\u7684\u6d4f\u89c8\u5668\u6807\u7b7e\u9875\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

      \u201cSora \u5c1a\u672a\u5168\u9762\u5f00\u653e\uff0c\u4f46 DiT \u67b6\u6784\u5df2\u7ecf\u8ba9\u5168\u7403\u5b9e\u9a8c\u5ba4\u5377…<\/p>\n","protected":false},"author":2,"featured_media":1876,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1878","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":77,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1878","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=1878"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1878\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/1876"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=1878"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=1878"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=1878"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}