{"id":4389,"date":"2025-11-17T20:08:11","date_gmt":"2025-11-17T12:08:11","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4389"},"modified":"2025-11-17T20:08:11","modified_gmt":"2025-11-17T12:08:11","slug":"202511%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bbblip-2-bootstrapping-language-image-pre-training-with-frozen-image-encoders-and-large-language-models","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202511%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bbblip-2-bootstrapping-language-image-pre-training-with-frozen-image-encoders-and-large-language-models\/","title":{"rendered":"202511\u8bba\u6587\u7814\u8bfbBLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models"},"content":{"rendered":"\n

2025\u5e7411\u670817\u65e5-gongchenhao<\/p>\n\n\n\n

\u6765\u6e90\uff1aICML<\/p>\n\n\n\n

\u4f5c\u8005\uff1aJunnan Li<\/p>\n\n\n\n

\u5355\u4f4d\uff1aSalesforce AI Research<\/p>\n\n\n\n

\u53d1\u8868\u65f6\u95f4\uff1a2023<\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/h2>\n\n\n\n

\u89c6\u89c9-\u8bed\u8a00\u9884\u8bad\u7ec3\uff08VLP\uff09\u7684\u5feb\u901f\u53d1\u5c55\u4f46\u6210\u672c\u6781\u9ad8
\u8fd1\u5e74\u6765\uff0c\u5927\u89c4\u6a21\u56fe\u6587\u6a21\u578b\uff08\u5982 CLIP\u3001BLIP\u3001SimVLM\u3001CoCa\uff09\u4e0d\u65ad\u5237\u65b0\u591a\u6a21\u6001\u4efb\u52a1\u6548\u679c\uff0c\u4f46\u5927\u591a\u4f9d\u8d56\uff1a
\u8d85\u5927\u89c4\u6a21\u6570\u636e\uff08\u6570\u4ebf\u7ea7\uff09
\u7aef\u5230\u7aef\u8bad\u7ec3\u89c6\u89c9\u6a21\u578b + \u6587\u672c\u6a21\u578b
\u5de8\u989d\u8bad\u7ec3\u6210\u672c\uff08\u8ba1\u7b97\u8d44\u6e90\u6602\u8d35<\/p>\n\n\n\n

\u5927\u6a21\u578b\u51fa\u73b0\u201c\u89c6\u89c9\u80fd\u529b\u7f3a\u5931\u201d\u95ee\u9898
LLM\uff08GPT\u3001OPT\u3001Flan-T5 \u7b49\uff09\u5177\u5907\u5f3a\u5927\u7684\u8bed\u8a00\u7406\u89e3\u4e0e\u751f\u6210\u80fd\u529b\uff0c\u4f46\uff1a
\u5b83\u4eec\u5728\u9884\u8bad\u7ec3\u65f6\u6ca1\u6709\u89c1\u8fc7\u56fe\u50cf
\u65e0\u6cd5\u76f4\u63a5\u8fdb\u884c\u89c6\u89c9\u7406\u89e3\u6216\u89c6\u89c9\u63a8\u7406
\u56e0\u6b64\u5982\u679c\u5f3a\u884c\u7aef\u5230\u7aef\u8bad\u7ec3\uff0c\u4f1a\u51fa\u73b0\uff1a
\u6a21\u6001\u5bf9\u9f50\u56f0\u96be
\u707e\u96be\u6027\u9057\u5fd8\uff08forgetting\uff09
\u8bad\u7ec3\u6781\u5176\u6602\u8d35<\/p>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u8d21\u732e<\/h2>\n\n\n\n
    \n
  1. \u63d0\u51fa Q-Former\uff1a\u8f7b\u91cf\u9ad8\u6548\u7684\u89c6\u89c9\u2014\u8bed\u8a00\u6865\u63a5\u6a21\u5757
    \u5229\u7528\u4e00\u7ec4\u53ef\u5b66\u4e60\u67e5\u8be2\u5411\u91cf\u4e3b\u52a8\u4ece\u51bb\u7ed3\u89c6\u89c9\u6a21\u578b\u4e2d\u63d0\u53d6\u8bed\u8a00\u76f8\u5173\u89c6\u89c9\u7279\u5f81\uff0c\u6784\u5efa\u201c\u4fe1\u606f\u74f6\u9888\u5f0f\u201d\u8de8\u6a21\u6001\u63a5\u53e3\u3002<\/li>\n\n\n\n
  2. \u521b\u65b0\u6027\u4e24\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565\uff0c\u89e3\u51b3\u51bb\u7ed3 LLM \u7684\u6a21\u6001\u5bf9\u9f50\u96be\u9898
    \u901a\u8fc7\u201c\u8868\u793a\u5b66\u4e60\uff08ITC\/ITG\/ITM\uff09+ \u751f\u6210\u5b66\u4e60\uff08LLM \u63a5\u5165\uff09\u201d\u9010\u6b65\u5bf9\u9f50\u89c6\u89c9\u4e0e\u8bed\u8a00\u7a7a\u95f4\uff0c\u5728\u4e0d\u66f4\u65b0 LLM \u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u7a33\u5b9a\u8de8\u6a21\u6001\u80fd\u529b\u3002<\/li>\n\n\n\n
  3. \u5229\u7528\u51bb\u7ed3\u89c6\u89c9\u6a21\u578b + \u51bb\u7ed3 LLM\uff0c\u5b9e\u73b0\u6781\u4f4e\u8bad\u7ec3\u6210\u672c\u7684\u9ad8\u6027\u80fd\u591a\u6a21\u6001\u6a21\u578b
    \u4ec5\u8bad\u7ec3 Q-Former \u5373\u53ef\u8fbe\u5230\u751a\u81f3\u8d85\u8fc7 Flamingo \u7b49\u5927\u89c4\u6a21\u7aef\u5230\u7aef\u6a21\u578b\u7684\u6548\u679c\uff0c\u8bad\u7ec3\u53c2\u6570\u91cf\u7f29\u5c0f 50\u00d7 \u4ee5\u4e0a\u3002<\/li>\n\n\n\n
  4. \u5efa\u7acb\u901a\u7528\u53ef\u590d\u7528\u7684\u591a\u6a21\u6001\u6846\u67b6
    \u89c6\u89c9\u6a21\u578b\u3001\u8bed\u8a00\u6a21\u578b\u5747\u53ef\u81ea\u7531\u66ff\u6362\uff0c\u5f62\u6210\u7075\u6d3b\u7684\u201c\u5373\u63d2\u5373\u7528\u5f0f\u201d\u591a\u6a21\u6001\u7cfb\u7edf\u67b6\u6784\u3002<\/li>\n\n\n\n
  5. \u5c55\u793a\u5f3a\u5927\u7684\u96f6\u6837\u672c\u89c6\u89c9\u4efb\u52a1\u80fd\u529b
    \u65e0\u9700\u5fae\u8c03\u5373\u53ef\u5b8c\u6210 VQA\u3001\u56fe\u50cf\u63cf\u8ff0\u3001\u89c6\u89c9\u5bf9\u8bdd\u3001\u5e38\u8bc6\u63a8\u7406\u7b49\u4efb\u52a1\uff0c\u8868\u660e\u6a21\u578b\u5177\u5907\u901a\u7528\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e09\u3001\u65b9\u6cd5<\/h2>\n\n\n\n
      \n
    1. \u6a21\u578b\u67b6\u6784\uff1aQ-Former\uff08Querying Transformer\uff09
      Q-Former \u662f BLIP-2 \u7684\u552f\u4e00\u53ef\u8bad\u7ec3\u6a21\u5757\uff0c\u7531\u4ee5\u4e0b\u7ec4\u6210\uff1a
      \u4e00\u7ec4 learnable queries\uff08\u53ef\u5b66\u4e60\u67e5\u8be2\u5411\u91cf\uff09
      \u591a\u5c42 Transformer
      \u4e0e\u89c6\u89c9\u6a21\u578b\u7684 \u8de8\u6ce8\u610f\u529b\uff08cross-attention\uff09\u5c42
      \u53ef\u4f5c\u4e3a\u7f16\u7801\u5668\/\u89e3\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u7ed3\u6784\uff08\u901a\u8fc7\u4e0d\u540c mask \u63a7\u5236\uff09
      \u5176\u4f5c\u7528\u662f\uff1a
      \u4e3b\u52a8\u4ece\u51bb\u7ed3\u89c6\u89c9\u6a21\u578b\u62bd\u53d6\u9ad8\u5ea6\u8bed\u8a00\u76f8\u5173\u7684\u89c6\u89c9\u8bed\u4e49\uff0c\u5e76\u751f\u6210\u4e00\u7ec4\u7d27\u51d1\u7684\u89c6\u89c9\u7279\u5f81\uff0832\u00d7768\uff09\u3002<\/li>\n\n\n\n
    2. \u7b2c\u4e00\u9636\u6bb5\uff1a\u89c6\u89c9\u2014\u8bed\u8a00\u8868\u793a\u5b66\u4e60\uff08Representation Learning\uff09
      \u76ee\u6807\uff1a
      \u8ba9 Q-Former \u5b66\u4f1a\u62bd\u53d6\u6587\u672c\u76f8\u5173\u7684\u89c6\u89c9\u7279\u5f81\u3002
      \u4f7f\u7528\u4e09\u4e2a\u4efb\u52a1\u8054\u5408\u8bad\u7ec3\uff08\u90fd\u4e0d\u66f4\u65b0\u89c6\u89c9\u6a21\u578b\uff09\uff1a
      \u2460 \u56fe\u6587\u5bf9\u6bd4\u5b66\u4e60\uff08ITC\uff09
      \u5bf9\u9f50\u56fe\u50cf\u4e0e\u6587\u672c\u7684\u5168\u5c40\u8bed\u4e49
      queries \u4e0e\u6587\u672c\u7684 [CLS] \u8fdb\u884c\u5bf9\u6bd4
      \u2461 \u56fe\u50cf\u5f15\u5bfc\u6587\u672c\u751f\u6210\uff08ITG\uff09
      \u5f3a\u8feb queries \u63d0\u53d6\u80fd\u652f\u6491\u6587\u672c\u751f\u6210\u7684\u4fe1\u606f
      \u91c7\u7528\u591a\u6a21\u6001 causal mask\uff0c\u4f7f\u6587\u672c token \u80fd\u770b\u89c1 queries
      \u2462 \u56fe\u6587\u5339\u914d\uff08ITM\uff09
      \u5b66\u4e60\u7ec6\u7c92\u5ea6\u8de8\u6a21\u6001\u5bf9\u9f50
      \u4f7f\u7528\u53cc\u5411\u6ce8\u610f\u529b mask\uff0c\u8ba9 queries \u4e0e\u6587\u672c\u6df1\u5ea6\u4ea4\u4e92
      \u4e09\u79cd\u635f\u5931\u5171\u540c\u4fc3\u4f7f queries \u5b66\u4f1a\u201c\u63d0\u53d6\u5bf9\u8bed\u8a00\u6700\u91cd\u8981\u7684\u89c6\u89c9\u4fe1\u606f\u201d\u3002<\/li>\n\n\n\n
    3. \u7b2c\u4e8c\u9636\u6bb5\uff1a\u89c6\u89c9 \u2192 \u8bed\u8a00\u751f\u6210\u5b66\u4e60\uff08Vision-to-Language Generative Learning\uff09
      \u76ee\u6807\uff1a
      \u8ba9 LLM \u80fd\u7406\u89e3 Q-Former \u7684\u89c6\u89c9\u8868\u793a\uff0c\u4f46\u4fdd\u6301 LLM \u5b8c\u5168\u51bb\u7ed3\u3002
      \u505a\u6cd5\uff1a
      \u5c06 Q-Former \u8f93\u51fa\u7684 queries \u7ebf\u6027\u6295\u5f71\u5230 LLM \u7684 embedding \u7ef4\u5ea6
      \u4f5c\u4e3a soft visual prompts \u62fc\u63a5\u5230\u8bed\u8a00\u8f93\u5165\u524d
      \u4f7f\u7528 LLM \u7684\u8bed\u8a00\u5efa\u6a21\u635f\u5931\u8bad\u7ec3 Q-Former\uff08LLM \u4e0d\u66f4\u65b0\uff09
      \u4e24\u79cd LLM \u65b9\u5f0f\uff1a
      Decoder-only\uff08OPT\uff09\uff1a\u76f4\u63a5\u505a LM loss
      Encoder-Decoder\uff08FlanT5\uff09\uff1aPrefix-LM loss\uff08\u524d\u7f00\u8f93\u5165\uff0c\u540e\u7f00\u9884\u6d4b\uff09
      \u6548\u679c\uff1a
      LLM \u83b7\u5f97\u89c6\u89c9\u7406\u89e3\u80fd\u529b\uff0c\u800c\u65e0\u9700\u4efb\u4f55\u53c2\u6570\u66f4\u65b0\uff0c\u4e5f\u907f\u514d\u707e\u96be\u6027\u9057\u5fd8\u3002<\/li>\n\n\n\n
    4. \u8bad\u7ec3\u7b56\u7565\u4e0e\u6574\u4f53\u6846\u67b6
      \u5168\u6d41\u7a0b\u4e2d\uff0c\u89c6\u89c9\u6a21\u578b\u548c LLM \u5747\u51bb\u7ed3
      \u4ec5\u8bad\u7ec3 Q-Former\uff08\u7ea6 188M \u53c2\u6570\uff09
      \u4e24\u9636\u6bb5\u8bad\u7ec3\u5206\u522b\u8fdb\u884c\uff1a
      250k \u6b65\uff08stage1\uff09
      80k \u6b65\uff08stage2\uff09<\/li>\n<\/ol>\n\n\n\n

      \u56db\u3001\u5b9e\u9a8c<\/h2>\n\n\n\n

      BLIP-2 \u5728\u591a\u4e2a\u89c6\u89c9\u8bed\u8a00\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u5168\u9762\u5b9e\u9a8c\uff0c\u9a8c\u8bc1\u5176\u5728\u51bb\u7ed3\u89c6\u89c9\u6a21\u578b + \u51bb\u7ed3 LLM + \u4ec5\u8bad\u7ec3 Q-Former\u7684\u8bbe\u5b9a\u4e0b\uff0c\u4ecd\u80fd\u53d6\u5f97\u5f3a\u4e8e\u73b0\u6709\u5927\u578b\u7aef\u5230\u7aef\u6a21\u578b\u7684\u6548\u679c\u3002<\/p>\n\n\n\n

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

      \u4e94\u3001\u603b\u7ed3<\/h2>\n\n\n\n

      \u4e00\u3001\u6838\u5fc3\u8303\u5f0f\uff1a\u5229\u7528\u4e0e\u6865\u63a5
      BLIP-2\u786e\u7acb\u4e86\u4e00\u79cd\u65b0\u8303\u5f0f\uff1a\u76f4\u63a5\u5229\u7528\u51bb\u7ed3\u7684\u3001\u6210\u719f\u7684\u5355\u6a21\u6001\u6a21\u578b\uff08\u5982CLIP\u3001OPT\uff09\uff0c\u5e76\u901a\u8fc7\u4e00\u4e2a\u8f7b\u91cf\u7684 Q-Former \u4f5c\u4e3a\u6865\u6881\u6765\u8fde\u63a5\u5b83\u4eec\u3002Q-Former\u5145\u5f53\u201c\u4fe1\u606f\u74f6\u9888\u201d\uff0c\u8d1f\u8d23\u5c06\u89c6\u89c9\u4fe1\u606f\u9ad8\u6548\u5730\u201c\u7ffb\u8bd1\u201d\u6210\u8bed\u8a00\u6a21\u578b\u80fd\u7406\u89e3\u7684\u201c\u89c6\u89c9\u63d0\u793a\u201d\u3002
      \u4e8c\u3001\u6838\u5fc3\u8d21\u732e\u4e0e\u6210\u679c
      \u9ad8\u6548\u5353\u8d8a\uff1a\u4ee5\u6781\u5c11\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\uff081.88\u4ebf\uff09\uff0c\u5728\u591a\u9879\u4efb\u52a1\u4e0a\u8fbe\u5230\u6700\u5148\u8fdb\u6027\u80fd\uff0c\u5b9e\u73b0\u4e86\u6548\u7387\u4e0e\u6548\u679c\u7684\u5353\u8d8a\u5e73\u8861\u3002
      \u6307\u4ee4\u8ddf\u968f\uff1a\u89e3\u9501\u4e86\u96f6\u6837\u672c\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\uff0c\u80fd\u6839\u636e\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u8fdb\u884c\u56fe\u50cf\u63cf\u8ff0\u4e0e\u5bf9\u8bdd\uff0c\u5c55\u73b0\u4e86\u901a\u5411\u591a\u6a21\u6001\u5bf9\u8bddAI\u7684\u6f5c\u529b\u3002
      \u901a\u7528\u53ef\u6269\u5c55\uff1a\u4f5c\u4e3a\u4e00\u4e2a\u901a\u7528\u6846\u67b6\uff0c\u53ef\u201c\u5373\u63d2\u5373\u7528\u201d\u5730\u517c\u5bb9\u66f4\u5f3a\u7684\u89c6\u89c9\u6216\u8bed\u8a00\u6a21\u578b\uff0c\u5177\u5907\u957f\u671f\u751f\u547d\u529b\u3002<\/p>\n\n\n\n

      \u516d\u3001\u601d\u8003<\/h2>\n\n\n\n

      1.0 \u6280\u672f\u521b\u65b0\u2014\u2014\u903b\u8f91\u601d\u7ef4\u63a8\u7406\u6846\u67b6<\/p>\n\n\n\n

      BLIP-2 \u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u63d0\u51fa Q-Former \u5e76\u91c7\u7528\u201c\u4e24\u9636\u6bb5\u8de8\u6a21\u6001\u5bf9\u9f50\u201d\u6846\u67b6\uff1a<\/p>\n\n\n\n

      \u8f7b\u91cf\u7ea7 Q-Former \u4f5c\u4e3a\u89c6\u89c9 \u2192 \u8bed\u8a00\u7684\u4fe1\u606f\u74f6\u9888\uff0c\u5b9e\u73b0\u4e3b\u52a8\u9009\u62e9\u9ad8\u4ef7\u503c\u89c6\u89c9\u8bed\u4e49\u3002<\/p>\n\n\n\n

      \u51bb\u7ed3\u89c6\u89c9\u6a21\u578b + \u51bb\u7ed3\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u53ea\u8bad\u7ec3 Q-Former\uff0c\u5927\u5e45\u964d\u4f4e\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n\n\n\n

      2.0 \u6280\u672f\u76ee\u6807\u2014\u2014\u8de8\u57df\u77e5\u8bc6\u7ed3\u6784\u5bf9\u6bd4<\/p>\n\n\n\n

      BLIP-2 \u7684\u6280\u672f\u76ee\u6807\u662f\uff1a<\/p>\n\n\n\n

      \u8ba9\u89c6\u89c9\u6a21\u578b\u8d1f\u8d23\u611f\u77e5\u3001LLM \u8d1f\u8d23\u63a8\u7406\u4e0e\u751f\u6210\uff0c<\/p>\n\n\n\n

      \u4e2d\u95f4\u7528 Q-Former \u5b9e\u73b0\u8de8\u57df\u77e5\u8bc6\u5bf9\u9f50\u4e0e\u538b\u7f29\uff0c<\/p>\n\n\n\n

      \u8fbe\u5230\u7528\u6781\u5c11\u53c2\u6570\u83b7\u5f97\u63a5\u8fd1\u7aef\u5230\u7aef\u5927\u6a21\u578b\u7684\u6548\u679c\u3002<\/p>\n\n\n\n

      3.0\u573a\u666f\u529f\u80fd\u2014\u56fe\u6587\u7f16\u8f91\u589e\u5f3a<\/p>\n\n\n\n

      \u5229\u7528BLIP-2\u7684\u96f6\u6837\u672c\u56fe\u50cf\u63cf\u8ff0\u4e0e\u89c6\u89c9\u95ee\u7b54\u80fd\u529b<\/p>\n\n\n\n

      \u81ea\u52a8\u68c0\u6d4b\u56fe\u7247\u5185\u5bb9\u4e0e\u6587\u6848\u662f\u5426\u4e00\u81f4<\/p>\n\n\n\n

      \u8bc6\u522b\u56fe\u7247\u4e2d\u672a\u63d0\u53ca\u3001\u63cf\u8ff0\u9519\u8bef\u6216\u9057\u6f0f\u7684\u4fe1\u606f<\/p>\n\n\n\n

      \u751f\u6210\u5dee\u5f02\u62a5\u544a\uff0c\u8f85\u52a9\u8fd0\u8425\u3001\u5a92\u4f53\u7f16\u8f91\u63d0\u9ad8\u5185\u5bb9\u51c6\u786e\u5ea6<\/p>\n","protected":false},"excerpt":{"rendered":"

      2025\u5e7411\u670817\u65e5-gongchenhao \u6765\u6e90\uff1aICML \u4f5c\u8005\uff1aJunnan Li \u5355\u4f4d\uff1aSalesfo […]<\/p>\n","protected":false},"author":73,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4389","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4389","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/73"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4389"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4389\/revisions"}],"predecessor-version":[{"id":4435,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4389\/revisions\/4435"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4389"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4389"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4389"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}