
{"id":9588,"date":"2026-06-25T03:11:05","date_gmt":"2026-06-24T19:11:05","guid":{"rendered":"https:\/\/infernews.com\/blog\/unlimited-ocr-works-welcome-the-era-of-one-shot-long-horizon-parsing\/"},"modified":"2026-06-25T03:18:53","modified_gmt":"2026-06-24T19:18:53","slug":"unlimited-ocr-works-welcome-the-era-of-one-shot-long-horizon-parsing","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/unlimited-ocr-works-welcome-the-era-of-one-shot-long-horizon-parsing\/","title":{"rendered":"Unlimited-OCR\uff1a\u9577\u6587\u4ef6 OCR \u65b0\u53d6\u5411"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/Unlimited-OCR.png\" alt=\"Baidu Inc.\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Unlimited-OCR \u662f\u4e00\u500b OCR \u8996\u89ba\u6587\u5b57\u8fa8\u8b58\u6a21\u578b\u9805\u76ee\uff0c\u4e5f\u53ef\u8996\u70ba\u4e00\u500b\u91dd\u5c0d\u9577\u6587\u4ef6\u89e3\u6790\u800c\u6539\u9020\u7684\u7814\u7a76\u539f\u578b\u3002\u5b83\u4e3b\u8981\u7528\u4f86\u628a\u5716\u7247\u6216 PDF \u5167\u7684\u5927\u91cf\u6587\u5b57\u8207\u7248\u9762\u5167\u5bb9\u4e00\u6b21\u904e\u8f49\u6210\u53ef\u8f38\u51fa\u7684\u89e3\u6790\u7d50\u679c\uff0c\u91cd\u9ede\u662f\u8655\u7406\u591a\u9801\u6587\u4ef6\u6642\u76e1\u91cf\u6e1b\u5c11\u8a18\u61b6\u9ad4\u8ca0\u64d4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709 end-to-end OCR \u505a\u6cd5\u4ee5 DeepSeek-OCR \u70ba\u4ee3\u8868\uff0c\u6703\u7528 large language model\uff08LLM\uff09\u4f5c decoder\uff0c\u512a\u9ede\u662f\u80fd\u501f\u52a9\u8a9e\u8a00\u5148\u9a57\u63d0\u5347\u8fa8\u8b58\u6548\u679c\uff0c\u4f46\u8f38\u51fa\u4e00\u9577\uff0cKV cache \u6703\u4e00\u8def\u7d2f\u7a4d\uff0c\u4ee4\u986f\u5b58\u9700\u6c42\u4e0a\u5347\u3001\u751f\u6210\u6108\u4f86\u6108\u6162\u3002Unlimited-OCR \u7684\u505a\u6cd5\u662f\u4fdd\u7559\u9ad8\u58d3\u7e2e encoder\uff0c\u518d\u628a decoder \u7684 attention \u5c64\u6539\u6210 Reference Sliding Window Attention\uff08R-SWA\uff09\uff0c\u8b93\u6bcf\u500b token \u6301\u7e8c\u95dc\u6ce8 reference tokens \u8207\u6709\u9650\u9577\u5ea6\u7684\u524d\u6587\uff0c\u76ee\u6a19\u662f\u628a KV cache \u7dad\u6301\u5728\u5e38\u6578\u898f\u6a21\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u53d6\u5411\u6700\u503c\u5f97\u7559\u610f\u7684\u5730\u65b9\uff0c\u4e0d\u662f\u55ae\u7d14\u8ffd\u6c42\u55ae\u9801\u6700\u9ad8\u7cbe\u5ea6\uff0c\u800c\u662f\u628a\u300cone-shot long-horizon parsing\u300d\u653e\u5728\u6838\u5fc3\u4f4d\u7f6e\u3002\u8ddf\u4e00\u822c full attention \u6bd4\uff0c\u5b83\u72a7\u7272\u7684\u662f\u50b3\u7d71\u5168\u57df\u6ce8\u610f\u529b\u5f62\u5f0f\uff0c\u63db\u4f86\u591a\u9801\u6587\u4ef6\u5728 32K \u9577\u5ea6\u4e0b\u4ecd\u53ef\u505a\u55ae\u6b21 forward pass\uff1b\u8ddf vanilla SWA \u6bd4\uff0c\u5b83\u53c8\u4fdd\u7559 visual tokens \u4f5c\u70ba\u7a69\u5b9a\u53c3\u7167\uff0c\u907f\u514d\u72c0\u614b\u50b3\u905e\u5f8c\u6108\u4f86\u6108\u6a21\u7cca\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u8def\u7dda\u76f8\u7576\u660e\u78ba\uff1a\u9805\u76ee\u63d0\u4f9b Hugging Face Transformers \u63a8\u7406\u65b9\u5f0f\uff0c\u6e2c\u8a66\u74b0\u5883\u5beb\u660e\u9700 NVIDIA GPU\uff0c\u4e26\u4ee5 Python 3.12.3\u3001CUDA 12.9 \u70ba\u57fa\u790e\uff1b\u55ae\u5f35\u5716\u7247\u53ef\u5728 gundam \u8207 base \u5169\u7a2e\u8a2d\u5b9a\u4e2d\u9078\u64c7\uff0c\u591a\u9801\u8207 PDF \u5247\u4f7f\u7528 base \u914d\u7f6e\u3002\u60f3\u5148\u4e86\u89e3\u6548\u679c\uff0c\u4e5f\u53ef\u76f4\u63a5\u770b Hugging Face Spaces demo \u6216 ModelScope \u7248\u672c\uff0c\u518d\u6c7a\u5b9a\u662f\u5426\u81ea\u884c\u843d\u5730\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u985e\u578b\u5b9a\u4f4d<\/strong>\uff1aOCR \u6a21\u578b\uff0f\u7814\u7a76\u539f\u578b\uff0c\u89e3\u6c7a\u9577\u6587\u4ef6\u3001\u591a\u9801\u89e3\u6790\u6642\u8a18\u61b6\u9ad4\u8207\u901f\u5ea6\u60e1\u5316\u554f\u984c<\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u5dee\u7570<\/strong>\uff1a\u4ee5 Reference Sliding Window Attention\uff08R-SWA\uff09\u53d6\u4ee3 decoder \u5168\u90e8 attention layers<\/li>\n\n\n\n<li><strong>\u9069\u5408\u60c5\u5883<\/strong>\uff1a\u9577 PDF\u3001\u6279\u91cf\u6587\u4ef6\u6578\u78bc\u5316\u3001\u9700\u8981\u7248\u9762\u89e3\u6790\u8207\u9577\u8f38\u51fa\u7684\u5718\u968a<\/li>\n\n\n\n<li><strong>\u76f8\u95dc\u6a21\u578b<\/strong>\uff1aDeepSeek-OCR\u3001Unlimited-OCR\uff1b\u6587\u4e2d\u4ea6\u63d0\u5230 R-SWA \u53ef\u5ef6\u4f38\u5230 ASR\u3001translation<\/li>\n\n\n\n<li><strong>\u9650\u5236\u5224\u65b7<\/strong>\uff1a\u76ee\u524d\u516c\u958b\u8cc7\u8a0a\u4e3b\u529b\u653e\u5728\u63a8\u7406\u8207\u65b9\u6cd5\u8a2d\u8a08\uff0c\u5177\u9ad4\u8a55\u6e2c\u6578\u5b57\u4ecd\u8981\u56de\u770b arXiv \u8ad6\u6587\u539f\u6587\u624d\u9069\u5408\u4f5c\u66f4\u7d30\u6bd4\u8f03<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u9700\u8981\u8655\u7406\u4fdd\u55ae\u3001\u5831\u8868\u3001\u6383\u63cf\u6a94\u3001\u66f8\u7c4d\u6216\u591a\u9801\u884c\u653f\u6587\u4ef6\u7684\u5718\u968a\uff0c\u9019\u500b\u9805\u76ee\u7684\u5438\u5f15\u529b\u6703\u6bd4\u4e00\u822c\u55ae\u9801 OCR \u66f4\u9ad8\u3002\u82e5\u4f60\u7684\u5de5\u4f5c\u91cd\u9ede\u662f\u77ed\u6587\u5b57\u622a\u5716\u3001\u624b\u6a5f\u5feb\u62cd\u8fa8\u8b58\uff0cUnlimited-OCR \u7684\u512a\u52e2\u672a\u5fc5\u5b8c\u5168\u767c\u63ee\uff0c\u4f46\u5c0d\u9577\u8f38\u51fa\u7a69\u5b9a\u6027\u8207\u90e8\u7f72\u5728 GPU \u74b0\u5883\u7684\u53ef\u884c\u6027\uff0c\u5b83\u5c55\u793a\u4e86\u4e00\u689d\u5f88\u6e05\u695a\u7684\u6539\u826f\u8def\u7dda\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/baidu\/Unlimited-OCR\" rel=\"noopener noreferrer\">https:\/\/github.com\/baidu\/Unlimited-OCR<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.23050\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.23050<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9019\u662f\u4e00\u500b\u4e3b\u6253\u9577\u7bc7\u6587\u4ef6\u89e3\u6790\u7684 OCR \u6a21\u578b\u9805\u76ee\u3002\u5b83\u91dd\u5c0d\u591a\u9801\u3001\u9577\u8f38\u51fa\u5167\u5bb9\u6642\u901f\u5ea6\u8207\u8a18\u61b6\u9ad4\u58d3\u529b\u8b8a\u5927\u7684\u554f\u984c\u4f5c\u51fa\u8abf\u6574\u3002<\/p>\n","protected":false},"author":8,"featured_media":9587,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,179,147,30,152,105,76,149,188,202],"tags":[],"class_list":["post-9588","post","type-post","status-publish","format-standard","hentry","category-133","category-nvidia","category-deepseek","category-image","category-python","category-python-nlp","category-76","category-149","category-meta","category-202"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9588","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=9588"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9588\/revisions"}],"predecessor-version":[{"id":9590,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9588\/revisions\/9590"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9587"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9588"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9588"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9588"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}