
{"id":12525,"date":"2026-09-22T22:41:54","date_gmt":"2026-09-22T14:41:54","guid":{"rendered":"https:\/\/infernews.com\/blog\/yue2-tokenizer\/"},"modified":"2026-09-22T22:42:55","modified_gmt":"2026-09-22T14:42:55","slug":"yue2-tokenizer","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/yue2-tokenizer\/","title":{"rendered":"YuE2 \u88dc\u4e0a\u771f\u5be6\u97f3\u8a0a Tokenizer\uff0c\u652f\u63f4\u6b4c\u66f2\u7ffb\u5531\u5fae\u8abf"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/pasted-973dab7d1598.jpg\" alt=\"Og image\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7576\u4f60\u9935\u7d66 YuE \u4e00\u6bb5\u53c3\u8003\u97f3\u8a0a\uff08\u4f8b\u5982\u67d0\u6b4c\u624b\u7684\u6f14\u5531\u7247\u6bb5\u6216\u7279\u5b9a\u6a02\u5668\u98a8\u683c\uff0c\u8981\u6c42\u63a5\u7e8c\u751f\u6210\uff09\u6642\uff1a\u5fc5\u9808\u5148\u7528 Tokenizer \u5c07\u53c3\u8003\u97f3\u8a0a\u300c\u58d3\u6210\u4ee3\u78bc\u300d\u50b3\u5165\u6a21\u578b\u3002\u82e5 Tokenizer \u54c1\u8cea\u4e0d\u4f73\uff0c\u97f3\u8272\u5728\u7b2c\u4e00\u6b65\u58d3\u7e2e\u6642\u5c31\u5931\u771f\u4e86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b v4 \u5de5\u5177\u80fd\u66f4\u9ad8\u4fdd\u771f\u5730\u5c07\u8f38\u5165\u8072\u97f3\u8f49\u70ba Token\uff0c\u8b93\u5f8c\u7e8c\u6a21\u578b\u6355\u6349\u5230\u7684\u97f3\u8272\u7279\u5fb5\u66f4\u8cbc\u8fd1\u539f\u97f3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9304\u97f3\u8f49\u63db\u3001\u6b4c\u624b\u98a8\u683c\u5fae\u8abf\u548c\u6b4c\u66f2\u7ffb\u5531\uff0c\u539f\u672c\u53d7\u9650\u65bc YuE2 \u6c92\u6709\u96a8\u9644\u7684 audio \u2192 semantic-token encoder\u3002\u9019\u500b\u9805\u76ee\u57fa\u65bc m-a-p\/YuE2-3B \u53ca m-a-p\/MERT-v2-FullSong\uff0c\u52a0\u5165\u97f3\u8a0a Tokenizer \u8207 NAR-branch LoRA\uff0c\u8b93 YuE2-3B \u53ef\u4ee5\u8655\u7406\u81ea\u5bb6\u9304\u97f3\uff0c\u4e26\u4ee5\u771f\u5be6\u88fd\u4f5c\u97f3\u8a0a\u8abf\u6574\u89e3\u78bc\u6548\u679c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tokenizer \u53d6\u7528 MERT-v2-FullSong \u7b2c 20 \u5c64\u7279\u5fb5\uff0c\u5148\u6309\u97f3\u8ecc\u4f5c instance normalization\uff0c\u518d\u4ee5 8 \u5c64\u3001d=512 \u7684 Transformer\uff0c\u5728 512-frame \u8996\u7a97\u5167\u8f49\u63db\u6210 32,768 \u7a2e YuE2 semantic codes\uff0c\u7279\u5fb5\u7387\u70ba 25 Hz\u3002\u7559\u51fa\u6b4c\u66f2\u6e2c\u8a66\u7684 top-1 exact match \u70ba 16.1%\uff0c\u4f46\u8fd1\u4f3c code \u7522\u751f\u7684\u8072\u97f3\u63a5\u8fd1\uff0cNAR round-trip \u807d\u611f\u6e2c\u8a66\u7d04\u9054 95%\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>tokenizer_head_joint_v4.pt<\/code>\u3001<code>.safetensors<\/code>\u3001<code>.bf16.safetensors<\/code>\uff1a\u97f3\u8a0a\u7de8\u78bc head<\/li>\n<li><code>nar_lora_joint_v4.pt<\/code>\u3001<code>.safetensors<\/code>\u3001<code>.bf16.safetensors<\/code>\uff1aNAR \u89e3\u78bc\u5206\u652f LoRA<\/li>\n<li><code>_comfyui.safetensors<\/code>\uff1a\u914d\u5408 ComfyUI \u7684 YuE2 \u6b0a\u91cd\u5e03\u5c40<\/li>\n<li>LoRA rank \u70ba 32\uff0c\u8986\u84cb 28 \u5c64 NAR attention\u3001MLP\uff0c\u4ee5\u53ca\u5b8c\u6574\u66ff\u63db\u7684 <code>vae2llm<\/code> \u548c <code>llm2vae<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><code>.bf16<\/code> \u7248\u672c\u7d04\u70ba\u5b8c\u6574\u7cbe\u5ea6\u6a94\u6848\u7684\u4e00\u534a\uff0chead \u8207 fp32 \u7684\u8f38\u51fa\u4e00\u81f4\u7387\u70ba 98.6%\u3002\u9805\u76ee\u9644\u6709 <code>scripts\/ckpt_io.py<\/code>\u3001\u8a13\u7df4\u8ff4\u5708\u548c\u63a8\u8ad6\u8173\u672c\uff1bComfyUI \u53ef\u628a LoRA \u8f09\u5165 YuE2 checkpoint \u7684 MODEL \u8f38\u51fa\uff0c\u5f37\u5ea6\u8a2d\u70ba 1.0\uff0cCLIP \u8f38\u5165\u4e0d\u53d7\u5f71\u97ff\u3002\u9801\u9762\u6c92\u6709\u63d0\u4f9b GGUF \u6a94\u6848\u3001\u6a94\u6848\u5927\u5c0f\u3001\u4e0a\u4e0b\u6587\u9577\u5ea6\u3001Ollama\u3001LM Studio\u3001llama.cpp \u6216 MTP draft speculation \u652f\u63f4\u8cc7\u6599\uff0c\u4ea6\u6c92\u6709\u5217\u51fa\u53ef\u4f9b\u9078\u64c7\u7684\u91cf\u5316\u7d1a\u5225\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u5148\u4ee5 4,765 \u9996 YuE2 \u81ea\u884c\u751f\u6210\u6b4c\u66f2\u8a13\u7df4\uff0c\u518d\u9069\u914d\u771f\u5be6\u97f3\u8a0a\uff1b\u56e0\u6b64\u5b83\u7684\u50f9\u503c\u96c6\u4e2d\u65bc\u63d0\u5347\u771f\u5be6\u9304\u97f3\u7684\u7de8\u78bc\u8207 NAR \u89e3\u78bc\uff0c\u800c\u975e\u53d6\u4ee3 YuE2-3B \u7684\u5b8c\u6574\u751f\u6210\u80fd\u529b\u3002\u6388\u6b0a\u70ba CC-BY-NC-4.0\uff0c\u5546\u696d\u7528\u9014\u9700\u8981\u53e6\u884c\u78ba\u8a8d\u6388\u6b0a\u689d\u4ef6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/huggingface.co\/Mothersuperior\/yue2-mothersuperior-realaudio-tokenizer-v4\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>YuE2\u53d6\u5f97\u97f3\u8a0a\u7de8\u78bc\u8207\u771f\u5be6\u9304\u97f3\u9069\u914d\u80fd\u529b\uff0c\u53ef\u5c07\u81ea\u5bb6\u9304\u97f3\u8f49\u6210\u8a9e\u7fa9\u78bc\uff0c\u518d\u5fae\u8abf\u6b4c\u624b\u98a8\u683c\u3002<\/p>\n","protected":false},"author":8,"featured_media":12524,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,125,76,127,29,38,211,213],"tags":[],"class_list":["post-12525","post","type-post","status-publish","format-standard","hentry","category-133","category-comfyui","category-76","category-127","category-audio","category-38","category-cantonese","category-tokenize"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12525","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=12525"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12525\/revisions"}],"predecessor-version":[{"id":12526,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12525\/revisions\/12526"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12524"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12525"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12525"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12525"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}