
{"id":8373,"date":"2026-05-19T22:53:39","date_gmt":"2026-05-19T14:53:39","guid":{"rendered":"https:\/\/infernews.com\/blog\/improving-text-and-face-fidelity-in-discrete-tokenization-for-autoregressive-ima\/"},"modified":"2026-05-19T22:53:39","modified_gmt":"2026-05-19T14:53:39","slug":"improving-text-and-face-fidelity-in-discrete-tokenization-for-autoregressive-ima","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/improving-text-and-face-fidelity-in-discrete-tokenization-for-autoregressive-ima\/","title":{"rendered":"InsightTok\uff1a\u4ee4\u751f\u5716\u6587\u5b57\u4eba\u81c9\u66f4\u6e05\u6670"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/Method-c8c407165c0e.jpg\" alt=\"Method\"><\/figure>\n<p>\u751f\u6210\u5716\u7247\u6642\uff0c\u6700\u5e38\u4ee4\u4eba\u51fa\u6232\u7684\u5f80\u5f80\u4e0d\u662f\u80cc\u666f\uff0c\u800c\u662f\u62db\u724c\u4e0a\u7684\u5b57\u3001\u6d77\u5831\u4e0a\u7684\u5b57\u53e5\uff0c\u6216\u8005\u4eba\u81c9\u4e94\u5b98\u7684\u7d30\u7bc0\u3002InsightTok \u91dd\u5c0d\u7684\u6b63\u6b63\u662f\u9019\u5169\u500b\u96e3\u4f4d\uff1a\u5b83\u4e0d\u662f\u76f4\u63a5\u6539\u6574\u500b\u751f\u5716\u6a21\u578b\uff0c\u800c\u662f\u5148\u6539\u5584\u5716\u7247\u88ab\u300c\u62c6\u89e3\u6210\u4ee3\u78bc\u300d\u7684\u65b9\u5f0f\uff0c\u8b93\u5f8c\u7e8c\u751f\u6210\u6642\u66f4\u6613\u4fdd\u7559\u91cd\u8981\u5167\u5bb9\u3002<\/p>\n<p>\u5be6\u969b\u4f7f\u7528\u4e0a\uff0c\u9019\u985e\u5c08\u6848\u8f03\u9069\u5408\u5df2\u7d93\u5728\u505a\u6587\u5b57\u751f\u6210\u5716\u7247\u3001\u7814\u7a76\u96e2\u6563\u8996\u89ba tokenization\uff0c\u6216\u4f7f\u7528\u81ea\u56de\u6b78\u5716\u50cf\u751f\u6210\u6d41\u7a0b\u7684\u958b\u767c\u8005\u8207\u7814\u7a76\u8005\u3002\u91cd\u9ede\u5728\u65bc\u628a\u539f\u6709\u6d41\u7a0b\u4e2d\u7684\u8996\u89ba tokenizer \u63db\u6210 InsightTok \u9019\u4e00\u985e\u65b9\u6848\uff0c\u5f8c\u9762\u7684\u751f\u6210\u6a21\u578b\u7406\u8ad6\u4e0a\u6bcb\u9808\u5927\u6539\uff0c\u5c0d\u73fe\u6709\u7ba1\u7dda\u7b97\u662f\u8f03\u5bb9\u6613\u63a5\u5165\u3002<\/p>\n<p>\u5b83\u7684\u7279\u5225\u4e4b\u8655\uff0c\u5728\u65bc\u6c92\u6709\u53ea\u7528\u4e00\u822c\u91cd\u5efa\u76ee\u6a19\uff0c\u800c\u662f\u66f4\u6709\u610f\u8b58\u5730\u7167\u9867\u5c40\u90e8\u800c\u4e14\u91cd\u8981\u7684\u5167\u5bb9\uff0c\u4f8b\u5982\u6587\u5b57\u5340\u57df\u8207\u4eba\u81c9\u5340\u57df\u3002\u6839\u64da\u5132\u5b58\u5eab\u63d0\u4f9b\u7684\u8aaa\u660e\uff0c\u5b83\u5728\u76f8\u540c\u58d3\u7e2e\u7387\u4e0b\uff0c\u80fd\u505a\u5230\u66f4\u597d\u7684\u6587\u5b57\u8207\u4eba\u81c9\u91cd\u5efa\uff0c\u540c\u6642\u53ea\u7528 16\u00d7 downsampling \u548c 16,384 \u500b\u689d\u76ee\u7684 codebook\uff0c\u984d\u5916\u8a13\u7df4\u6210\u672c\u4ea6\u76f8\u5c0d\u6709\u9650\u3002<\/p>\n<ul>\n<li>\u4e3b\u8981\u6539\u5584\u5716\u7247\u4e2d\u7684<strong>\u6587\u5b57\u53ef\u8b80\u6027<\/strong>\u8207<strong>\u4eba\u81c9\u7d30\u7bc0\u9084\u539f<\/strong><\/li>\n<li>\u53ef\u517c\u5bb9\u6a19\u6e96\u81ea\u56de\u6b78\u5716\u50cf\u751f\u6210\u6d41\u7a0b\uff0c\u6bcb\u9808\u9023\u4e0b\u6e38\u6a21\u578b\u4e00\u4f75\u91cd\u5beb<\/li>\n<li>\u5728\u76f8\u540c\u58d3\u7e2e\u689d\u4ef6\u4e0b\uff0c\u91cd\u5efa\u6548\u679c\u4e3b\u6253\u66f4\u6e05\u6670\u3001\u66f4\u5fe0\u65bc\u539f\u5716<\/li>\n<li>\u8a2d\u8a08\u4e0a\u5ef6\u7e8c VQGAN \u98a8\u683c tokenizer \u601d\u8def\uff0c\u4f46\u52a0\u5f37\u91cd\u9ede\u5340\u57df supervision<\/li>\n<li>\u76f8\u95dc\u8108\u7d61\u5305\u62ec\u96e2\u6563\u8996\u89ba tokenizer\u3001VQGAN \u98a8\u683c\u65b9\u6cd5\uff0c\u4ee5\u53ca\u6587\u5b57\u751f\u6210\u5716\u7247\u6a21\u578b<\/li>\n<\/ul>\n<p>\u6574\u9ad4\u4f86\u770b\uff0cInsightTok \u5438\u5f15\u4e4b\u8655\u4e0d\u5728\u82b1\u5de7\u529f\u80fd\uff0c\u800c\u5728\u65bc\u5b83\u6293\u5230\u751f\u5716\u6700\u5e38\u88ab\u6279\u8a55\u7684\u75db\u9ede\uff0c\u518d\u7528\u76f8\u5c0d\u52d9\u5be6\u7684\u65b9\u6cd5\u88dc\u5f37\u3002\u5c0d\u4e00\u822c\u8b80\u8005\u4f86\u8aaa\uff0c\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\u4e00\u500b\u300c\u4ee4 AI \u66f4\u8b58\u5f97\u4fdd\u4f4f\u5b57\u540c\u6a23\u8c8c\u300d\u7684\u5e95\u5c64\u96f6\u4ef6\uff1b\u5c0d\u6280\u8853\u5718\u968a\u800c\u8a00\uff0c\u5b83\u8f03\u50cf\u662f\u4e00\u500b\u53ef\u76f4\u63a5\u63d0\u5347\u756b\u9762\u53ef\u7528\u6027\u7684\u57fa\u790e\u7d44\u4ef6\u3002<\/p>\n<p><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/LeapLabTHU\/InsightTok\" rel=\"noopener noreferrer\">https:\/\/github.com\/LeapLabTHU\/InsightTok<\/a><\/p>\n<p><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2605.14333\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2605.14333<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9019\u500b\u5c08\u6848\u5c08\u6ce8\u6539\u5584\u751f\u6210\u5716\u7247\u4e2d\u6700\u6613\u5931\u771f\u7684\u6587\u5b57\u8207\u4eba\u81c9\u7d30\u7bc0\u3002\u5c0d\u5df2\u6709\u81ea\u56de\u6b78\u751f\u5716\u6d41\u7a0b\u7684\u4eba\u4f86\u8aaa\uff0c\u7279\u5225\u6709\u53c3\u8003\u50f9\u503c\u3002<\/p>\n","protected":false},"author":8,"featured_media":8372,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,173,149,141,189,196],"tags":[],"class_list":["post-8373","post","type-post","status-publish","format-standard","hentry","category-133","category-173","category-149","category-141","category-189","category-196"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8373","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=8373"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8373\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8372"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8373"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8373"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8373"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}