{"id":8936,"date":"2026-06-08T23:50:34","date_gmt":"2026-06-08T15:50:34","guid":{"rendered":"https:\/\/infernews.com\/blog\/phaselock\/"},"modified":"2026-06-08T23:50:34","modified_gmt":"2026-06-08T15:50:34","slug":"phaselock","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/phaselock\/","title":{"rendered":"PhaseLock\uff1a\u7528\u5169\u6b65\u9396\u4f4f\u5f71\u7247\u7269\u7406\u611f"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/fig1-dae68bbf3a3f.jpg\" alt=\"yonsei emblem\"><\/figure>\n<p>PhaseLock \u662f\u4e00\u500b\u91dd\u5c0d Image-to-Video diffusion models \u7684\u65b9\u6cd5\uff0c\u91cd\u9ede\u662f\u4fee\u6b63\u5f71\u7247\u751f\u6210\u4e2d\u5e38\u898b\u7684\u7269\u7406\u932f\u8aa4\u3002\u4e3b\u8981\u662f\u91dd\u5c0d inference-time method \/ sampling strategy\u3002\u6a21\u578b\u5728\u5b8c\u6574 50 \u6b65\u53bb\u566a\u6642\u96d6\u7136\u756b\u9762\u66f4\u7d30\u7dfb\uff0c\u4f46\u52d5\u4f5c\u53cd\u800c\u53ef\u80fd\u504f\u96e2\u7269\u7406\u898f\u5f8b\uff1b\u76f8\u5c0d\u5730\uff0c\u53ea\u505a 2 \u6b65\u53bb\u566a\u6642\uff0c\u52d5\u4f5c\u5148\u9a57\u66f4\u53ef\u4fe1\uff0c\u53ea\u662f\u8cea\u611f\u8f03\u7c97\u7cd9\u3002<\/p>\n<p>\u9805\u76ee\u7684\u6838\u5fc3\u505a\u6cd5\u662f\u5169\u968e\u6bb5\u6d41\u7a0b\uff0c\u800c\u4e14\u4e0d\u9700\u8981\u984d\u5916\u8a13\u7df4\u3002\u5b83\u6703\u5148\u7528 2 \u6b65\u53bb\u566a\u62bd\u51fa motion prior\uff0c\u6587\u4e2d\u4ee5 \u0394 phys \u8868\u793a\uff0c\u518d\u5728 50 \u6b65\u5b8c\u6574\u751f\u6210\u671f\u9593\u4ee5 Latent Delta Guidance \u91cd\u65b0\u6ce8\u5165\uff0c\u76ee\u6a19\u662f\u5728\u9ad8\u4fdd\u771f\u756b\u9762\u4e2d\u4fdd\u7559\u8f03\u5408\u7406\u7684\u52d5\u614b\u7d50\u679c\u3002<\/p>\n<p>\u4f8b\u5b50\u5f88\u76f4\u89c0\uff0c\u4f8b\u5982\u975e\u78c1\u6027\u7684\u7db2\u7403\u4e0d\u61c9\u88ab\u5e36\u78c1\u7684\u7c43\u5b50\u5438\u8d77\u3002\u57fa\u7dda\u7d50\u679c\u6703\u7522\u751f\u9055\u53cd\u5e38\u8b58\u7684\u52d5\u4f5c\uff0cPhaseLock \u5247\u8f03\u80fd\u7dad\u6301\u7269\u4ef6\u61c9\u6709\u7684\u79fb\u52d5\u65b9\u5f0f\u3002\u9019\u985e\u60c5\u6cc1\u5f88\u9069\u5408\u7528\u65bc\u9700\u8981\u57fa\u672c\u7269\u7406\u5408\u7406\u6027\u7684\u5f71\u7247\u751f\u6210\u9805\u76ee\uff0c\u4f8b\u5982\u7269\u4ef6\u4e92\u52d5\u3001\u6389\u843d\u3001\u6293\u53d6\u6216\u63a5\u89f8\u5834\u666f\u3002<\/p>\n<p>\u91cd\u9ede\u53ef\u6b78\u7d0d\u70ba\uff1a<br \/>\n&#8211; \u4ee5 <strong>training-free<\/strong> \u65b9\u5f0f\u6539\u5584\u5f71\u7247\u4e2d\u7684\u7269\u7406\u4e00\u81f4\u6027<br \/>\n&#8211; \u767c\u73fe 2-step generation \u7684 physics \u53ef\u80fd\u6bd4 50-step output \u66f4\u597d<br \/>\n&#8211; \u900f\u904e <strong>Latent Delta Guidance<\/strong> \u628a\u65e9\u671f motion prior \u9396\u56de\u6700\u7d42\u7d50\u679c<br \/>\n&#8211; \u5831\u544a\u6307\u51fa physical consistency \u5e73\u5747\u63d0\u5347 <strong>+6.2 points<\/strong><br \/>\n&#8211; \u984d\u5916\u6210\u672c\u76f8\u5c0d\u6709\u9650\uff0c\u7d04 <strong>1.06\u00d7 time<\/strong>\u3001<strong>1.02\u00d7 memory<\/strong><\/p>\n<p>\u5982\u679c\u4f60\u672c\u8eab\u5df2\u5728\u7528\u5f71\u7247\u64f4\u6563\u6a21\u578b\uff0c\u9019\u500b\u9805\u76ee\u7684\u4f7f\u7528\u6982\u5ff5\u4e0d\u7b97\u8907\u96dc\uff1a\u5148\u8dd1\u77ed\u6b65\u6578\u7d50\u679c\u53d6\u51fa\u52d5\u4f5c\u8a0a\u865f\uff0c\u518d\u914d\u5408\u5b8c\u6574\u6b65\u6578\u751f\u6210\u3002\u5f9e\u73fe\u6709\u5167\u5bb9\u770b\uff0cPhaseLock \u7684\u50f9\u503c\u4e0d\u5728\u65bc\u66f4\u63db\u4e3b\u6a21\u578b\uff0c\u800c\u662f\u5728\u540c\u4e00\u6a21\u578b\u4e4b\u4e0a\u88dc\u56de\u88ab\u5f8c\u671f\u53bb\u566a\u300c\u78e8\u8d70\u300d\u7684\u52d5\u4f5c\u5148\u9a57\u3002\u6587\u4e2d\u63d0\u5230\u6e2c\u8a66\u7528\u7684\u6a21\u578b\u5305\u62ec <strong>Wan 2.1<\/strong>\u3002<\/p>\n<p><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/dnwjddl\/phaselock\" rel=\"noopener noreferrer\">https:\/\/github.com\/dnwjddl\/phaselock<\/a><\/p>\n<p><strong>\u9805\u76ee\uff1a<\/strong> <a href=\"https:\/\/dnwjddl.github.io\/phaselock\/\" rel=\"noopener noreferrer\">https:\/\/dnwjddl.github.io\/phaselock\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9019\u9805\u76ee\u7528\u514d\u8a13\u7df4\u65b9\u6cd5\uff0c\u6539\u5584\u5f71\u7247\u751f\u6210\u4e2d\u7684\u4e0d\u5408\u7406\u52d5\u4f5c\uff0c\u540c\u6642\u76e1\u91cf\u4fdd\u7559\u9ad8\u756b\u8cea\u3002<\/p>\n","protected":false},"author":8,"featured_media":8935,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,179,184,197],"tags":[],"class_list":["post-8936","post","type-post","status-publish","format-standard","hentry","category-133","category-nvidia","category-robotic","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8936","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=8936"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8936\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8935"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8936"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8936"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8936"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}