{"id":12514,"date":"2026-09-22T18:16:34","date_gmt":"2026-09-22T10:16:34","guid":{"rendered":"https:\/\/infernews.com\/blog\/explore-physics-in-video-generation-models\/"},"modified":"2026-09-22T18:16:34","modified_gmt":"2026-09-22T10:16:34","slug":"explore-physics-in-video-generation-models","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/explore-physics-in-video-generation-models\/","title":{"rendered":"physics\uff1a\u62c6\u89e3 Wan2.1 \u7684\u7269\u7406\u5931\u8aa4"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/cross_attention.png\" alt=\"teaser\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u500b\u7c43\u7403\u7531\u534a\u7a7a\u8dcc\u843d\u5f8c\uff0c\u7406\u61c9\u649e\u5730\u518d\u5f48\u8d77\uff0cWan2.1-T2V \u537b\u53ef\u80fd\u751f\u6210\u6f02\u6d6e\u3001\u7a81\u7136\u505c\u9813\u6216\u534a\u7a7a\u53cd\u5f48\u3002\u9019\u500b GitHub \u9805\u76ee\u5c6c\u65bc\u5f71\u7247\u751f\u6210\u6a21\u578b\u7684\u53ef\u89e3\u91cb\u6027\u7814\u7a76\uff0c\u5be6\u969b\u8655\u7406\u7684\u662f text-to-video diffusion models \u5982\u4f55\u5728\u65e9\u671f\u53bb\u566a\u968e\u6bb5\u898f\u5283\u7269\u4ef6\u8ecc\u8de1\uff0c\u4ee5\u53ca\u932f\u8aa4\u4f4d\u7f6e\u70ba\u4f55\u6703\u4e00\u8def\u9396\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u628a\u8996\u7dda\u96c6\u4e2d\u5728 Wan2.1-T2V-1.3B \u7684\u6700\u521d\u4e94\u500b\u53bb\u566a\u6b65\u9a5f\uff0c\u900f\u904e video-to-object-token cross-attention \u8ffd\u8e64\u7269\u4ef6\u4f4d\u7f6e\uff0c\u518d\u914d\u5408 convergence speed \u548c causal contribution \u627e\u51fa\u8ca0\u8cac\u5f62\u6210\u8ecc\u8de1\u7684 attention heads\u3002\u5206\u6790\u6307\u51fa\uff0cSelf-attention \u4e2d\u7684 Rotary Position Embedding (RoPE) \u6703\u4ee4\u7a7a\u9593\u6ce8\u610f\u529b\u8870\u6e1b\u904e\u5feb\uff0c\u65e9\u671f\u932f\u8aa4\u5019\u9078\u4f4d\u7f6e\u56e0\u6b64\u58d3\u904e\u9130\u8fd1\u5f71\u683c\u8f03\u5408\u7406\u7684\u8def\u5f91\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4fee\u6b63\u65b9\u6cd5\u4e0d\u4f9d\u8cf4\u5916\u90e8\u7269\u7406\u6a21\u64ec\u5668\u6216\u984d\u5916 foundation-model teacher\uff0c\u800c\u662f\u5728\u65e9\u671f\u6b65\u9a5f\u964d\u4f4e height\/width RoPE frequency\uff0c\u6e2c\u8a66\u6642\u524d\u4e94\u6b65\u4f7f\u7528 \u03bb^{h\/w}=0.70\uff1b\u9700\u8981\u8a13\u7df4\u6642\u5247\u4ee5\u53ea\u4f5c\u7528\u65bc self-attention \u548c cross-attention \u7684 Low-Rank Adaptation (LoRA) \u914d\u5408 \u03bb^{h\/w}=0.75\u3002LoRA \u8a2d\u5b9a\u70ba r=64\u3001alpha=32\uff0c\u4e26\u4ee5\u9996 10% \u53bb\u566a\u6b65\u9a5f\u4f54 0.9 \u6a5f\u7387\u7684\u6df7\u5408 timestep sampler \u52a0\u5f37\u65e9\u671f\u5b78\u7fd2\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6e2c\u8a66\u9019\u500b\u9805\u76ee\u6642\u53ef\u6cbf\u7528 Wan \u5b98\u65b9\u6a21\u578b\u8207\u63d0\u793a\u8a5e\uff0c\u4ee5\u4e0d\u540c random seeds \u6bd4\u8f03\u539f\u6a21\u578b\u3001training-free \u983b\u7387\u8abf\u6574\u53ca LoRA \u7248\u672c\u7684\u7c43\u7403\u8dcc\u843d\u7b49\u7247\u6bb5\uff1b\u7a0b\u5f0f\u900f\u904e runtime monkey patches \u4fee\u6539\u884c\u70ba\uff0c\u4e0d\u9700\u8981\u6539\u52d5\u5b98\u65b9 Wan source\u3002\u63d0\u4f9b\u7684\u8cc7\u8a0a\u78ba\u8a8d\u6709 training-free \u548c training-based \u5be6\u9a57\uff0c\u4f46\u672a\u5217\u51fa\u5b8c\u6574\u6578\u503c\u6307\u6a19\uff0c\u56e0\u6b64\u66f4\u9069\u5408\u8996\u70ba\u6a5f\u5236\u5206\u6790\u8207\u53ef\u91cd\u73fe\u7684\u7814\u7a76\u539f\u578b\uff0c\u800c\u975e\u5df2\u8b49\u660e\u80fd\u8655\u7406\u6240\u6709\u7269\u7406\u5834\u666f\u7684\u901a\u7528\u4fee\u88dc\u65b9\u6848\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7269\u4ef6\u8ecc\u8de1\u4e3b\u8981\u5728\u6700\u521d\u4e94\u500b\u53bb\u566a\u6b65\u9a5f\u5f62\u6210<\/li>\n<li>Cross-attention \u8ca0\u8cac\u5f15\u5165\u7269\u4ef6\u8a9e\u7fa9\uff0cself-attention \u5354\u52a9\u9078\u64c7\u5019\u9078\u5340\u57df<\/li>\n<li>RoPE \u7a7a\u9593\u8870\u6e1b\u904e\u5f37\uff0c\u6703\u4ee4\u932f\u8aa4\u4f4d\u7f6e\u904e\u65e9\u56fa\u5b9a<\/li>\n<li>\u53ef\u7528\u65e9\u671f\u983b\u7387\u7e2e\u653e\uff0c\u6216\u52a0\u5165\u53ea\u6539\u6ce8\u610f\u529b\u5c64\u7684 LoRA<\/li>\n<li>\u9069\u5408\u7814\u7a76\u5f71\u7247\u751f\u6210\u3001\u6a21\u578b\u53ef\u89e3\u91cb\u6027\u53ca\u7269\u7406\u4e00\u81f4\u6027\u7684\u5718\u968a<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/siriuslala.github.io\/physics\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Siriuslala\/physics\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7c43\u7403\u5f48\u8df3\u8b8a\u6210\u534a\u7a7a\u6f02\u6d6e\uff0c\u554f\u984c\u53ef\u80fd\u65e9\u5728\u53bb\u566a\u982d\u5e7e\u6b65\u7531\u6ce8\u610f\u529b\u6c7a\u5b9a\u3002\u9019\u500b\u7814\u7a76\u4ee5 RoPE \u8abf\u6574\u6539\u5584\u5f71\u7247\u52d5\u614b\u3002<\/p>\n","protected":false},"author":8,"featured_media":12513,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,162,131,76,141,127,31,161],"tags":[],"class_list":["post-12514","post","type-post","status-publish","format-standard","hentry","category-133","category-ai-productions","category-embedding","category-76","category-141","category-127","category-video","category-161"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12514","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=12514"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12514\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12513"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12514"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12514"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12514"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}