{"id":9840,"date":"2026-07-03T21:51:46","date_gmt":"2026-07-03T13:51:46","guid":{"rendered":"https:\/\/infernews.com\/blog\/multi-resolution-flow-matching-training-free-diffusion-acceleration-via-staged-s\/"},"modified":"2026-07-03T21:51:46","modified_gmt":"2026-07-03T13:51:46","slug":"multi-resolution-flow-matching-training-free-diffusion-acceleration-via-staged-s","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/multi-resolution-flow-matching-training-free-diffusion-acceleration-via-staged-s\/","title":{"rendered":"MrFlow\uff1a\u6587\u5b57\u751f\u6210\u5716\u7247\u63d0\u901f\u65b0\u8def\u7dda"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/framework.jpg\" alt=\"MrFlow framework\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">MrFlow \u662f\u4e00\u500b<strong>\u8a13\u7df4\u514d\u9664\u7684\u5716\u50cf\u751f\u6210\u52a0\u901f\u65b9\u6cd5<\/strong>\uff0c\u5c6c\u65bc\u91dd\u5c0d flow-matching text-to-image diffusion models \u7684\u7814\u7a76\u539f\u578b\u8207\u5be6\u4f5c\u3002\u5b83\u8981\u89e3\u6c7a\u7684\u554f\u984c\u5f88\u76f4\u63a5\uff1a\u9ad8\u89e3\u6790\u5ea6\u51fa\u5716\u592a\u6162\uff0c\u56e0\u6b64\u5148\u5728\u4f4e\u89e3\u6790\u5ea6\u5b8c\u6210\u5927\u90e8\u5206\u751f\u6210\uff0c\u518d\u7528\u8f03\u77ed\u7684\u9ad8\u89e3\u6790\u5ea6\u4fee\u88dc\u6d41\u7a0b\u88dc\u56de\u7d30\u7bc0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u591a\u89e3\u6790\u5ea6\u52a0\u901f\u65b9\u6cd5\uff0c\u901a\u5e38\u6703\u5728 latent space \u505a\u4e0a\u63a1\u6a23\uff0c\u6216\u8005\u53ea\u6539\u52d5\u90e8\u5206\u5340\u57df\uff1b\u4f5c\u8005\u8a8d\u70ba\u9019\u7a2e\u56fa\u5b9a\u7bc4\u5f0f\u5bb9\u6613\u5e36\u4f86\u6a21\u7cca\u611f\u8207 artifact\u3002MrFlow \u6539\u7528 pixel space \u7684 Real-ESRGAN \u505a\u8d85\u89e3\u6790\u5ea6\uff0c\u4e4b\u5f8c\u91cd\u65b0\u7de8\u78bc\u3001\u6ce8\u5165\u8207 scheduler \u4e00\u81f4\u7684\u4f4e\u5f37\u5ea6\u96dc\u8a0a\uff0c\u518d\u505a\u77ed\u6b65\u6578 refinement\uff0c\u5c07\u6602\u8cb4\u7684\u9ad8\u89e3\u6790\u5ea6 denoising \u6210\u672c\uff0c\u5927\u5e45\u8f49\u79fb\u5230\u8f03\u4fbf\u5b9c\u7684\u4f4e\u89e3\u6790\u5ea6\u968e\u6bb5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u7684\u53d6\u5411\u76f8\u7576\u52d9\u5be6\uff0c\u56e0\u70ba\u5b83\u4e0d\u8981\u6c42 finetuning\u3001learned upsampler\uff0c\u4ea6\u5514\u9700\u8981 model-specific retraining \u6216 custom kernels\uff0c\u76f4\u63a5\u5efa\u7acb\u5728 PyTorch\u3001Diffusers \u8207\u65e2\u6709 scheduler \u63a7\u5236\u4e4b\u4e0a\u3002\u90e8\u7f72\u7406\u89e3\u4e0a\uff0c\u5b83\u6bd4\u8f03\u50cf\u4e00\u689d\u53ef\u63d2\u5165\u73fe\u6210\u6a21\u578b\u6d41\u7a0b\u7684 sampling pipeline\uff1a\u5148\u6e96\u5099 Diffusers \u76f8\u5bb9\u74b0\u5883\u3001\u5c0d\u61c9\u7684\u9810\u8a13\u7df4\u6a21\u578b\u6b0a\u91cd\uff0c\u4ee5\u53ca Real-ESRGAN \u7684 x2 \u6b0a\u91cd\uff0c\u518d\u628a README \u5167\u793a\u4f8b\u8173\u672c\u7684 checkpoint \u8def\u5f91\u63db\u6210\u672c\u5730\u8a2d\u5b9a\u5373\u53ef\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6548\u80fd\u6578\u5b57\u662f\u9019\u500b\u9805\u76ee\u7684\u6838\u5fc3\u8ce3\u9ede\u3002\u8cc7\u6599\u6307\u51fa\uff0cMrFlow \u5728 Qwen-Image \u53ef\u505a\u5230\u8d85\u904e 10x end-to-end speedup\uff0c\u8ad6\u6587\u4ea6\u63d0\u5230\u76f8\u5c0d\u52a0\u901f\u524d\u7684 OneIG \u5dee\u8ddd\u53ef\u63a7\u5236\u5728 1% \u4ee5\u5167\uff1b\u518d\u914d\u5408 timestep-distilled models\uff0c\u4f8b\u5982 Pi-Flow \u8207 FLUX-schnell\uff0c\u6574\u9ad4\u52a0\u901f\u53ef\u9032\u4e00\u6b65\u53bb\u5230 25x\u3002\u9019\u7a2e\u8a2d\u8a08\u4e5f\u5df2\u5c55\u793a\u53ef\u8f49\u79fb\u5230 Qwen-Image\u3001FLUX.1-dev\u3001FLUX.2 Klein \u8207 Z-Image family\uff0c\u4ee3\u8868\u5b83\u4e0d\u662f\u53ea\u7d81\u6b7b\u55ae\u4e00\u6a21\u578b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u985e\u578b\u5b9a\u4f4d<\/strong>\uff1a\u8a13\u7df4\u514d\u9664\u7684 staged sampling \u52a0\u901f\u65b9\u6cd5\uff0c\u5514\u4fc2\u65b0\u5e95\u6a21\u672c\u8eab<\/li>\n<li><strong>\u4e3b\u8981\u5dee\u7570<\/strong>\uff1a\u907f\u958b latent space \u4e0a\u63a1\u6a23\u8def\u7dda\uff0c\u6539\u7528 pixel space \u8d85\u89e3\u6790\u5ea6\u52a0\u77ed\u7a0b\u9ad8\u89e3\u6790\u5ea6\u4fee\u88dc<\/li>\n<li><strong>\u90e8\u7f72\u91cd\u9ede<\/strong>\uff1a\u4f9d\u8cf4 PyTorch\u3001Diffusers\u3001Transformers\u3001Real-ESRGAN\uff0c\u4e26\u9700\u81ea\u884c\u914d\u7f6e\u6a21\u578b\u8207\u6b0a\u91cd\u8def\u5f91<\/li>\n<li><strong>\u9069\u5408\u5834\u666f<\/strong>\uff1a\u8981\u4fdd\u7559\u756b\u8cea\u3001\u53c8\u60f3\u7e2e\u77ed\u751f\u6210\u6642\u9593\u7684\u5716\u50cf\u751f\u6210\u5718\u968a\u8207\u7814\u7a76\u4eba\u54e1<\/li>\n<li><strong>\u76f8\u95dc\u6a21\u578b<\/strong>\uff1aQwen-Image\u3001FLUX.1-dev\u3001FLUX.2 Klein\u3001Z-Image\u3001Pi-Flow\u3001FLUX-schnell<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">MrFlow \u6700\u9069\u5408\u653e\u5728\u5df2\u6709\u6587\u5b57\u751f\u6210\u5716\u7247\u6d41\u7a0b\u7684\u9805\u76ee\u4e4b\u4e2d\uff0c\u4f5c\u70ba\u52a0\u901f\u5c64\u800c\u4e0d\u662f\u5b8c\u6574\u66ff\u4ee3\u54c1\u3002\u5b83\u7684\u9650\u5236\u4ea6\u5f88\u6e05\u695a\uff1a\u4ecd\u7136\u4f9d\u8cf4\u5916\u90e8\u8d85\u89e3\u6790\u5ea6\u6a21\u578b\u8207\u65e2\u6709 backbone \u54c1\u8cea\uff0c\u91cd\u9ede\u5728\u65bc\u91cd\u65b0\u5206\u914d\u7b97\u529b\u6210\u672c\uff0c\u672a\u5fc5\u7b49\u65bc\u6240\u6709\u63d0\u793a\u8a5e\u3001\u6240\u6709\u756b\u98a8\u90fd\u80fd\u7121\u4ee3\u50f9\u8907\u88fd\u539f\u672c\u9ad8\u89e3\u6790\u5ea6\u9577\u6b65\u6578\u751f\u6210\u7684\u7d50\u679c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/Xingyu-Zheng\/MrFlow\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.01642\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>MrFlow \u7528\u5206\u968e\u6bb5\u53d6\u6a23\u52a0\u901f flow-matching \u5716\u50cf\u751f\u6210\uff0c\u91cd\u9ede\u662f\u5514\u4f7f\u91cd\u65b0\u8a13\u7df4\u6a21\u578b\u3002\u5b83\u8ffd\u6c42\u9ad8\u500d\u63d0\u901f\uff0c\u540c\u6642\u76e1\u91cf\u5b88\u4f4f\u756b\u9762\u7d30\u7bc0\u3002<\/p>\n","protected":false},"author":8,"featured_media":9839,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,179,33,30,152,129,189],"tags":[],"class_list":["post-9840","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-nvidia","category-stablediffusion","category-image","category-python","category-txt2img","category-189"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9840","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=9840"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9840\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9839"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9840"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9840"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9840"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}