{"id":8912,"date":"2026-06-07T14:35:45","date_gmt":"2026-06-07T06:35:45","guid":{"rendered":"https:\/\/infernews.com\/blog\/stable-layers-vlm\/"},"modified":"2026-06-07T14:35:45","modified_gmt":"2026-06-07T06:35:45","slug":"stable-layers-vlm","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/stable-layers-vlm\/","title":{"rendered":"Stable-Layers\uff1a\u9760 VLM \u8a55\u5206\u5f37\u5316\u5b78\u7fd2\uff0c\u6bcb\u9808\u914d\u5c0d\u6578\u64da\u4e5f\u80fd\u6539\u826f\u5716\u5c64\u5206\u96e2"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-c709b53db0ab.jpg\" alt=\"Hero image preview\"><\/figure>\n<p>\u5716\u50cf\u5716\u5c64\u5206\u89e3\uff08image layer decomposition\uff09\u662f\u6307\u628a\u4e00\u5f35\u5716\u5206\u6210\u5e7e\u5c64\u53ef\u7368\u7acb\u7de8\u8f2f\u7684 RGBA \u5f71\u50cf\uff0c\u518d\u5408\u4f75\u56de\u539f\u672c\u7684\u756b\u9762\u3002\u9019\u500b\u6280\u8853\u662f\u5c08\u696d\u4fee\u5716\u8207\u5408\u6210\u5de5\u4f5c\u6d41\u7684\u57fa\u672c\u5de5\u5e8f\uff0c\u4f46\u8981\u8a13\u7df4\u6a21\u578b\u505a\u9019\u4ef6\u4e8b\u4e26\u4e0d\u5bb9\u6613\uff1a\u540c\u4e00\u5f35\u5716\u5f80\u5f80\u5b58\u5728\u591a\u7a2e\u5408\u7406\u5206\u5c64\u65b9\u5f0f\uff0c\u800c\u4e14\u54c1\u8cea\u597d\u58de\u53d6\u6c7a\u65bc\u4e0b\u6e38\u662f\u5426\u597d\u7528\uff0c\u4f8b\u5982\u8a9e\u610f\u5206\u5c64\u662f\u5426\u6e05\u6670\u3001alpha \u906e\u7f69\u662f\u5426\u4e7e\u6de8\u3001\u662f\u5426\u6709\u5197\u9918\u5c64\uff0c\u4ee5\u53ca\u88ab\u906e\u64cb\u7684\u90e8\u5206\u80fd\u5426\u88ab\u5408\u7406\u586b\u56de\u3002<\/p>\n<p>\u50b3\u7d71\u505a\u6cd5\u6703\u7528\u5408\u6210\u7684\u914d\u5c0d\u6578\u64da\u96c6\uff08\u5373\u540c\u4e00\u5f35\u5716\u540c\u6642\u63d0\u4f9b\u300c\u539f\u5716\u300d\u8207\u300c\u6b63\u78ba\u5206\u5c64\u300d\uff09\u4f86\u76e3\u7763\u6a21\u578b\u5b78\u7fd2\uff0c\u4f46\u9019\u6703\u5e36\u4f86\u5148\u5929\u9650\u5236\uff1a\u7576\u591a\u7a2e\u5206\u5c64\u90fd\u540c\u6a23\u5408\u7406\u6642\uff0c\u5f37\u8feb\u6a21\u578b\u64ec\u5408\u55ae\u4e00\u6a19\u6e96\u7b54\u6848\uff0c\u7b49\u65bc\u61f2\u7f70\u4e86\u5176\u4ed6\u53ef\u884c\u7684\u89e3\u6cd5\u3002Stable-Layers \u5617\u8a66\u7e5e\u904e\u9019\u500b\u9650\u5236\uff0c\u6539\u7528\u5f37\u5316\u5b78\u7fd2\uff08reinforcement learning\uff09\u8b93\u6a21\u578b\u76f4\u63a5\u671d\u300c\u770b\u8d77\u4f86\u54c1\u8cea\u597d\u300d\u7684\u65b9\u5411\u512a\u5316\uff0c\u76e3\u7763\u8a0a\u865f\u53ea\u4f86\u81ea\u4e00\u500b\u8996\u89ba\u8a9e\u8a00\u6a21\u578b\uff08vision-language model\uff0cVLM\uff09\u3002<\/p>\n<p>\u5177\u9ad4\u800c\u8a00\uff0c\u9805\u76ee\u4ee5 Qwen-Image-Layered \u70ba\u8d77\u9ede\uff0c\u7d50\u5408 Flow-GRPO \u8207 LoRA\uff08Low-Rank Adaptation\uff0c\u4f4e\u79e9\u9069\u61c9\uff09\u5fae\u8abf\uff0c\u91dd\u5c0d\u6bcf\u5f35\u5716\u63a1\u6a23\u591a\u500b\u5019\u9078\u5206\u5c64\uff0c\u518d\u7528 VLM \u8a55\u5206\uff0c\u5f9e\u7fa4\u7d44\u76f8\u5c0d\u512a\u52e2\uff08group-relative advantages\uff09\u4f86\u66f4\u65b0\u7b56\u7565\u3002\u7576\u4e2d\u6700\u5927\u7684\u6311\u6230\u662f\u8a2d\u8a08\u53ef\u9760\u7684\u734e\u52f5\u8a0a\u865f\uff1aVLM \u55ae\u7368\u8a55\u5206\u6642\u5bb9\u6613\u628a\u6240\u6709\u6a23\u672c\u64e0\u9032\u4e00\u500b\u72f9\u7a84\u7684\u5206\u6578\u5340\u9593\uff0c\u5c0e\u81f4 GRPO \u7f3a\u4e4f\u7d44\u5167\u8b8a\u5316\u53ef\u4ee5\u5b78\u7fd2\u3002\u70ba\u6b64\uff0cStable-Layers \u63a1\u7528\u5169\u968e\u6bb5\u8a55\u4f30\u6d41\u7a0b\u2014\u2014\u5148\u6309\u4e94\u500b\u7de8\u8f2f\u5411\u6a19\u6e96\u9010\u9805\u8a55\u5206\uff0c\u518d\u628a\u6240\u6709\u5019\u9078\u4e26\u6392\u653e\u5728\u6a19\u8a18\u597d\u7684\u6bd4\u8f03\u7db2\u683c\u4e0a\u91cd\u65b0\u8a55\u5206\u4e00\u6b21\uff0c\u85c9\u6b64\u53d6\u5f97\u66f4\u7d30\u7dfb\u7684\u76f8\u5c0d\u6821\u6e96\u3002<\/p>\n<p><strong>Stable-Layers \u91cd\u9ede\u6458\u8981\uff1a<\/strong><\/p>\n<ul>\n<li>\u6bcb\u9808\u914d\u5c0d\u6578\u64da\uff1a\u5728\u5b8c\u5168\u6c92\u6709\u6a19\u8a3b\u7684\u5716\u50cf\u4e0a\u8a13\u7df4\uff0c\u89e3\u6c7a\u5408\u6210\u6578\u64da\u96c6\u5e36\u4f86\u7684\u504f\u8aa4\u554f\u984c<\/li>\n<li>VLM \u64d4\u4efb\u8a55\u5be9\uff1a\u5229\u7528\u8996\u89ba\u8a9e\u8a00\u6a21\u578b\u5c31\u4e94\u500b\u7de8\u8f2f\u6a19\u6e96\u6253\u5206\uff0c\u63d0\u4f9b\u734e\u52f5\u8a0a\u865f<\/li>\n<li>\u5169\u968e\u6bb5\u8a55\u4f30\uff1a\u5148\u7368\u7acb\u8a55\u5206\uff0c\u518d\u4ee5\u6bd4\u8f03\u7db2\u683c\u91cd\u65b0\u6821\u6e96\uff0c\u907f\u514d\u5206\u6578\u904e\u5ea6\u96c6\u4e2d<\/li>\n<li>\u5f37\u5316\u5b78\u7fd2\u5fae\u8abf\uff1a\u7d50\u5408 Flow-GRPO \u8207 LoRA\uff0c\u5f9e Qwen-Image-Layered \u958b\u59cb\u6539\u826f<\/li>\n<li>\u5be6\u6e2c\u8868\u73fe\uff1a\u5728 Crello \u6578\u64da\u96c6\u4e0a\uff0c\u5716\u5c64\u5206\u96e2\u5ea6\u66f4\u9ad8\u3001\u7a7a\u767d\u6216\u5e36\u7455\u75b5\u7684\u5c64\u66f4\u5c11\u3001\u6bcf\u5c64\u91cd\u5efa\u8aa4\u5dee\u4e5f\u8f03\u4f4e<\/li>\n<\/ul>\n<p><strong>\u9069\u7528\u5834\u666f\u8207\u5c0d\u8c61\uff1a<\/strong>\u9019\u500b\u9805\u76ee\u9069\u5408\u505a\u5716\u50cf\u7de8\u8f2f\u3001\u5408\u6210\u6216\u8a2d\u8a08\u5de5\u5177\u7684\u7814\u7a76\u8005\u8207\u5de5\u7a0b\u5e2b\uff0c\u5c24\u5176\u662f\u624b\u4e0a\u6c92\u6709\u5927\u91cf\u914d\u5c0d\u5206\u5c64\u6578\u64da\u3001\u53c8\u60f3\u63d0\u5347\u5206\u5c64\u54c1\u8cea\u7684\u5718\u968a\u3002\u5c0d\u5f37\u5316\u5b78\u7fd2\u61c9\u7528\u65bc\u8996\u89ba\u751f\u6210\u611f\u8208\u8da3\u7684\u4eba\uff0c\u4e5f\u80fd\u5f9e\u5b83\u8655\u7406\u300c\u7d44\u5167\u8b8a\u5316\u4e0d\u8db3\u300d\u7684\u8a2d\u8a08\u4e2d\u5f97\u5230\u555f\u767c\u3002<\/p>\n<p><strong>\u6548\u80fd\u8207\u8a55\u4f30\uff1a<\/strong>\u5718\u968a\u5728 Crello \u6578\u64da\u96c6\u4e0a\u6e2c\u8a66\uff0c\u7d50\u679c\u986f\u793a Stable-Layers \u76f8\u6bd4\u57fa\u790e\u6a21\u578b\uff0c\u5716\u5c64\u5206\u96e2\u66f4\u660e\u78ba\u3001\u51fa\u73fe\u7a7a\u767d\u6216\u5e36\u96dc\u8a0a\u7684\u5c64\u66f4\u5c11\uff0c\u800c\u4e14\u6bcf\u5c64\u7684\u91cd\u5efa\u8aa4\u5dee\u4e5f\u66f4\u4f4e\u3002\u8ad6\u6587\u7372 NeurIPS 2026 \u63a5\u6536\uff08arXiv:2605.30257v1\uff09\u3002<\/p>\n<p><strong>\u5f15\u7528\u7684\u6a21\u578b\uff1a<\/strong>Qwen-Image-Layered\uff08\u57fa\u790e\u5206\u5c64\u6a21\u578b\uff09\u3001Flow-GRPO\uff08\u5f37\u5316\u5b78\u7fd2\u7b97\u6cd5\uff09\u3001LoRA\uff08\u9ad8\u6548\u5fae\u8abf\u65b9\u6cd5\uff09\u3001\u8996\u89ba\u8a9e\u8a00\u6a21\u578b\u8a55\u5be9\u3002<\/p>\n<p><strong>\u9805\u76ee\uff1a<\/strong> <a href=\"https:\/\/stability-ai.github.io\/stable-layers.github.io\/\" rel=\"noopener noreferrer\">https:\/\/stability-ai.github.io\/stable-layers.github.io\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Stability AI \u63d0\u51fa Stable-Layers \u6846\u67b6\uff0c\u5229\u7528 VLM \u8a55\u5206\u8207 Flow-GRPO \u5fae\u8abf\u5716\u5c64\u5206\u89e3\u6a21\u578b\uff0c\u89e3\u6c7a\u914d\u5c0d\u6578\u64da\u4e0d\u8db3\u7684\u9650\u5236\u3002<\/p>\n","protected":false},"author":8,"featured_media":8911,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,157,120,76,160],"tags":[],"class_list":["post-8912","post","type-post","status-publish","format-standard","hentry","category-133","category-157","category-120","category-76","category-160"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8912","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=8912"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8912\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8911"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8912"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8912"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8912"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}