{"id":4667,"date":"2025-12-16T22:37:09","date_gmt":"2025-12-16T14:37:09","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4667"},"modified":"2025-12-16T22:37:09","modified_gmt":"2025-12-16T14:37:09","slug":"202512%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-pixart-%ce%b1-fast-training-of-diffusion-transformer-for-photorealistic-text-to-image-synthesis","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202512%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-pixart-%ce%b1-fast-training-of-diffusion-transformer-for-photorealistic-text-to-image-synthesis\/","title":{"rendered":"202512\u8bba\u6587\u7814\u8bfb-PIXART-\u03b1: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis"},"content":{"rendered":"\n

\u4f5c\u8005: Junsong Chen<\/p>\n\n\n\n

\u5355\u4f4d: Huawei Noah\u2019s Ark Lab, Dalian University of Technology, HKU, HKUST<\/p>\n\n\n\n

\u6765\u6e90: arXiv preprint<\/p>\n\n\n\n

\u65f6\u95f4: 2023.12<\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/h2>\n\n\n\n

1.\u73b0\u6709\u6a21\u578b\u7684\u9ad8\u6602\u8bad\u7ec3\u6210\u672c \u5c3d\u7ba1 DALL-E 2\u3001Imagen \u548c Stable Diffusion \u7b49 T2I \u6a21\u578b\u5f00\u542f\u4e86\u903c\u771f\u56fe\u50cf\u5408\u6210\u7684\u65b0\u65f6\u4ee3\uff0c\u4f46\u5b83\u4eec\u7684\u8bad\u7ec3\u9700\u8981\u5de8\u5927\u7684\u8ba1\u7b97\u8d44\u6e90 \u3002
2.\u73af\u5883\u5f71\u54cd\u4e0e\u78b3\u6392\u653e \u5de8\u5927\u7684\u8ba1\u7b97\u91cf\u5bfc\u81f4\u4e86\u5927\u91cf\u7684\u4e8c\u6c27\u5316\u78b3\uff08CO2\uff09\u6392\u653e\uff0c\u7ed9\u73af\u5883\u5e26\u6765\u538b\u529b \u3002\u4f8b\u5982\uff0cRAPHAEL \u7684\u8bad\u7ec3\u4ea7\u751f\u7684\u4e8c\u6c27\u5316\u78b3\u6392\u653e\u91cf\u9ad8\u8fbe 35 \u5428\uff0c\u76f8\u5f53\u4e8e\u4e00\u4e2a\u4eba 7 \u5e74\u7684\u6392\u653e\u91cf \u3002
3.\u963b\u788d\u521b\u65b0\u4e0e\u793e\u533a\u53d1\u5c55 \u9ad8\u6602\u7684\u6210\u672c\u548c\u8d44\u6e90\u9700\u6c42\u4e3a\u7814\u7a76\u793e\u533a\u548c\u521d\u521b\u4f01\u4e1a\u8bbe\u7acb\u4e86\u6781\u9ad8\u7684\u95e8\u69db\uff0c\u4e25\u91cd\u963b\u788d\u4e86 AIGC\uff08\u4eba\u5de5\u667a\u80fd\u751f\u6210\u5185\u5bb9\uff09\u793e\u533a\u7684\u57fa\u7840\u521b\u65b0 \u3002\u7edd\u5927\u591a\u6570\u7814\u7a76\u8005\u548c\u521b\u4e1a\u8005\u96be\u4ee5\u627f\u62c5\u4ece\u5934\u6784\u5efa\u9ad8\u8d28\u91cf\u751f\u6210\u6a21\u578b\u7684\u6210\u672c\u3002
4.\u73b0\u6709\u8bad\u7ec3\u6d41\u7a0b\u4e0e\u6570\u636e\u7684\u4f4e\u6548 \u8bba\u6587\u6307\u51fa\u5bfc\u81f4\u8bad\u7ec3\u7f13\u6162\u7684\u539f\u56e0\u4e3b\u8981\u5728\u4e8e\u8bad\u7ec3\u6d41\u7a0b\u548c\u6570\u636e\u4e24\u4e2a\u65b9\u9762\uff1a
\u4efb\u52a1\u7ea0\u7f20\uff1a \u73b0\u6709\u65b9\u6cd5\u901a\u5e38\u5c06\u50cf\u7d20\u4f9d\u8d56\u5b66\u4e60\u3001\u6587\u672c-\u56fe\u50cf\u5bf9\u9f50\u548c\u7f8e\u5b66\u8d28\u91cf\u63d0\u5347\u8fd9\u4e09\u4e2a\u4efb\u52a1\u6df7\u5408\u5728\u4e00\u8d77\uff0c\u4f7f\u7528\u6d77\u91cf\u6570\u636e\u4ece\u5934\u8bad\u7ec3\uff0c\u5bfc\u81f4\u6548\u7387\u4f4e\u4e0b \u3002
\u6570\u636e\u8d28\u91cf\u95ee\u9898\uff1a\u73b0\u6709\u7684\u6587\u672c-\u56fe\u50cf\u5bf9\u6570\u636e\u96c6\uff08\u5982 LAION\uff09\u5b58\u5728\u6587\u672c\u63cf\u8ff0\u7f3a\u5931\u3001\u8bcd\u6c47\u9891\u7387\u4f4e\u548c\u957f\u5c3e\u6548\u5e94\u7b49\u95ee\u9898\u3002\u6587\u672c\u5f80\u5f80\u53ea\u80fd\u63cf\u8ff0\u56fe\u50cf\u4e2d\u7684\u90e8\u5206\u5bf9\u8c61\uff0c\u8fd9\u79cd\u4fe1\u606f\u5bc6\u5ea6\u4f4e\u7684\u201c\u5f31\u201d\u6587\u672c-\u56fe\u50cf\u5bf9\u4e25\u91cd\u963b\u788d\u4e86\u6a21\u578b\u7684\u8bad\u7ec3\u6548\u7387\uff0c\u5bfc\u81f4\u9700\u8981\u6570\u767e\u4e07\u6b21\u8fed\u4ee3\u624d\u80fd\u5b9e\u73b0\u7a33\u5b9a\u7684\u5bf9\u9f50\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u8d21\u732e<\/h2>\n\n\n\n

\u8bad\u7ec3\u7b56\u7565\u5206\u89e3\uff1a\u5c06\u4f20\u7edf\u7684 T2I \u8bad\u7ec3\u4efb\u52a1\u5206\u89e3\u4e3a \u4e09\u4e2a\u5b50\u4efb\u52a1\uff0c\u4ece\u800c\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\uff1a\u50cf\u7d20\u4f9d\u8d56\u6027\u5b66\u4e60\uff1a\u901a\u8fc7\u4f4e\u6210\u672c\u7684\u7c7b\u6761\u4ef6\u6a21\u578b\uff08\u4f8b\u5982\u4f7f\u7528 ImageNet \u9884\u8bad\u7ec3\uff09\u6765\u6355\u83b7\u81ea\u7136\u56fe\u50cf\u7684\u50cf\u7d20\u5206\u5e03\u3002\u6587\u672c-\u56fe\u50cf\u5bf9\u9f50\uff1a\u7cbe\u786e\u5730\u5bf9\u9f50\u6587\u672c\u548c\u56fe\u50cf\uff0c\u91c7\u7528\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u7684\u6587\u672c-\u56fe\u50cf\u5bf9\u6765\u52a0\u901f\u5bf9\u9f50\u5b66\u4e60\u3002\u7f8e\u5b66\u8d28\u91cf\u63d0\u5347\uff1a\u901a\u8fc7\u4f7f\u7528\u9ad8\u8d28\u91cf\u7684\u7f8e\u5b66\u6570\u636e\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u751f\u6210\u56fe\u50cf\u7684\u827a\u672f\u611f\u3002
\u9ad8\u6548\u7684 T2I Transformer\uff1a\u57fa\u4e8e Diffusion Transformer (DiT)\uff0cPIXART-\u03b1 \u5f15\u5165\u4e86 \u8de8\u6ce8\u610f\u529b\u673a\uff08Cross-Attention\uff09 \u6765\u6ce8\u5165\u6587\u672c\u6761\u4ef6\uff0c\u540c\u65f6\u7b80\u5316\u4e86\u4f20\u7edf\u7684\u8ba1\u7b97\u5bc6\u96c6\u578b\u7c7b\u522b\u6761\u4ef6\u5206\u652f\uff0c\u51cf\u5c11\u4e86\u8ba1\u7b97\u91cf\u548c\u8d44\u6e90\u6d88\u8017\u3002
\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u6570\u636e\uff1a\u4f20\u7edf\u7684\u6587\u672c-\u56fe\u50cf\u6570\u636e\u96c6\u5b58\u5728\u6587\u672c\u63cf\u8ff0\u4e0d\u5b8c\u6574\u548c\u4fe1\u606f\u5bc6\u5ea6\u4f4e\u7684\u95ee\u9898\u3002PIXART-\u03b1 \u4f7fLLaVA \u6a21\u578b\u81ea\u52a8\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u6587\u672c\u63cf\u8ff0\uff0c\u63d0\u9ad8\u4e86\u6570\u636e\u96c6\u7684\u6982\u5ff5\u5bc6\u5ea6\uff0c\u5e2e\u52a9\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u66f4\u9ad8\u6548\u5730\u5bf9\u9f50\u6587\u672c\u548c\u56fe\u50cf\u3002
\u8bad\u7ec3\u6548\u7387\uff1aPIXART-\u03b1 \u663e\u8457\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\u3002\u4f8b\u5982\uff0cPIXART-\u03b1 \u4ec5\u7528\u4e86 12% \u7684 Stable Diffusion \u8bad\u7ec3\u65f6\u95f4\uff0c\u4e14 CO2 \u6392\u653e\u51cf\u5c11\u4e86 90%\uff0c\u6210\u672c\u5927\u5e45\u964d\u4f4e\u3002\u76f8\u6bd4 RAPHAEL \u7b49\u5927\u578b\u6a21\u578b\uff0cPIXART-\u03b1 \u7684\u8bad\u7ec3\u6210\u672c\u53ea\u6709 1%\uff0c\u4e14\u8282\u7701\u4e86 \u8fd1 300,000 \u7f8e\u5143\u3002
\u56fe\u50cf\u8d28\u91cf\uff1a\u5c3d\u7ba1\u8bad\u7ec3\u6210\u672c\u548c\u8d44\u6e90\u6d88\u8017\u5927\u5e45\u964d\u4f4e\uff0cPIXART-\u03b1 \u5728\u56fe\u50cf\u8d28\u91cf\u65b9\u9762\u4ecd\u8868\u73b0\u51fa\u8272\u3002\u5b83\u7684\u751f\u6210\u56fe\u50cf\u5728 FID \u7b49\u6307\u6807\u4e0a\u8868\u73b0\u4f18\u5f02\uff0c\u5e76\u4e14\u5728 T2I-CompBench \u4e2d\u5c55\u793a\u4e86\u5728\u6587\u672c-\u56fe\u50cf\u5bf9\u9f50\u3001\u5c5e\u6027\u7ed1\u5b9a\u3001\u5bf9\u8c61\u5173\u7cfb\u7b49\u591a\u4e2a\u7ef4\u5ea6\u7684\u4f18\u52bf\u3002
\u5b9a\u5236\u5316\u529f\u80fd\uff1aPIXART-\u03b1 \u8fd8\u652f\u6301\u4e0e DreamBooth \u548c ControlNet \u96c6\u6210\uff0c\u63d0\u4f9b\u5b9a\u5236\u5316\u56fe\u50cf\u751f\u6210\u80fd\u529b\uff0c\u53ef\u4ee5\u7cbe\u786e\u4fee\u6539\u56fe\u50cf\u7684\u989c\u8272\u3001\u98ce\u683c\u7b49\u5c5e\u6027\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u5176\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u7075\u6d3b\u6027\uff08\u4f8b\u5982\u5b9a\u5236\u7279\u5b9a\u98ce\u683c\u6216\u4fee\u6539\u5bf9\u8c61\u989c\u8272\uff09\u3002
\u603b\u7ed3\uff1aPIXART-\u03b1 \u901a\u8fc7\u5176\u521b\u65b0\u7684\u8bbe\u8ba1\uff0c\u4f7f\u5f97\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u4e0d\u4ec5\u66f4\u52a0\u9ad8\u6548\uff0c\u800c\u4e14\u5927\u5e45\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\u548c\u73af\u5883\u5f71\u54cd\u3002\u5b83\u4e3a\u7814\u7a76\u4eba\u5458\u3001\u521d\u521b\u516c\u53f8\u7b49\u63d0\u4f9b\u4e86\u4e00\u4e2a \u4f4e\u6210\u672c\u3001\u9ad8\u8d28\u91cf \u7684\u751f\u6210\u6a21\u578b\u6784\u5efa\u65b9\u6848\uff0c\u4e3a AIGC\uff08\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\u5185\u5bb9\uff09\u793e\u533a\u5e26\u6765\u4e86\u65b0\u7684\u7a81\u7834\u548c\u601d\u8def\u3002<\/p>\n\n\n\n

\u4e09\u3001\u65b9\u6cd5<\/h2>\n\n\n\n

1.\u8bad\u7ec3\u7b56\u7565\u5206\u89e3\uff08Training Strategy Decomposition\uff09
PIXART-\u03b1 \u5c06\u8bad\u7ec3\u8fc7\u7a0b\u5206\u4e3a \u4e09\u4e2a\u9636\u6bb5\uff0c\u6bcf\u4e2a\u9636\u6bb5\u805a\u7126\u4e8e\u4e0d\u540c\u7684\u4efb\u52a1\uff0c\u4ece\u800c\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\uff1a
\u9636\u6bb51\uff1a\u50cf\u7d20\u4f9d\u8d56\u6027\u5b66\u4e60
\u8be5\u9636\u6bb5\u7684\u76ee\u6807\u662f\u5b66\u4e60\u56fe\u50cf\u7684\u50cf\u7d20\u5206\u5e03\uff0c\u5e76\u901a\u8fc7\u7c7b\u6761\u4ef6\u6a21\u578b\u521d\u59cb\u5316\uff0c\u4ee5\u4f4e\u6210\u672c\u8fdb\u884c\u81ea\u7136\u56fe\u50cf\u751f\u6210\u3002
\u9636\u6bb52\uff1a\u6587\u672c-\u56fe\u50cf\u5bf9\u9f50\u5b66\u4e60
\u5728\u6b64\u9636\u6bb5\uff0c\u6a21\u578b\u7684\u76ee\u6807\u662f\u786e\u4fdd\u751f\u6210\u7684\u56fe\u50cf\u4e0e\u8f93\u5165\u6587\u672c\u7684\u63cf\u8ff0\u7cbe\u786e\u5bf9\u9f50\uff0cPIXART-\u03b1 \u901a\u8fc7\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u7684\u6587\u672c-\u56fe\u50cf\u5bf9\u6765\u52a0\u901f\u8fd9\u4e00\u8fc7\u7a0b\u3002
\u9636\u6bb53\uff1a\u9ad8\u5206\u8fa8\u7387\u548c\u7f8e\u5b66\u8d28\u91cf\u751f\u6210
\u6700\u540e\uff0c\u6a21\u578b\u901a\u8fc7\u4f7f\u7528\u9ad8\u8d28\u91cf\u7684\u7f8e\u5b66\u6570\u636e\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u3001\u89c6\u89c9\u8d28\u91cf\u4f18\u79c0\u7684\u56fe\u50cf\u3002<\/p>\n\n\n\n

2.\u9ad8\u6548\u7684 T2I Transformer\uff08Efficient T2I Transformer\uff09
PIXART-\u03b1 \u57fa\u4e8e Diffusion Transformer (DiT)\uff0c\u5e76\u5bf9\u5176\u8fdb\u884c\u4e86\u4f18\u5316\u4ee5\u63d0\u9ad8\u6548\u7387\uff1a
\u8de8\u6ce8\u610f\u529b\u673a\u5236\uff08Cross-Attention\uff09\uff1a\u6bcf\u4e2a Transformer \u5757\u4e2d\u52a0\u5165\u4e86\u591a\u5934\u8de8\u6ce8\u610f\u529b\u5c42\uff0c\u4ee5\u4fbf\u66f4\u6709\u6548\u5730\u4e0e\u6587\u672c\u5d4c\u5165\u8fdb\u884c\u4ea4\u4e92\u3002
AdaLN-single\uff1a\u901a\u8fc7\u53ea\u4f7f\u7528\u65f6\u95f4\u5d4c\u5165\uff08time embedding\uff09\u5e76\u6d88\u9664\u7c7b\u522b\u6761\u4ef6\uff0c\u51cf\u5c11\u4e86\u6a21\u578b\u7684\u53c2\u6570\u91cf\uff0c\u63d0\u5347\u4e86\u6548\u7387\u3002<\/p>\n\n\n\n

3.\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u6570\u636e\uff08High-Informative Data\uff09
PIXART-\u03b1 \u901a\u8fc7 LLaVA \u6a21\u578b\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u6587\u672c\u63cf\u8ff0\uff0c\u663e\u8457\u63d0\u9ad8\u4e86\u6570\u636e\u96c6\u7684\u6982\u5ff5\u5bc6\u5ea6\uff0c\u5e76\u51cf\u5c11\u4e86\u6b67\u4e49\uff0c\u4ece\u800c\u52a0\u901f\u4e86\u6587\u672c\u548c\u56fe\u50cf\u7684\u5bf9\u9f50\u8fc7\u7a0b\u3002
4.\u6570\u636e\u96c6\u6784\u5efa\uff08Dataset Construction\uff09
PIXART-\u03b1 \u4f7f\u7528\u4e86\u591a\u6837\u5316\u7684\u6570\u636e\u96c6\uff0c\u5305\u542b\u4e86 SAM \u6570\u636e\u96c6 \u548c\u5176\u4ed6\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u96c6\uff0c\u5982 JourneyDB \u548c \u5185\u90e8\u6570\u636e\u96c6\uff0c\u4ee5\u786e\u4fdd\u9ad8\u6548\u8bad\u7ec3\u548c\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u751f\u6210\u3002<\/p>\n\n\n\n

\u56db\u3001\u5b9e\u9a8c<\/h2>\n\n\n\n

<\/p>\n\n\n\n

1.\u5b9e\u9a8c\u51c6\u5907\u57fa\u7840\u67b6\u6784\uff1a \u91c7\u7528 DiT-XL\/2 \u4f5c\u4e3a\u57fa\u7840\u7f51\u7edc\u67b6\u6784 \u3002
\u6587\u672c\u7f16\u7801\u5668\uff1a \u4f7f\u7528 T5 large language model (4.3B Flan-T5-XXL) \u63d0\u53d6\u6587\u672c\u7279\u5f81\uff0c\u8fd9\u6bd4 CLIP \u7684\u6587\u672c\u7406\u89e3\u80fd\u529b\u66f4\u5f3a \u3002
Token \u957f\u5ea6\uff1a \u5c06\u63d0\u53d6\u7684\u6587\u672c Token \u957f\u5ea6\u8c03\u6574\u4e3a 120\uff08\u901a\u5e38\u4e3a 77\uff09\uff0c\u56e0\u4e3a\u8be5\u7814\u7a76\u4f7f\u7528\u4e86\u4fe1\u606f\u5bc6\u5ea6\u66f4\u9ad8\u7684\u5bc6\u96c6\u63cf\u8ff0\uff08Dense Captions\uff09\u3002
VAE\uff1a \u4f7f\u7528\u6765\u81ea LDM \u7684\u9884\u8bad\u7ec3\u4e14\u51bb\u7ed3\u7684 VAE \u3002
\u786c\u4ef6\u4e0e\u65f6\u95f4\uff1a \u6700\u7ec8\u6a21\u578b\u5728 64 \u5f20 V100 GPU \u4e0a\u8bad\u7ec3\u4e86\u7ea6 26 \u5929 \u3002\u6362\u7b97\u4e3a A100 GPU \u5927\u7ea6\u4e3a 753 GPU days \u3002<\/p>\n\n\n\n

2.\u6027\u80fd\u5bf9\u6bd4 (Performance Comparisons)\u5b9e\u9a8c\u4ece\u4e09\u4e2a\u4e3b\u8981\u7ef4\u5ea6\u8fdb\u884c\u4e86\u8bc4\u4f30\uff1a\u56fe\u50cf\u4fdd\u771f\u5ea6 (Fidelity)\u3001\u6587\u672c\u5bf9\u9f50 (Alignment) \u548c \u7528\u6237\u504f\u597d (User Study)\u3002
A. \u56fe\u50cf\u4fdd\u771f\u5ea6 (Fidelity) – FID \u6307\u6807 \u7ed3\u679c\uff1a PIXART-\u03b1 \u5728 MSCOCO \u6570\u636e\u96c6\u4e0a\u7684 Zero-shot FID-30K \u5f97\u5206\u4e3a 7.32 \u3002\u5bf9\u6bd4\uff1a\u4f18\u4e8e Stable Diffusion v1.5 (9.62) \u548c GigaGAN (9.09)\u3002\u63a5\u8fd1\u62e5\u6709\u5de8\u5927\u8d44\u6e90\u6d88\u8017\u7684 RAPHAEL (6.61)\u3002\u6548\u7387\u4f18\u52bf\uff1a \u8fbe\u5230\u8fd9\u4e00\u6027\u80fd\u4ec5\u6d88\u8017\u4e86 SD v1.5 12% \u7684\u8bad\u7ec3\u65f6\u95f4\uff0c\u4ee5\u53ca RAPHAEL 1% \u7684\u8bad\u7ec3\u6210\u672c\uff08\u6570\u636e\u91cf\u4ec5\u4e3a RAPHAEL \u7684 0.5%\uff09\u3002
B. \u6587\u672c\u5bf9\u9f50 (Alignment) – T2I-CompBench \u6d4b\u8bd5\u57fa\u51c6\uff1a \u4f7f\u7528 T2I-CompBench \u8bc4\u4f30\u7ec4\u5408\u751f\u6210\u80fd\u529b\uff08\u5982\u5c5e\u6027\u7ed1\u5b9a\u3001\u7269\u4f53\u5173\u7cfb\u3001\u590d\u6742\u7ec4\u5408\u7b49\uff09\u3002\u7ed3\u679c\uff1a PIXART-\u03b1 \u5728 6 \u9879\u8bc4\u4f30\u6307\u6807\u4e2d\u6709 5 \u9879 \u8868\u73b0\u4f18\u5f02\uff08\u5982\u989c\u8272\u7ed1\u5b9a\u3001\u5f62\u72b6\u7ed1\u5b9a\u3001\u7a7a\u95f4\u5173\u7cfb\u7b49\uff09\u3002\u4f18\u52bf\uff1a \u4f18\u4e8e SDXL \u548c DALL-E 2\uff0c\u8fd9\u5f52\u529f\u4e8e\u7b2c\u4e8c\u9636\u6bb5\u8bad\u7ec3\u4e2d\u4f7f\u7528\u4e86\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u7684\u7cbe\u786e\u56fe\u6587\u5bf9\uff0c\u6781\u5927\u5730\u63d0\u5347\u4e86\u7ec6\u7c92\u5ea6\u7684\u8bed\u4e49\u5bf9\u9f50\u80fd\u529b \u3002
C. \u7528\u6237\u7814\u7a76 (User Study)\u65b9\u6cd5\uff1a \u9009\u53d6 300 \u4e2a\u56fa\u5b9a\u63d0\u793a\u8bcd\uff0c\u9080\u8bf7 50 \u540d\u8bc4\u4f30\u8005\u5bf9\u6bd4 PIXART-\u03b1 \u4e0e DALL-E 2\u3001SDv2\u3001SDXL \u548c DeepFloyd \u7684\u751f\u6210\u7ed3\u679c \u3002
\u7ed3\u679c\uff1a\u5728\u611f\u77e5\u8d28\u91cf (Quality) \u548c\u8bed\u4e49\u5bf9\u9f50 (Alignment) \u4e0a\uff0cPIXART-\u03b1 \u5747\u4f18\u4e8e\u5bf9\u6bd4\u6a21\u578b \u3002
\u4f8b\u5982\uff0c\u76f8\u6bd4 SDv2\uff0cPIXART-\u03b1 \u5728\u56fe\u50cf\u8d28\u91cf\u4e0a\u63d0\u5347\u4e86 7.2%\uff0c\u5728\u5bf9\u9f50\u6027\u4e0a\u5927\u5e45\u63d0\u5347\u4e86 42.4% \u3002<\/p>\n\n\n\n

\u6d88\u878d\u5b9e\u9a8c (Ablation Study)
\u4e3a\u4e86\u9a8c\u8bc1\u67b6\u6784\u6539\u8fdb\u7684\u6709\u6548\u6027\uff0c\u4f5c\u8005\u91cd\u70b9\u5bf9\u6bd4\u4e86\u4e0d\u540c\u7684\u8bbe\u8ba1\u9009\u62e9\uff1a
\u91cd\u53c2\u6570\u5316 (Re-parameterization) \u7684\u5fc5\u8981\u6027\uff1a
\u5982\u679c\u4e0d\u4f7f\u7528\u91cd\u53c2\u6570\u5316\uff08w\/o re-param\uff09\uff0c\u6a21\u578b\u65e0\u6cd5\u6709\u6548\u5229\u7528 ImageNet \u9884\u8bad\u7ec3\u6743\u91cd\uff0c\u5bfc\u81f4\u751f\u6210\u7684\u56fe\u50cf\u626d\u66f2\u4e14\u7f3a\u4e4f\u7ec6\u8282 \u3002
adaLN-single vs. adaLN\uff1a
\u6548\u679c\uff1a \u8bba\u6587\u63d0\u51fa\u7684 adaLN-single \u7ed3\u6784\u5728\u89c6\u89c9\u6548\u679c\u4e0a\u4e0e\u6807\u51c6\u7684 adaLN \u76f8\u5f53 \u3002
\u6548\u7387\uff1a adaLN-single \u51cf\u5c11\u4e86 21% \u7684\u663e\u5b58\u5360\u7528\uff0829G -> 23G\uff09\u548c 26% \u7684\u53c2\u6570\u91cf\uff08833M -> 611M\uff09\u3002\u8fd9\u8bc1\u660e\u4e86\u7cbe\u7b80\u67b6\u6784\u662f\u9ad8\u6548\u8bad\u7ec3\u7684\u5173\u952e\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
    \n
  1. \u6269\u5c55\u5e94\u7528 (Application Extensions)
    \u5b9e\u9a8c\u8fd8\u5c55\u793a\u4e86 PIXART-\u03b1 \u826f\u597d\u7684\u6269\u5c55\u6027\u548c\u517c\u5bb9\u6027\uff0c\u7c7b\u4f3c\u4e8e Stable Diffusion \u751f\u6001\uff1a<\/li>\n<\/ol>\n\n\n\n

    DreamBooth (\u4e2a\u6027\u5316\u5b9a\u5236):
    \u6a21\u578b\u53ef\u4ee5\u8f7b\u677e\u7ed3\u5408 DreamBooth \u8fdb\u884c\u5fae\u8c03\u3002\u5b9e\u9a8c\u5c55\u793a\u4e86\u4ec5\u9700\u51e0\u5f20\u56fe\u7247\uff0c\u5c31\u80fd\u751f\u6210\u7279\u5b9a\u7269\u4f53\uff08\u5982\u7279\u5b9a\u7684\u72d7\u6216\u6c7d\u8f66\uff09\u5728\u4e0d\u540c\u573a\u666f\u4e0b\u7684\u9ad8\u4fdd\u771f\u56fe\u50cf\uff0c\u751a\u81f3\u6539\u53d8\u7269\u4f53\u989c\u8272 \u3002<\/p>\n\n\n\n

    ControlNet (\u7ed3\u6784\u63a7\u5236):
    \u901a\u8fc7\u51bb\u7ed3 DiT \u5757\u5e76\u6dfb\u52a0\u53ef\u8bad\u7ec3\u526f\u672c\uff08ControlNet \u67b6\u6784\uff09\uff0c\u6a21\u578b\u80fd\u591f\u63a5\u53d7\u8fb9\u7f18\u56fe\uff08HED Edge\uff09\u7b49\u4f5c\u4e3a\u63a7\u5236\u4fe1\u53f7\u3002\u5b9e\u9a8c\u663e\u793a PIXART-\u03b1 \u80fd\u7cbe\u51c6\u5730\u6839\u636e\u8fb9\u7f18\u56fe\u751f\u6210\u5bf9\u5e94\u7ed3\u6784\u7684\u56fe\u50cf \u3002<\/p>\n\n\n\n

    \u4e94\u3001\u603b\u7ed3<\/h2>\n\n\n\n

    \u603b\u7ed3\uff1aPIXART-\u03b1 \u63d0\u51fa\u4e86\u4e00\u4e2a\u65b0\u7684\u9ad8\u6548\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u6846\u67b6\uff0c\u901a\u8fc7\u521b\u65b0\u7684 \u8bad\u7ec3\u7b56\u7565\u5206\u89e3\u3001\u9ad8\u6548\u7684 Transformer \u67b6\u6784 \u548c \u9ad8\u4fe1\u606f\u5bc6\u5ea6\u6570\u636e\uff0c\u5728\u663e\u8457\u964d\u4f4e\u8bad\u7ec3\u6210\u672c\u548c\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u7684\u540c\u65f6\uff0c\u4ecd\u80fd\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf\u3002\u5b83\u7684\u9ad8\u6548\u8bad\u7ec3\u65b9\u5f0f\u548c\u5353\u8d8a\u7684\u751f\u6210\u80fd\u529b\u4f7f\u5176\u6210\u4e3a\u7814\u7a76\u4eba\u5458\u548c\u521d\u521b\u516c\u53f8\u6784\u5efa\u4f4e\u6210\u672c\u9ad8\u8d28\u91cf\u751f\u6210\u6a21\u578b\u7684\u6709\u529b\u5de5\u5177\u3002
    \u672a\u6765\u65b9\u5411\uff1a\u4f5c\u8005\u8fd8\u6307\u51fa\uff0c\u672a\u6765\u53ef\u4ee5\u901a\u8fc7\u6269\u5c55 PIXART-\u03b1 \u7684\u89c4\u6a21\u548c\u6027\u80fd\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u751f\u6210\u8d28\u91cf\uff0c\u5e76\u9488\u5bf9\u751f\u6210\u7ec6\u8282\u548c\u6587\u672c\u751f\u6210\u80fd\u529b\u8fdb\u884c\u6539\u8fdb\u3002\u6b64\u5916\uff0cPIXART-\u03b1 \u4e3a\u56fe\u50cf\u751f\u6210\u4e2d\u7684 \u5b9a\u5236\u5316\u529f\u80fd \u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u548c\u5de5\u5177\uff0c\u672a\u6765\u53ef\u4ee5\u63a2\u7d22\u66f4\u5e7f\u6cdb\u7684\u5e94\u7528\u573a\u666f\u3002<\/p>\n\n\n\n

    \u516d\u3001\u5bf9\u5176\u601d\u8003<\/h2>\n\n\n\n

    1.0 \u6280\u672f\u521b\u65b0\u2014\u2014\u8bad\u7ec3\u7b56\u7565\u89e3\u8026\u4e0e\u9ad8\u6548\u67b6\u6784\u6846\u67b6\u3002 PIXART-\u03b1 \u7684\u6838\u5fc3\u5728\u4e8e\u201c\u4efb\u52a1\u62c6\u89e3 + \u67b6\u6784\u7cbe\u7b80 + \u6570\u636e\u589e\u5bc6\u201d\u7684\u9ad8\u6548\u8bad\u7ec3\u8303\u5f0f\u3002 \u67b6\u6784\u7cbe\u7b80\uff08Efficient DiT\uff09\uff1a \u5728 DiffusionTransformer\uff08DiT)\u57fa\u7840\u4e0a\u5f15\u5165 Cross-Attention \u6ce8\u5165\u6587\u672c\u6761\u4ef6\uff0c\u5e76\u521b\u65b0\u6027\u63d0\u51fa adaLN-single \u6a21\u5757\uff0c\u5728\u4fdd\u6301\u751f\u6210\u6548\u679c\u7684\u540c\u65f6\u51cf\u5c11\u4e86\u7ea6 26% \u7684\u53c2\u6570\u91cf\u548c 21% \u7684\u663e\u5b58\u5360\u7528\u3002\u6570\u636e\u589e\u5bc6\uff1a\u5229\u7528\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08LLaVA\uff09\u81ea\u52a8\u6807\u6ce8 SAM \u6570\u636e\u96c6\uff0c\u751f\u6210\u9ad8\u4fe1\u606f\u5bc6\u5ea6\u7684\u4f2a\u63cf\u8ff0\uff0c\u89e3\u51b3\u4e86\u4f20\u7edf\u6570\u636e\u63cf\u8ff0\u7a00\u758f\u5bfc\u81f4\u7684\u5bf9\u9f50\u6548\u7387\u4f4e\u95ee\u9898 \u3002
    2.0 \u6280\u672f\u76ee\u6807\u2014\u2014\u4f4e\u6210\u672c\u9ad8\u4fdd\u771f\u751f\u6210\u3002 PIXART-\u03b1 \u7684\u6280\u672f\u76ee\u6807\u662f\u5728\u6781\u4f4e\u7684\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u4e0b\uff0c\u5b9e\u73b0\u5ab2\u7f8e\u751a\u81f3\u8d85\u8d8a SOTA \u5546\u4e1a\u7ea7\u6a21\u578b\u7684\u56fe\u50cf\u751f\u6210\u8d28\u91cf\u3002 \u5176\u6838\u5fc3\u5728\u4e8e\u6253\u7834\u201c\u9ad8\u8d28\u91cf=\u9ad8\u6210\u672c\u201d\u7684\u5b9a\u5f0f\uff0c\u5c06\u8bad\u7ec3\u6210\u672c\u964d\u4f4e\u81f3\u540c\u7c7b\u6a21\u578b\u7684 1% \u5de6\u53f3 \u3002
    3.0 \u5e94\u7528\u573a\u666f\u2014\u2014\u53ef\u63a7\u4e0e\u5b9a\u5236\u5316\u56fe\u50cf\u5408\u6210\u3002PIXART-\u03b1 \u4e0d\u4ec5\u652f\u6301\u539f\u751f\u7684\u9ad8\u5206\u8fa8\u7387\uff08\u6700\u9ad8 1024\u00d71024\uff09\u5199\u5b9e\u56fe\u50cf\u751f\u6210 \uff0c\u8fd8\u5c55\u73b0\u4e86\u6781\u5f3a\u7684\u4e0b\u6e38\u4efb\u52a1\u6269\u5c55\u80fd\u529b\u3002 \u5b83\u53ef\u4ee5\u65e0\u7f1d\u7ed3\u5408 DreamBooth\uff0c\u4ec5\u9700\u5c11\u91cf\u6837\u672c\u5373\u53ef\u5b9e\u73b0\u7279\u5b9a\u4e3b\u4f53\u7684\u9ad8\u4fdd\u771f\u4e2a\u6027\u5316\u751f\u6210\u4e0e\u5c5e\u6027\u4fee\u6539\uff1b\u540c\u65f6\u652f\u6301\u7ed3\u5408 ControlNet\uff0c\u901a\u8fc7\u8fb9\u7f18\u68c0\u6d4b\u56fe\u7b49\u6761\u4ef6\u4fe1\u53f7\uff0c\u7cbe\u51c6\u63a7\u5236\u751f\u6210\u56fe\u50cf\u7684\u7a7a\u95f4\u7ed3\u6784\u4e0e\u5e03\u5c40\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u4f5c\u8005: Junsong Chen \u5355\u4f4d: Huawei Noah\u2019s Ark Lab, Dalian Univ […]<\/p>\n","protected":false},"author":73,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4667","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4667","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/73"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4667"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4667\/revisions"}],"predecessor-version":[{"id":4678,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4667\/revisions\/4678"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4667"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4667"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4667"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}