{"id":6454,"date":"2023-03-08T16:13:16","date_gmt":"2023-03-08T08:13:16","guid":{"rendered":"https:\/\/i007.vip:15443\/wordpress\/?p=6454"},"modified":"2023-03-08T16:13:16","modified_gmt":"2023-03-08T08:13:16","slug":"stable-diffusion%e7%9a%84%e5%85%a5%e9%97%a8%e4%bb%8b%e7%bb%8d%e5%92%8c%e4%bd%bf%e7%94%a8%e6%95%99%e7%a8%8b","status":"publish","type":"post","link":"https:\/\/i007.cc\/wordpress\/archives\/6454","title":{"rendered":"Stable Diffusion\u7684\u5165\u95e8\u4ecb\u7ecd\u548c\u4f7f\u7528\u6559\u7a0b"},"content":{"rendered":"
\n

\u539f\u6587\u5730\u5740<\/a><\/p>\n

\u00a0<\/p>\n

<\/a>\u6570\u636e\u6d3eTHU<\/span>\u00a02023-01-02 17:30<\/em>\u00a0Posted on\u00a0\u5317\u4eac<\/span><\/em><\/div>\n
<\/p>\n

\"Image\"<\/p>\n

\n
\n
\n
\n
\u6765\u6e90\uff1aDeepHuhb IMBA<\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n
\n
\n
\u672c\u6587\u7ea64800\u5b57<\/strong>\uff0c\u5efa\u8bae\u9605\u8bfb9<\/strong>\u5206\u949f<\/strong><\/section>\n
\u672c\u6587\u4ecb\u7ecd\u4e86Stable Diffusion\u662f\u4e00\u4e2a\u6587\u672c\u5230\u56fe\u50cf\u7684\u6f5c\u5728\u6269\u6563\u6a21\u578b\u7684\u5165\u95e8\u4ecb\u7ecd\u53ca\u4f7f\u7528\u6559\u7a0b\u3002<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\u00a0<\/p>\n

Stable Diffusion\u662f\u4e00\u4e2a\u6587\u672c\u5230\u56fe\u50cf\u7684\u6f5c\u5728\u6269\u6563\u6a21\u578b\uff0c\u7531CompVis\u3001Stability AI\u548cLAION\u7684\u7814\u7a76\u4eba\u5458\u548c\u5de5\u7a0b\u5e08\u521b\u5efa\u3002\u5b83\u4f7f\u7528\u6765\u81eaLAION-5B\u6570\u636e\u5e93\u5b50\u96c6\u7684512×512\u56fe\u50cf\u8fdb\u884c\u8bad\u7ec3\u3002\u4f7f\u7528\u8fd9\u4e2a\u6a21\u578b\uff0c\u53ef\u4ee5\u751f\u6210\u5305\u62ec\u4eba\u8138\u5728\u5185\u7684\u4efb\u4f55\u56fe\u50cf\uff0c\u56e0\u4e3a\u6709\u5f00\u6e90\u7684\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u6240\u4ee5\u6211\u4eec\u4e5f\u53ef\u4ee5\u5728\u81ea\u5df1\u7684\u673a\u5668\u4e0a\u8fd0\u884c\u5b83\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n

\"Image\"<\/p>\n

\u5982\u679c\u4f60\u8db3\u591f\u806a\u660e\u548c\u6709\u521b\u9020\u529b\uff0c\u4f60\u53ef\u4ee5\u521b\u9020\u4e00\u7cfb\u5217\u7684\u56fe\u50cf\uff0c\u7136\u540e\u5f62\u6210\u4e00\u4e2a\u89c6\u9891\u3002\u4f8b\u5982\uff0cXander Steenbrugge\u4f7f\u7528\u5b83\u548c\u4e0a\u56fe\u6240\u793a\u7684\u8f93\u5165\u63d0\u793a\u521b\u5efa\u4e86\u4ee4\u4eba\u60ca\u53f9\u7684\u300a\u7a7f\u8d8a\u65f6\u95f4\u300b\u3002<\/p>\n

<\/section>\n
\n
\n
\n
<\/div>\n<\/div>\n
\n
\n
<\/div>\n
deephub\u6df1\u5ea6\u5b66\u4e60<\/div>\n<\/div>\n
\n
<\/div>\n
\uff0c\u8d5e<\/span>32<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/section>\n

\u4ee5\u4e0b\u662f\u4ed6\u7528\u6765\u521b\u4f5c\u8fd9\u5e45\u521b\u9020\u6027\u827a\u672f\u4f5c\u54c1\u7684\u7075\u611f\u548c\u6587\u672c:<\/p>\n

\"Image\"<\/p>\n

\u672c\u6587\u9996\u5148\u4ecb\u7ecd\u4ec0\u4e48\u662fStable Diffusion\uff0c\u5e76\u8ba8\u8bba\u5b83\u7684\u4e3b\u8981\u7ec4\u6210\u90e8\u5206\u3002\u7136\u540e\u6211\u4eec\u5c06\u4f7f\u7528\u6a21\u578b\u4ee5\u4e09\u79cd\u4e0d\u540c\u7684\u65b9\u5f0f\u521b\u5efa\u56fe\u50cf\uff0c\u8fd9\u4e09\u79cd\u65b9\u5f0f\u4ece\u66f4\u7b80\u5355\u5230\u590d\u6742\u3002<\/p>\n

Stable Diffusion<\/strong><\/h2>\n

Stable Diffusion\u662f\u4e00\u79cd\u673a\u5668\u5b66\u4e60\u6a21\u578b\uff0c\u5b83\u7ecf\u8fc7\u8bad\u7ec3\u53ef\u4ee5\u9010\u6b65\u5bf9\u968f\u673a\u9ad8\u65af\u566a\u58f0\u8fdb\u884c\u53bb\u566a\u4ee5\u83b7\u5f97\u611f\u5174\u8da3\u7684\u6837\u672c\uff0c\u4f8b\u5982\u751f\u6210\u56fe\u50cf\u3002<\/p>\n

\u6269\u6563\u6a21\u578b\u6709\u4e00\u4e2a\u4e3b\u8981\u7684\u7f3a\u70b9\u5c31\u662f\u53bb\u566a\u8fc7\u7a0b\u7684\u65f6\u95f4\u548c\u5185\u5b58\u6d88\u8017\u90fd\u975e\u5e38\u6602\u8d35\u3002\u8fd9\u4f1a\u4f7f\u8fdb\u7a0b\u53d8\u6162\uff0c\u5e76\u6d88\u8017\u5927\u91cf\u5185\u5b58\u3002\u4e3b\u8981\u539f\u56e0\u662f\u5b83\u4eec\u5728\u50cf\u7d20\u7a7a\u95f4\u4e2d\u8fd0\u884c\uff0c\u7279\u522b\u662f\u5728\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u3002<\/p>\n

Latent diffusion\u901a\u8fc7\u5728\u8f83\u4f4e\u7ef4\u5ea6\u7684\u6f5c\u7a7a\u95f4\u4e0a\u5e94\u7528\u6269\u6563\u8fc7\u7a0b\u800c\u4e0d\u662f\u4f7f\u7528\u5b9e\u9645\u7684\u50cf\u7d20\u7a7a\u95f4\u6765\u51cf\u5c11\u5185\u5b58\u548c\u8ba1\u7b97\u6210\u672c\u3002\u6240\u4ee5Stable Diffusion\u5f15\u5165\u4e86Latent diffusion\u7684\u65b9\u5f0f\u6765\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\u8ba1\u7b97\u4ee3\u4ef7\u6602\u8d35\u7684\u95ee\u9898\u3002<\/p>\n

1\u3001Latent diffusion\u7684\u4e3b\u8981\u7ec4\u6210\u90e8\u5206<\/strong><\/p>\n

Latent diffusion\u6709\u4e09\u4e2a\u4e3b\u8981\u7ec4\u6210\u90e8\u5206:<\/p>\n

\u81ea\u52a8\u7f16\u7801\u5668(VAE)<\/strong><\/p>\n

\u81ea\u52a8\u7f16\u7801\u5668(VAE)\u7531\u4e24\u4e2a\u4e3b\u8981\u90e8\u5206\u7ec4\u6210:\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u3002\u7f16\u7801\u5668\u5c06\u628a\u56fe\u50cf\u8f6c\u6362\u6210\u4f4e\u7ef4\u7684\u6f5c\u5728\u8868\u793a\u5f62\u5f0f\uff0c\u8be5\u8868\u793a\u5f62\u5f0f\u5c06\u4f5c\u4e3a\u4e0b\u4e00\u4e2a\u7ec4\u4ef6U_Net\u7684\u8f93\u5165\u3002\u89e3\u7801\u5668\u5c06\u505a\u76f8\u53cd\u7684\u4e8b\u60c5\uff0c\u5b83\u5c06\u628a\u6f5c\u5728\u7684\u8868\u793a\u8f6c\u6362\u56de\u56fe\u50cf\u3002<\/p>\n

\u5728Latent diffusion\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u5229\u7528\u7f16\u7801\u5668\u83b7\u5f97\u6b63\u5411\u6269\u6563\u8fc7\u7a0b\u4e2d\u8f93\u5165\u56fe\u50cf\u7684\u6f5c\u8868\u793a(latent)\u3002\u800c\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0cVAE\u89e3\u7801\u5668\u5c06\u628a\u6f5c\u4fe1\u53f7\u8f6c\u6362\u56de\u56fe\u50cf\u3002<\/p>\n

U-Net<\/strong><\/p>\n

\"Image\"<\/p>\n

U-Net\u4e5f\u5305\u62ec\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u4e24\u90e8\u5206\uff0c\u4e24\u8005\u90fd\u7531ResNet\u5757\u7ec4\u6210\u3002\u7f16\u7801\u5668\u5c06\u56fe\u50cf\u8868\u793a\u538b\u7f29\u4e3a\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf\uff0c\u89e3\u7801\u5668\u5c06\u4f4e\u5206\u8fa8\u7387\u89e3\u7801\u56de\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u3002<\/p>\n

\u4e3a\u4e86\u9632\u6b62U-Net\u5728\u4e0b\u91c7\u6837\u65f6\u4e22\u5931\u91cd\u8981\u4fe1\u606f\uff0c\u901a\u5e38\u5728\u7f16\u7801\u5668\u7684\u4e0b\u91c7\u6837\u7684ResNet\u548c\u89e3\u7801\u5668\u7684\u4e0a\u91c7\u6837ResNet\u4e4b\u95f4\u6dfb\u52a0\u4e86\u6377\u5f84\u7684\u8fde\u63a5\u3002<\/p>\n

\u5728Stable Diffusion\u7684U-Net\u4e2d\u6dfb\u52a0\u4e86\u4ea4\u53c9\u6ce8\u610f\u5c42\u5bf9\u6587\u672c\u5d4c\u5165\u7684\u8f93\u51fa\u8fdb\u884c\u8c03\u8282\u3002\u4ea4\u53c9\u6ce8\u610f\u5c42\u88ab\u6dfb\u52a0\u5230U-Net\u7684\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668ResNet\u5757\u4e4b\u95f4\u3002<\/p>\n

Text-Encoder<\/strong><\/p>\n

\"Image\"<\/p>\n

\u6587\u672c\u7f16\u7801\u5668\u5c06\u628a\u8f93\u5165\u6587\u5b57\u63d0\u793a\u8f6c\u6362\u4e3aU-Net\u53ef\u4ee5\u7406\u89e3\u7684\u5d4c\u5165\u7a7a\u95f4\uff0c\u8fd9\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u57fa\u4e8etransformer\u7684\u7f16\u7801\u5668\uff0c\u5b83\u5c06\u6807\u8bb0\u5e8f\u5217\u6620\u5c04\u5230\u6f5c\u5728\u6587\u672c\u5d4c\u5165\u5e8f\u5217\u3002\u4ece\u8fd9\u91cc\u53ef\u4ee5\u770b\u5230\u4f7f\u7528\u826f\u597d\u7684\u6587\u5b57\u63d0\u793a\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u9884\u671f\u8f93\u51fa\u3002<\/p>\n

\u4e3a\u4ec0\u4e48Latent Diffusion\u5feb\u901f\u6709\u6548<\/strong><\/p>\n

Latent Diffusion\u4e4b\u6240\u4ee5\u5feb\u901f\u6709\u6548\uff0c\u662f\u56e0\u4e3a\u5b83\u7684U-Net\u662f\u5728\u4f4e\u7ef4\u7a7a\u95f4\u4e0a\u5de5\u4f5c\u7684\u3002\u4e0e\u50cf\u7d20\u7a7a\u95f4\u6269\u6563\u76f8\u6bd4\uff0c\u8fd9\u964d\u4f4e\u4e86\u5185\u5b58\u548c\u8ba1\u7b97\u590d\u6742\u5ea6\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a(3,512,512)\u7684\u56fe\u50cf\u5728\u6f5c\u5728\u7a7a\u95f4\u4e2d\u4f1a\u53d8\u6210(4,64,64)\uff0c\u5185\u5b58\u5c06\u4f1a\u51cf\u5c1164\u500d\u3002<\/p>\n

Stable Diffusion\u7684\u63a8\u7406\u8fc7\u7a0b<\/strong><\/p>\n

\"Image\"<\/p>\n

\u9996\u5148\uff0c\u6a21\u578b\u5c06\u6f5c\u5728\u7a7a\u95f4\u7684\u968f\u673a\u79cd\u5b50\u548c\u6587\u672c\u63d0\u793a\u540c\u65f6\u4f5c\u4e3a\u8f93\u5165\u3002\u7136\u540e\u4f7f\u7528\u6f5c\u5728\u7a7a\u95f4\u7684\u79cd\u5b50\u751f\u6210\u5927\u5c0f\u4e3a64\u00d764\u7684\u968f\u673a\u6f5c\u5728\u56fe\u50cf\u8868\u793a\uff0c\u901a\u8fc7CLIP\u7684\u6587\u672c\u7f16\u7801\u5668\u5c06\u8f93\u5165\u7684\u6587\u672c\u63d0\u793a\u8f6c\u6362\u4e3a\u5927\u5c0f\u4e3a77\u00d7768\u7684\u6587\u672c\u5d4c\u5165\u3002<\/p>\n

\u7136\u540e\uff0c\u4f7f\u7528U-Net \u5728\u4ee5\u6587\u672c\u5d4c\u5165\u4e3a\u6761\u4ef6\u7684\u540c\u65f6\u8fed\u4ee3\u5730\u5bf9\u968f\u673a\u6f5c\u5728\u56fe\u50cf\u8868\u793a\u8fdb\u884c\u53bb\u566a\u3002U-Net \u7684\u8f93\u51fa\u662f\u566a\u58f0\u7684\u6b8b\u5dee\uff0c\u7528\u4e8e\u901a\u8fc7scheduler \u7a0b\u5e8f\u7b97\u6cd5\u8ba1\u7b97\u53bb\u566a\u7684\u6f5c\u5728\u56fe\u50cf\u8868\u793a\u3002scheduler \u7b97\u6cd5\u6839\u636e\u5148\u524d\u7684\u566a\u58f0\u8868\u793a\u548c\u9884\u6d4b\u7684\u566a\u58f0\u6b8b\u5dee\u8ba1\u7b97\u9884\u6d4b\u7684\u53bb\u566a\u56fe\u50cf\u8868\u793a\u3002<\/p>\n

\u8bb8\u591a\u4e0d\u540c\u7684scheduler \u7b97\u6cd5\u53ef\u4ee5\u7528\u4e8e\u8fd9\u4e2a\u8ba1\u7b97\uff0c\u6bcf\u4e00\u4e2a\u90fd\u6709\u5b83\u7684\u4f18\u70b9\u548c\u7f3a\u70b9\u3002\u5bf9\u4e8eStable Diffusion\uff0c\u5efa\u8bae\u4f7f\u7528\u4ee5\u4e0b\u5176\u4e2d\u4e4b\u4e00:<\/p>\n

    \n
  • PNDM scheduler (\u9ed8\u8ba4)<\/li>\n
  • DDIM scheduler<\/li>\n
  • K-LMS scheduler<\/li>\n<\/ul>\n

    \u53bb\u566a\u8fc7\u7a0b\u91cd\u590d\u7ea650\u6b21\uff0c\u8fd9\u6837\u53ef\u4ee5\u9010\u6b65\u68c0\u7d22\u66f4\u597d\u7684\u6f5c\u5728\u56fe\u50cf\u8868\u793a\u3002\u4e00\u65e6\u5b8c\u6210\uff0c\u6f5c\u5728\u56fe\u50cf\u8868\u793a\u5c31\u4f1a\u7531\u53d8\u5206\u81ea\u7f16\u7801\u5668\u7684\u89e3\u7801\u5668\u90e8\u5206\u8fdb\u884c\u89e3\u7801\u3002<\/p>\n

    \u4f7f\u7528 Hugging Face\u7684API<\/strong><\/h2>\n

    Hugging Face\u63d0\u4f9b\u4e86\u4e00\u4e2a\u975e\u5e38\u7b80\u5355\u7684API\u6765\u4f7f\u7528\u6211\u4eec\u7684\u6a21\u578b\u751f\u6210\u56fe\u50cf\u3002\u5728\u4e0b\u56fe\u4e2d\u53ef\u4ee5\u770b\u5230\u6211\u4f7f\u7528\u4e86\u201castronaut riding a horse\u201d\u4f5c\u4e3a\u8f93\u5165\u5f97\u5230\u8f93\u51fa\u56fe\u50cf:<\/p>\n

    \"Image\"<\/p>\n

    \u4ed6\u63d0\u4f9b\u7684\u6a21\u578b\u8fd8\u5305\u542b\u4e86\u4e00\u4e9b\u53ef\u7528\u7684\u9ad8\u7ea7\u9009\u9879\u6765\u6539\u53d8\u751f\u6210\u7684\u56fe\u50cf\u7684\u8d28\u91cf\uff0c\u5982\u4e0b\u56fe\u6240\u793a:<\/p>\n

    \"Image\"<\/p>\n

    \u8fd9\u91cc\u7684\u56db\u4e2a\u9009\u9879\u8bf4\u660e\u5982\u4e0b:<\/p>\n

    images:\u8be5\u9009\u9879\u63a7\u5236\u7684\u751f\u6210\u56fe\u50cf\u6570\u91cf\u6700\u591a\u4e3a4\u4e2a\u3002<\/p>\n

    Steps:\u6b64\u9009\u9879\u9009\u62e9\u60f3\u8981\u7684\u6269\u6563\u8fc7\u7a0b\u7684\u6b65\u9aa4\u6570\u3002\u6b65\u9aa4\u8d8a\u591a\uff0c\u751f\u6210\u7684\u56fe\u50cf\u8d28\u91cf\u8d8a\u597d\u3002\u5982\u679c\u60f3\u8981\u9ad8\u8d28\u91cf\uff0c\u53ef\u4ee5\u9009\u62e9\u53ef\u7528\u7684\u6700\u5927\u6b65\u9aa4\u6570\uff0c\u537350\u3002\u5982\u679c\u4f60\u60f3\u8981\u66f4\u5feb\u7684\u7ed3\u679c\uff0c\u90a3\u4e48\u8003\u8651\u51cf\u5c11\u6b65\u9aa4\u7684\u6570\u91cf\u3002<\/p>\n

    Guidance Scale:Guidance Scale\u662f\u751f\u6210\u7684\u56fe\u50cf\u4e0e\u8f93\u5165\u63d0\u793a\u7684\u7d27\u5bc6\u7a0b\u5ea6\u4e0e\u8f93\u5165\u7684\u591a\u6837\u6027\u4e4b\u95f4\u7684\u6743\u8861\u3002\u5b83\u7684\u5178\u578b\u503c\u57287.5\u5de6\u53f3\u3002\u589e\u52a0\u7684\u6bd4\u4f8b\u8d8a\u591a\uff0c\u56fe\u50cf\u7684\u8d28\u91cf\u5c31\u4f1a\u8d8a\u9ad8\uff0c\u4f46\u662f\u4f60\u5f97\u5230\u7684\u8f93\u51fa\u5c31\u4f1a\u8d8a\u5c11\u3002<\/p>\n

    Seed:\u968f\u673a\u79cd\u5b50\u591f\u63a7\u5236\u751f\u6210\u7684\u6837\u672c\u7684\u591a\u6837\u6027\u3002<\/p>\n

    \u4f7f\u7528Diffuser \u5305<\/strong><\/h2>\n

    \u7b2c\u4e8c\u79cd\u4f7f\u7528\u7684\u65b9\u6cd5\u662f\u4f7f\u7528Hugging Face\u7684Diffusers\u5e93\uff0c\u5b83\u5305\u542b\u4e86\u76ee\u524d\u53ef\u7528\u7684\u5927\u90e8\u5206\u7a33\u5b9a\u6269\u6563\u6a21\u578b\uff0c\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u5728\u8c37\u6b4c\u7684Colab\u4e0a\u8fd0\u884c\u5b83\u3002<\/p>\n

    \u7b2c\u4e00\u6b65\u662f\u6253\u5f00\u8c37\u6b4ccollab\uff0c\u68c0\u67e5\u662f\u5426\u8fde\u63a5\u5230GPU\uff0c\u53ef\u4ee5\u5728\u8d44\u6e90\u6309\u94ae\u4e2d\u67e5\u770b\uff0c\u5982\u4e0b\u56fe\u6240\u793a:<\/p>\n

    \"Image\"<\/p>\n

    \u53e6\u4e00\u4e2a\u9009\u62e9\u662f\u4ece\u8fd0\u884c\u65f6\u83dc\u5355\u4e2d\u9009\u62e9\u66f4\u6539\u8fd0\u884c\u65f6\u7c7b\u578b\uff0c\u7136\u540e\u68c0\u67e5\u786c\u4ef6\u52a0\u901f\u5668\u88ab\u9009\u62e9\u4e3aGPU:<\/p>\n

    \"Image\"<\/p>\n

    \u6211\u4eec\u786e\u4fdd\u4f7f\u7528GPU\u8fd0\u884c\u65f6\u540e\uff0c\u4f7f\u7528\u4e0b\u9762\u7684\u4ee3\u7801\uff0c\u67e5\u770b\u6211\u4eec\u5f97\u5230\u7684GPU\uff1a<\/p>\n

    \u00a0<\/p>\n

    \n
      \n
    • <\/li>\n<\/ul>\n
       !nvidia-smi<\/span><\/span><\/code><\/pre>\n<\/section>\n
      \"Image\"<\/p>\n
      \u975e\u5e38\u4e0d\u5e78\u6211\u4eec\u53ea\u5206\u914d\u5230\u4e86\u4e00\u4e2aT4\uff0c\u5982\u679c\u4f60\u80fd\u5206\u914d\u5230\u4e00\u5757P100\uff0c\u90a3\u4e48\u4f60\u7684\u63a8\u7406\u901f\u5ea6\u4f1a\u53d8\u5f97\u66f4\u5feb<\/p>\n
      \u4e0b\u9762\u6211\u4eec\u5b89\u88c5\u4e00\u4e9b\u9700\u8981\u7684\u5305\uff1adiffusers ,scipy, ftfy\u548ctransformer:<\/p>\n<\/section>\n
      \n
        \n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n<\/ul>\n
         !pip install<\/span> diffusers==0.4<\/span>.0<\/span><\/span><\/code> !pip install<\/span> transformers scipy ftfy<\/span><\/code> !pip install<\/span> \"ipywidgets>=7,<8\"<\/span><\/span><\/code><\/pre>\n<\/section>\n
        \u8fd9\u91cc\u9700\u8981\u7684\u989d\u5916\u64cd\u4f5c\u662f\u5fc5\u987b\u540c\u610f\u6a21\u578b\u534f\u8bae\uff0c\u8fd8\u8981\u901a\u8fc7\u52fe\u9009\u590d\u9009\u6846\u6765\u63a5\u53d7\u6a21\u578b\u8bb8\u53ef\u3002\u201cHugging Face\u201d\u4e0a\u6ce8\u518c\uff0c\u5e76\u83b7\u5f97\u8bbf\u95ee\u4ee4\u724c\u7b49\u7b49\u3002<\/p>\n
        \u53e6\u5916\u5bf9\u4e8e\u8c37\u6b4ccollab\uff0c\u5b83\u5df2\u7ecf\u7981\u7528\u4e86\u5916\u90e8\u5c0f\u90e8\u4ef6\uff0c\u6240\u4ee5\u9700\u8981\u542f\u7528\u5b83\u3002\u8fd0\u884c\u4ee5\u4e0b\u4ee3\u7801\u8fd9\u6837\u624d\u80fd\u591f\u4f7f\u7528\u201cnotebook_login\u201d<\/p>\n<\/section>\n
        \n
          \n
        • <\/li>\n
        • <\/li>\n<\/ul>\n
           from<\/span> google.colab import<\/span> output<\/span><\/code>\u00a0output.enable_custom_widget_manager()<\/span><\/code><\/pre>\n<\/section>\n
          \u73b0\u5728\u5c31\u53ef\u4ee5\u4ece\u7684\u8d26\u6237\u4e2d\u83b7\u5f97\u7684\u8bbf\u95ee\u4ee4\u724c\u767b\u5f55Hugging Face\u4e86:<\/p>\n<\/section>\n
          \n
            \n
          • <\/li>\n
          • <\/li>\n<\/ul>\n
             from<\/span> huggingface_hub import<\/span> notebook_login<\/span><\/code>\u00a0notebook_login()<\/span><\/code><\/pre>\n<\/section>\n
            \u4ecediffusers\u5e93\u52a0\u8f7dStableDiffusionPipeline\u3002<\/p>\n
            StableDiffusionPipeline\u662f\u4e00\u4e2a\u7aef\u5230\u7aef\u63a8\u7406\u7ba1\u9053\uff0c\u53ef\u7528\u4e8e\u4ece\u6587\u672c\u751f\u6210\u56fe\u50cf\u3002<\/p>\n
            \u6211\u4eec\u5c06\u52a0\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b\u6743\u91cd\u3002\u6a21\u578bid\u5c06\u662fCompVis\/ stable-diffusion-v1-4\uff0c\u6211\u4eec\u4e5f\u5c06\u4f7f\u7528\u4e00\u4e2a\u7279\u5b9a\u7c7b\u578b\u7684\u4fee\u8ba2\u7248torch_dtype\u51fd\u6570\u3002\u8bbe\u7f6erevision = \u201cfp16\u201d\u4ece\u534a\u7cbe\u5ea6\u5206\u652f\u52a0\u8f7d\u6743\u91cd\uff0c\u5e76\u8bbe\u7f6etorch_dtype = ” torch\u3002torch_dtype = \u201ctorch.float16\u201d\u544a\u8bc9\u6a21\u578b\u4f7f\u7528fp16\u7684\u6743\u91cd\u3002<\/p>\n
            \u50cf\u8fd9\u6837\u8bbe\u7f6e\u53ef\u4ee5\u51cf\u5c11\u5185\u5b58\uff0c\u5e76\u4e14\u8fd0\u884c\u7684\u66f4\u5feb\u3002<\/p>\n<\/section>\n
            \n
              \n
            • <\/li>\n
            • <\/li>\n
            • <\/li>\n
            • <\/li>\n
            • <\/li>\n<\/ul>\n
               import<\/span> torch<\/span><\/code> from<\/span> diffusers import<\/span> StableDiffusionPipeline<\/span><\/code> <\/code> # make sure you're logged in with `huggingface-cli login`<\/span><\/span><\/code> pipe = StableDiffusionPipeline.from_pretrained(\"CompVis\/stable-diffusion-v1-4\"<\/span>, revision=\"fp16\"<\/span>, torch_dtype=torch.float16)  <\/span><\/code><\/pre>\n<\/section>\n
              \u4e0b\u9762\u8bbe\u7f6eGPU\uff1a<\/p>\n<\/section>\n
              \n
                \n
              • <\/li>\n<\/ul>\n
                 pipe<\/span> = pipe.to(\"cuda\"<\/span>)<\/span><\/code><\/pre>\n<\/section>\n
                \n\u73b0\u5728\u5c31\u53ef\u4ee5\u751f\u6210\u56fe\u7247\u4e86\u3002\u6211\u4eec\u5c06\u7f16\u5199\u4e00\u4e2a\u63d0\u793a\u6587\u672c\u5e76\u5c06\u5176\u4ea4\u7ed9\u7ba1\u9053\u5e76\u6253\u5370\u8f93\u51fa\u3002\u8fd9\u91cc\u7684\u8f93\u5165\u63d0\u793a\u662f\u201can astronaut riding a horse\u201d\uff0c\u8ba9\u770b\u770b\u8f93\u51fa:<\/p>\n<\/section>\n
                \n
                  \n
                • <\/li>\n
                • <\/li>\n
                • <\/li>\n
                • <\/li>\n
                • <\/li>\n<\/ul>\n
                   prompt = \"a photograph of an astronaut riding a horse\"<\/span><\/code> image = pipe(prompt).images[0<\/span>]  # image here is in [PIL format<\/span>](https:\/\/pillow.readthedocs.io\/en\/stable\/<\/span>)<\/span><\/code> <\/code> # Now to display an image you can do either save it such as:<\/span><\/code> image.save(f\"astronaut_rides_<\/span>horse.png\")<\/span><\/code><\/pre>\n<\/section>\n
                  \"Image\"<\/p>\n
                  \u6bcf\u6b21\u8fd0\u884c\u4e0a\u9762\u7684\u4ee3\u7801\uff0c\u90fd\u4f1a\u5f97\u5230\u4e0d\u540c\u7684\u56fe\u50cf\u3002\u4e3a\u4e86\u6bcf\u6b21\u90fd\u5f97\u5230\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u4f60\u53ef\u4ee5\u5411\u4f20\u9012\u4e00\u4e2a\u968f\u673a\u79cd\u5b50\uff0c\u5982\u4e0b\u9762\u7684\u4ee3\u7801\u6240\u793a:<\/p>\n<\/section>\n
                  \n
                    \n
                  • <\/li>\n
                  • <\/li>\n
                  • <\/li>\n
                  • <\/li>\n<\/ul>\n
                     import<\/span> torch<\/span><\/span><\/code> generator<\/span> = torch.Generator(\"cuda\").manual_seed(1024)<\/span><\/span><\/code> image<\/span> = pipe(prompt, generator=generator).images[0]<\/span><\/span><\/code> image<\/span><\/span><\/code><\/pre>\n<\/section>\n
                    \u8fd8\u53ef\u4ee5\u4f7f\u7528num_inference_steps\u53c2\u6570\u66f4\u6539\u6b65\u9aa4\u7684\u6570\u91cf\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u63a8\u7406\u6b65\u9aa4\u8d8a\u591a\uff0c\u751f\u6210\u7684\u56fe\u50cf\u8d28\u91cf\u8d8a\u9ad8\uff0c\u4f46\u751f\u6210\u7ed3\u679c\u9700\u8981\u66f4\u591a\u7684\u65f6\u95f4\u3002\u5982\u679c\u4f60\u60f3\u8981\u66f4\u5feb\u7684\u7ed3\u679c\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u66f4\u5c11\u7684\u6b65\u9aa4\u3002<\/p>\n
                    \u4e0b\u9762\u7684\u5355\u5143\u683c\u4f7f\u7528\u4e0e\u524d\u9762\u76f8\u540c\u7684\u79cd\u5b50\uff0c\u4f46\u6b65\u9aa4\u66f4\u5c11\u3002\u6ce8\u610f\u4e00\u4e9b\u7ec6\u8282\uff0c\u5982\u9a6c\u5934\u6216\u5934\u76d4\uff0c\u6bd4\u524d\u4e00\u5f20\u56fe\u5b9a\u4e49\u5f97\u66f4\u6a21\u7cca:<\/p>\n<\/section>\n
                    \n
                      \n
                    • <\/li>\n
                    • <\/li>\n
                    • <\/li>\n
                    • <\/li>\n<\/ul>\n
                       import<\/span> torch<\/span><\/span><\/code> generator<\/span> = torch.Generator(\"cuda\").manual_seed(1024)<\/span><\/span><\/code> image<\/span> = pipe(prompt, num_inference_steps=15, generator=generator).images[0]<\/span><\/span><\/code> image<\/span><\/span><\/code><\/pre>\n<\/section>\n
                      \n\"Image\"<\/p>\n
                      \u53e6\u4e00\u4e2a\u53c2\u6570\u662fGuidance Scale\u3002\u8fd9\u662f\u4e00\u79cd\u63d0\u9ad8\u5bf9\u6761\u4ef6\u4fe1\u53f7\u7684\u4f9d\u4ece\u6027\u7684\u65b9\u6cd5\uff0c\u5728\u6269\u6563\u6a21\u578b\u7684\u60c5\u51b5\u4e0b\u5b83\u662f\u6587\u672c\u548c\u6574\u4f53\u6837\u672c\u8d28\u91cf\u3002<\/p>\n
                      \u7b80\u5355\u5730\u8bf4\uff0c\u65e0\u5206\u7c7b\u4fe1\u606f\u7684\u5f15\u5bfc\u8feb\u4f7f\u751f\u6210\u4e0e\u6587\u672c\u63d0\u793a\u66f4\u597d\u5730\u5339\u914d\u3002\u50cf7\u62168.5\u8fd9\u6837\u7684\u6570\u5b57\u53ef\u4ee5\u7ed9\u51fa\u5f88\u597d\u7684\u7ed3\u679c\u3002\u5982\u679c\u4f7f\u7528\u7684\u6570\u5b57\u975e\u5e38\u5927\u56fe\u50cf\u53ef\u80fd\u770b\u8d77\u6765\u5f88\u597d\uff0c\u4f46\u4f1a\u51cf\u5c11\u591a\u6837\u6027\u3002<\/p>\n
                      \u5982\u679c\u8981\u4e3a\u76f8\u540c\u7684\u6587\u672c\u63d0\u793a\u751f\u6210\u591a\u4e2a\u56fe\u50cf\uff0c\u53ea\u9700\u91cd\u590d\u591a\u6b21\u8f93\u5165\u76f8\u540c\u7684\u6587\u672c\u5373\u53ef\u3002\u6211\u4eec\u53ef\u4ee5\u628a\u6587\u672c\u7684\u5217\u8868\u53d1\u9001\u5230\u6a21\u578b\u4e2d\uff0c\u8ba9\u6211\u4eec\u7f16\u5199\u4e00\u4e2a\u52a9\u624b\u51fd\u6570\u6765\u663e\u793a\u591a\u4e2a\u56fe\u50cf\u3002<\/p>\n<\/section>\n
                      \n
                        \n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n
                      • <\/li>\n<\/ul>\n
                         from<\/span> PIL import<\/span> Image<\/span><\/code> <\/code> def<\/span> image_grid<\/span>(imgs, rows, cols)<\/span>:<\/span><\/span><\/code>     assert<\/span> len(imgs) == rows*cols<\/span><\/code> <\/code>     w, h = imgs[0<\/span>].size<\/span><\/code>     grid = Image.new('RGB'<\/span>, size=(cols*w, rows*h))<\/span><\/code>     grid_w, grid_h = grid.size<\/span><\/code>     <\/code>     for<\/span> i, img in<\/span> enumerate(imgs):<\/span><\/code>         grid.paste(img, box=(i%cols*w, i\/\/cols*h))<\/span><\/code>     return<\/span> grid<\/span><\/code><\/pre>\n<\/section>\n
                        \u73b0\u5728\uff0c\u6211\u4eec\u53ef\u4ee5\u751f\u6210\u591a\u4e2a\u56fe\u50cf\u5e76\u4e00\u8d77\u5c55\u793a\u4e86\u3002<\/p>\n<\/section>\n
                        \n
                          \n
                        • <\/li>\n
                        • <\/li>\n
                        • <\/li>\n
                        • <\/li>\n
                        • <\/li>\n
                        • <\/li>\n
                        • <\/li>\n<\/ul>\n
                           num_images<\/span> = 3<\/span><\/span><\/code> prompt<\/span> = [\"a photograph of an astronaut riding a horse\"] * num_images<\/span><\/span><\/code> <\/code> images<\/span> = pipe(prompt).images<\/span><\/span><\/code> <\/code> grid<\/span> = image_grid(images, rows=1, cols=3)<\/span><\/span><\/code> grid<\/span><\/span><\/code><\/pre>\n<\/section>\n
                          \"Image\"<\/p>\n
                          \u8fd8\u53ef\u4ee5\u751f\u6210n*m\u5f20\u56fe\u50cf:<\/p>\n<\/section>\n
                          \n
                            \n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n
                          • <\/li>\n<\/ul>\n
                             num_cols<\/span> = 3<\/span><\/span><\/code> num_rows<\/span> = 4<\/span><\/span><\/code> <\/code> prompt<\/span> = [\"a photograph of an astronaut riding a horse\"] * num_cols<\/span><\/span><\/code> <\/code> all_images<\/span> = []<\/span><\/span><\/code> for<\/span> i in range(num_rows):<\/span><\/span><\/code>   images<\/span> = pipe(prompt).images<\/span><\/span><\/code>   all_images.extend(images)<\/span><\/span><\/code> <\/code> grid<\/span> = image_grid(all_images, rows=num_rows, cols=num_cols)<\/span><\/span><\/code> grid<\/span><\/span><\/code><\/pre>\n<\/section>\n
                            \"Image\"<\/p>\n
                            \u751f\u6210\u7684\u56fe\u50cf\u9ed8\u8ba4\u5927\u5c0f\u4e3a512*512\u50cf\u7d20\u3002\u53ef\u4ee5\u4f7f\u7528height\u548cwidth\u53c2\u6570\u6765\u66f4\u6539\u751f\u6210\u56fe\u50cf\u7684\u9ad8\u5ea6\u548c\u5bbd\u5ea6\u3002\u8fd9\u91cc\u6709\u4e00\u4e9b\u9009\u62e9\u597d\u7684\u56fe\u7247\u5927\u5c0f\u7684\u6280\u5de7:<\/p>\n
                            \u5c06height\u548cwidth\u53c2\u6570\u90fd\u9009\u62e9\u4e3a8\u7684\u500d\u6570\u3002\u9ad8\u5ea6\u548c\u5bbd\u5ea6\u8bbe\u7f6e\u4e3a\u5c0f\u4e8e512,\u53ef\u80fd\u4f1a\u5bfc\u81f4\u8d28\u91cf\u6bd4\u8f83\u5dee\u5982\u679c\u4e24\u4e2a\u90fd\u8bbe\u7f6e\u4e3a512\u4ee5\u4e0a\u53ef\u80fd\u4f1a\u51fa\u73b0\u5168\u5c40\u8fde\u8d2f\u6027(Global Coherence)\uff0c\u6240\u4ee5\u5982\u679c\u9700\u8981\u5927\u56fe\u50cf\u53ef\u4ee5\u8bd5\u8bd5\u9009\u4e00\u4e2a\u503c\u56fa\u5b9a\u7684512\uff0c\u800c\u53e6\u4e00\u4e2a\u5927\u4e8e512\u3002\u4f8b\u5982\u4e0b\u9762\u7684\u5927\u5c0f\uff1a<\/p>\n<\/section>\n
                            \n
                              \n
                            • <\/li>\n
                            • <\/li>\n
                            • <\/li>\n
                            • <\/li>\n<\/ul>\n
                               prompt<\/span> = \"a photograph of an astronaut riding a horse\"<\/span><\/span><\/code> <\/code> image = pipe(prompt, height=512<\/span>, width=768<\/span>).images[0<\/span>]<\/span><\/code> image<\/span><\/code><\/pre>\n<\/section>\n
                              \n\"Image\"<\/p>\n
                              <\/h2>\n
                              \u5efa\u7acb\u4f60\u81ea\u5df1\u7684\u5904\u7406\u7ba1\u9053<\/strong><\/h2>\n
                              \u6211\u4eec\u4e5f\u53ef\u4ee5\u901a\u8fc7Diffusers\u81ea\u5b9a\u4e49\u6269\u6563\u7ba1\u9053\u4e0e\u6269\u6563\u5668\u3002\u8fd9\u91cc\u5c06\u6f14\u793a\u5982\u4f55\u4f7f\u7528\u4e0d\u540c\u7684scheduler\uff0c\u5373Katherine Crowson\u7684K-LMS\u8c03\u5ea6\u5668\u3002<\/p>\n
                              \u6211\u4eec\u5148\u770b\u4e00\u4e0bStableDiffusionPipeline:<\/p>\n
                              \u00a0<\/p>\n<\/section>\n
                              \n
                                \n
                              • <\/li>\n
                              • <\/li>\n<\/ul>\n
                                \u00a0import<\/span>\u00a0torch<\/span><\/code> torch_device = \"cuda\"<\/span> if<\/span> torch.cuda.is_available() else<\/span> \"cpu\"<\/span><\/span><\/code><\/pre>\n<\/section>\n
                                \u9884\u8bad\u7ec3\u7684\u6a21\u578b\u5305\u62ec\u5efa\u7acb\u4e00\u4e2a\u5b8c\u6574\u7684\u7ba1\u9053\u6240\u9700\u7684\u6240\u6709\u7ec4\u4ef6\u3002\u5b83\u4eec\u5b58\u653e\u5728\u4ee5\u4e0b\u6587\u4ef6\u5939\u4e2d:<\/p>\n
                                text_encoder:Stable Diffusion\u4f7f\u7528CLIP\uff0c\u4f46\u5176\u4ed6\u6269\u6563\u6a21\u578b\u53ef\u80fd\u4f7f\u7528\u5176\u4ed6\u7f16\u7801\u5668\uff0c\u5982BERT\u3002<\/p>\n
                                tokenizer:\u5b83\u5fc5\u987b\u4e0etext_encoder\u6a21\u578b\u4f7f\u7528\u7684\u6807\u8bb0\u5668\u5339\u914d\u3002<\/p>\n
                                scheduler:\u7528\u4e8e\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9010\u6b65\u5411\u56fe\u50cf\u6dfb\u52a0\u566a\u58f0\u7684scheduler\u7b97\u6cd5\u3002<\/p>\n
                                U-Net:\u7528\u4e8e\u751f\u6210\u8f93\u5165\u7684\u6f5c\u5728\u8868\u793a\u7684\u6a21\u578b\u3002<\/p>\n
                                VAE\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u5b83\u5c06\u6f5c\u5728\u7684\u8868\u793a\u89e3\u7801\u4e3a\u771f\u5b9e\u7684\u56fe\u50cf\u3002<\/p>\n
                                \u53ef\u4ee5\u901a\u8fc7\u5f15\u7528\u7ec4\u4ef6\u88ab\u4fdd\u5b58\u7684\u6587\u4ef6\u5939\uff0c\u4f7f\u7528from_pretraining\u7684\u5b50\u6587\u4ef6\u5939\u53c2\u6570\u6765\u52a0\u8f7d\u7ec4\u4ef6\u3002<\/p>\n<\/section>\n
                                \n
                                  \n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n
                                • <\/li>\n<\/ul>\n
                                   from<\/span> transformers import<\/span> CLIPTextModel, CLIPTokenizer<\/span><\/code> from<\/span> diffusers import<\/span> AutoencoderKL, UNet2DConditionModel, PNDMScheduler<\/span><\/code> <\/code> # 1. Load the autoencoder model which will be used to decode the latents into image space. <\/span><\/span><\/code> vae = AutoencoderKL.from_pretrained(\"CompVis\/stable-diffusion-v1-4\"<\/span>, subfolder=\"vae\"<\/span>)<\/span><\/code> <\/code> # 2. Load the tokenizer and text encoder to tokenize and encode the text. <\/span><\/span><\/code> tokenizer = CLIPTokenizer.from_pretrained(\"openai\/clip-vit-large-patch14\"<\/span>)<\/span><\/code> text_encoder = CLIPTextModel.from_pretrained(\"openai\/clip-vit-large-patch14\"<\/span>)<\/span><\/code> <\/code> # 3. The UNet model for generating the latents.<\/span><\/span><\/code> unet = UNet2DConditionModel.from_pretrained(\"CompVis\/stable-diffusion-v1-4\"<\/span>, subfolder=\"unet\"<\/span>)<\/span><\/code><\/pre>\n<\/section>\n
                                  \n\u73b0\u5728\uff0c\u6211\u4eec\u4e0d\u52a0\u8f7d\u9884\u5b9a\u4e49\u7684scheduler\uff0c\u800c\u662f\u52a0\u8f7dK-LMS\uff1a<\/p>\n<\/section>\n
                                  \n
                                    \n
                                  • <\/li>\n
                                  • <\/li>\n
                                  • <\/li>\n<\/ul>\n
                                     from<\/span> diffusers import<\/span> LMSDiscreteScheduler<\/span><\/code> <\/code> scheduler = LMSDiscreteScheduler(beta_start=0.00085<\/span>, beta_end=0.012<\/span>, beta_schedule=\"scaled_linear\"<\/span>, num_train_timesteps=1000<\/span>)<\/span><\/code><\/pre>\n<\/section>\n
                                    \u5c06\u6a21\u578b\u79fb\u52a8\u5230GPU\u4e0a\u3002<\/p>\n<\/section>\n
                                    \n
                                      \n
                                    • <\/li>\n
                                    • <\/li>\n
                                    • <\/li>\n<\/ul>\n
                                       vae<\/span> = vae.to(torch_device)<\/span><\/span><\/code> text_encoder<\/span> = text_encoder.to(torch_device)<\/span><\/span><\/code> unet<\/span> = unet.to(torch_device)<\/span><\/span><\/code><\/pre>\n<\/section>\n
                                      \n\u5b9a\u4e49\u7528\u4e8e\u751f\u6210\u56fe\u50cf\u7684\u53c2\u6570\u3002\u4e0e\u524d\u9762\u7684\u793a\u4f8b\u76f8\u6bd4\uff0c\u8bbe\u7f6enum_inference_steps = 100\u6765\u83b7\u5f97\u66f4\u660e\u786e\u7684\u56fe\u50cf\u3002<\/p>\n<\/section>\n
                                      \n
                                        \n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n
                                      • <\/li>\n<\/ul>\n
                                         prompt<\/span> = [\"a photograph of an astronaut riding a horse\"<\/span>]<\/span><\/code> <\/code> height = 512<\/span>                        # default height of Stable Diffusion<\/span><\/span><\/code> width = 512<\/span>                         # default width of Stable Diffusion<\/span><\/span><\/code> <\/code> num_inference_steps = 100<\/span>            # Number of denoising steps<\/span><\/span><\/code> <\/code> guidance_scale = 7<\/span>.5<\/span>                # Scale for classifier-free guidance<\/span><\/span><\/code> <\/code> generator = torch.manual_seed(32<\/span>)   # Seed generator to create the inital latent noise<\/span><\/span><\/code> <\/code> batch_size = 1<\/span><\/span><\/code><\/pre>\n<\/section>\n
                                        \u83b7\u53d6\u6587\u672c\u63d0\u793a\u7684text_embeddings\u3002\u7136\u540e\u5c06\u5d4c\u5165\u7528\u4e8e\u8c03\u6574U-Net\u6a21\u578b\u3002<\/p>\n<\/section>\n
                                        \n
                                          \n
                                        • <\/li>\n
                                        • <\/li>\n
                                        • <\/li>\n
                                        • <\/li>\n<\/ul>\n
                                           text_input = tokenizer(prompt, padding=\"max_length\"<\/span>, max_length=tokenizer.model_max_length, truncation=True<\/span>, return_tensors=\"pt\"<\/span>)<\/span><\/code> <\/code> with<\/span> torch.no_grad():<\/span><\/code>   text_embeddings = text_encoder(text_input.input_ids.to(torch_device))[0<\/span>]<\/span><\/code><\/pre>\n<\/section>\n
                                          \n\u83b7\u5f97\u7528\u4e8e\u65e0\u5206\u7c7b\u5668\u5f15\u5bfc\u7684\u65e0\u6761\u4ef6\u6587\u672c\u5d4c\u5165\uff0c\u8fd9\u53ea\u662f\u586b\u5145\u4ee4\u724c(\u7a7a\u6587\u672c)\u7684\u5d4c\u5165\u3002\u5b83\u4eec\u9700\u8981\u5177\u6709\u4e0etext_embeddings (batch_size\u548cseq_length)\u76f8\u540c\u7684\u5f62\u72b6\u3002<\/p>\n
                                          \u00a0<\/p>\n<\/section>\n
                                          \n