{"id":6239,"date":"2025-10-07T15:44:10","date_gmt":"2025-10-07T07:44:10","guid":{"rendered":"https:\/\/nullthought.net\/?p=6239"},"modified":"2025-10-07T15:44:11","modified_gmt":"2025-10-07T07:44:11","slug":"diffusion-transformer-dit","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6239","title":{"rendered":"Diffusion Transformer (DiT)"},"content":{"rendered":"\n<p>\u6269\u6563\u6a21\u578b\u5728\u56fe\u50cf\u751f\u6210\u4e0a\u957f\u671f\u4ee5\u5377\u79ef\u5f0f U-Net \u4e3a\u4e3b\u5e72\uff0c\u4f46\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2212.09748\" target=\"_blank\" rel=\"noreferrer noopener\">Scalable Diffusion Models with Transformers<\/a><\/strong>\u63d0\u51fa\u4ee5 Transformer \u4f5c\u4e3a\u6269\u6563\u6a21\u578b\u7684\u4e3b\u5e72\uff08Diffusion Transformer\uff0c\u7b80\u79f0 DiT\uff09\uff0c\u5e76\u7cfb\u7edf\u7814\u7a76\u5176\u201c\u53ef\u6269\u5c55\u6027\u201d\uff1a\u5f53\u901a\u8fc7\u589e\u5927\u7f51\u7edc\u6df1\/\u5bbd\u6216\u589e\u52a0\u8f93\u5165 token \u6570\u4f7f\u524d\u5411\u8ba1\u7b97\u91cf\uff08\u4ee5 Gflops \u8ba1\uff09\u589e\u52a0\u65f6\uff0c\u751f\u6210\u8d28\u91cf\uff08\u4ee5 FID \u8ba1\uff09\u662f\u5426\u7a33\u5b9a\u63d0\u5347\u3002\u4f5c\u8005\u7ed9\u51fa\u660e\u786e\u8bc1\u636e\uff1a\u66f4\u9ad8\u7684\u524d\u5411 Gflops \u4e0e\u66f4\u4f4e\u7684 FID \u5f3a\u8d1f\u76f8\u5173\uff08\u76f8\u5173\u7cfb\u6570\u7ea6 \u22120.93\uff09\uff0c\u540c\u65f6\u5728 ImageNet 256\u00d7256 \u4e0e 512\u00d7512 \u4e0a\u53d6\u5f97\u5f53\u65f6\u7684 SOTA\uff0c\u5176\u4e2d 256\u00d7256 \u7684 FID \u964d\u81f3 2.27\u3002\u8bba\u6587\u8fd8\u5c55\u793a\u4e86\u5728\u6f5c\u7a7a\u95f4\uff08LDM \u6846\u67b6\uff09\u4e2d\u4ee5 ViT \u98ce\u683c patch \u5316\u7684\u65b9\u5f0f\u66ff\u6362 U-Net \u7684\u53ef\u884c\u6027\u4e0e\u6548\u7387\u4f18\u52bf\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aWilliam Peebles, Saining Xie\u3002<\/p>\n\n\n\n<p>\u4e00\u3001\u65b9\u6cd5\u603b\u89c8\uff1aLatent-space DiT<br>\u4f5c\u8005\u91c7\u7528\u201c\u6f5c\u7a7a\u95f4\u6269\u6563\u201d\u8303\u5f0f\uff1a\u5148\u7528\u9884\u8bad\u7ec3 VAE \u5c06\u56fe\u50cf\u538b\u7f29\u4e3a\u8f83\u5c0f\u7684\u7a7a\u95f4\u8868\u793a z\uff08\u5982 256\u00d7256\u00d73 \u56fe\u50cf\u88ab\u7f16\u7801\u4e3a 32\u00d732\u00d74 \u7684 latent\uff09\uff0c\u6269\u6563\u6a21\u578b\u5728 z \u7a7a\u95f4\u8bad\u7ec3\u4e0e\u91c7\u6837\uff0c\u6700\u540e\u7528 VAE \u89e3\u7801\u56de\u50cf\u7d20\u3002DiT \u4ee5 ViT \u4e3a\u84dd\u672c\uff1a\u5c06 z \u6309 p\u00d7p patch \u7ebf\u6027\u5d4c\u5165\u4e3a token \u5e8f\u5217\uff08\u957f\u5ea6 T=(I\/p)\u00b2\uff09\uff0c\u53e0\u52a0\u6b63\u5f26\/\u4f59\u5f26\u4f4d\u7f6e\u7f16\u7801\u540e\uff0c\u9001\u5165\u4e00\u4e32 Transformer blocks \u8fdb\u884c\u566a\u58f0\u4e0e\u5bf9\u89d2\u534f\u65b9\u5dee\u9884\u6d4b\uff0c\u6700\u7ec8\u7ebf\u6027\u6295\u5f71\u5e76\u6309\u539f\u7a7a\u95f4\u6392\u5e03\u8fd8\u539f\u81f3 p\u00d7p\u00d72C\uff08\u5bf9\u5e94\u566a\u58f0\u4e0e\u534f\u65b9\u5dee\uff09\u3002p \u8d8a\u5c0f\uff0ctoken \u8d8a\u591a\uff0cGflops \u589e\u957f\u66f4\u5feb\uff0c\u4f46\u53c2\u6570\u91cf\u57fa\u672c\u4e0d\u53d8\uff0c\u4ece\u800c\u53ef\u5c06\u201c\u7b97\u529b\u201d\u4e0e\u201c\u53c2\u6570\u91cf\u201d\u89e3\u8026\u5230\u4e00\u5b9a\u7a0b\u5ea6\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u6761\u4ef6\u6ce8\u5165\u4e0e DiT Block \u8bbe\u8ba1<br>\u8bba\u6587\u7cfb\u7edf\u6bd4\u8f83\u4e86\u56db\u79cd\u5c06\u6761\u4ef6\uff08\u6269\u6563\u65f6\u95f4\u6b65 t\u3001\u7c7b\u522b\u6807\u7b7e c \u7b49\uff09\u6ce8\u5165 Transformer \u7684\u65b9\u5f0f\uff1a<br>1\uff09In-context conditioning\uff1a\u5c06 t\u3001c \u7684\u5411\u91cf\u76f4\u63a5\u4f5c\u4e3a\u989d\u5916 tokens \u62fc\u5230\u5e8f\u5217\u4e2d\uff0c\u51e0\u4e4e\u4e0d\u589e\u52a0 Gflops\u3002<br>2\uff09Cross-Attention\uff1a\u5728\u81ea\u6ce8\u610f\u529b\u540e\u52a0\u5165\u591a\u5934\u4ea4\u53c9\u6ce8\u610f\u529b\uff0c\u7528\u6761\u4ef6\u5e8f\u5217\u67e5\u8be2\u56fe\u50cf tokens\uff0c\u7ea6\u5e26\u6765 15% \u7684\u989d\u5916 Gflops\u3002<br>3\uff09Adaptive LayerNorm\uff08adaLN\uff09\uff1a\u7528(t+c) \u7684\u5d4c\u5165\u56de\u5f52\u5c42\u5f52\u4e00\u5316\u4e2d\u7684 \u03b3\u3001\u03b2\uff0c\u5bf9\u6240\u6709 tokens \u65bd\u52a0\u540c\u4e00\u6761\u4ef6\u4eff\u5c04\u53d8\u6362\uff0c\u989d\u5916 Gflops \u6781\u5c0f\u3002<br>4\uff09adaLN-Zero\uff1a\u5728 adaLN \u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u6b8b\u5dee\u524d\u7684\u9010\u7ef4\u7f29\u653e \u03b1\uff0c\u5e76\u5c06\u5176 MLP \u521d\u59cb\u5316\u4e3a\u5168\u96f6\uff0c\u4f7f\u6bcf\u4e2a Transformer block \u521d\u59cb\u4e3a\u201c\u6052\u7b49\u201d\u6620\u5c04\uff08identity-like initialization\uff09\uff0c\u8bad\u7ec3\u66f4\u7a33\u66f4\u5feb\u3002\u5b9e\u9a8c\u663e\u793a\uff0cadaLN-Zero \u5728\u5168\u7a0b\u8bad\u7ec3\u4e2d FID \u6700\u4f18\u4e14\u8ba1\u7b97\u6700\u7701\u3002<\/p>\n\n\n\n<p>\u4e09\u3001\u6a21\u578b\u89c4\u683c\u4e0e\u7f29\u653e\u7ef4\u5ea6<br>\u4f5c\u8005\u6cbf\u7528 ViT \u7684 S\u3001B\u3001L\u3001XL \u56db\u6863\u914d\u7f6e\uff08\u5c42\u6570 N\u3001\u9690\u5c42\u5bbd d\u3001\u6ce8\u610f\u529b\u5934\u6570\u968f\u4e4b\u589e\u957f\uff09\uff0c\u5e76\u4e0e patch \u5927\u5c0f p\u2208{2,4,8} \u7ec4\u5408\uff0c\u5f62\u6210 12 \u4e2a\u6a21\u578b\uff08\u5982\u201cDiT-XL\/2\u201d\u8868\u793a XL \u914d\u7f6e\u4e14 p=2\uff09\u3002\u4e24\u6761\u4e3b\u8981\u7f29\u653e\u8f74\uff1a<br>\uff08A\uff09\u589e\u5927\u6a21\u578b\u6df1\u5bbd\uff08\u63d0\u9ad8\u6bcf\u5c42\u7ef4\u5ea6\u4e0e\u5c42\u6570\uff09\uff1b<br>\uff08B\uff09\u51cf\u5c0f patch \u5c3a\u5bf8\u3001\u589e\u52a0 token \u6570\uff08\u4e3b\u8981\u589e\u52a0 Gflops \u800c\u51e0\u4e4e\u4e0d\u53d8\u53c2\u6570\u91cf\uff09\u3002\u4e24\u8f74\u5747\u663e\u8457\u964d\u4f4e FID\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u53ef\u6269\u5c55\u6027\u5173\u952e\u53d1\u73b0<br>1\uff09\u201cGflops \u800c\u975e\u53c2\u6570\u91cf\u201d\u9a71\u52a8\u8d28\u91cf\u63d0\u5347\uff1a\u5728\u56fa\u5b9a\u53c2\u6570\u91cf\uff08\u56fa\u5b9a\u914d\u7f6e\uff09\u4e0b\uff0c\u4ec5\u901a\u8fc7\u51cf\u5c0f p \u63d0\u9ad8 token \u6570\u3001\u589e\u52a0\u524d\u5411 Gflops\uff0c\u4e5f\u80fd\u5927\u5e45\u964d FID\uff1b\u4e0d\u540c\u914d\u7f6e\u53ea\u8981 Gflops \u76f8\u8fd1\uff0c\u5176 FID \u5f80\u5f80\u63a5\u8fd1\uff08\u5982 S\/2 \u4e0e B\/4\uff09\u3002<br>2\uff09\u5f3a\u76f8\u5173\u6027\uff1a\u5728 400k \u6b65\u8bad\u7ec3\u540e\uff0c\u5404\u6a21\u578b\u7684 FID-50K \u4e0e Transformer \u524d\u5411 Gflops \u5448\u663e\u8457\u8d1f\u76f8\u5173\uff08\u2248\u22120.93\uff09\u3002<br>3\uff09\u66f4\u5927\u6a21\u578b\u7684\u201c\u8bad\u7ec3\u7b97\u529b\u5229\u7528\u7387\u201d\u66f4\u9ad8\uff1a\u4ee5\u201c\u603b\u8bad\u7ec3\u7b97\u529b = \u6a21\u578b Gflops \u00d7 batch \u00d7 steps \u00d73\u201d\u8fd1\u4f3c\u8861\u91cf\u65f6\uff0c\u5c0f\u6a21\u578b\u5373\u4fbf\u5ef6\u957f\u8bad\u7ec3\u4e5f\u4f1a\u5728\u5355\u4f4d\u7b97\u529b\u6548\u7387\u4e0a\u843d\u540e\u4e8e\u5927\u6a21\u578b\uff1b\u540c\u4e3a XL \u914d\u7f6e\u65f6\uff0cp=2 \u7684 XL\/2 \u5728\u7ea6 10\u00b9\u2070 Gflops \u540e\u5c31\u8d85\u8fc7 p=4 \u7684 XL\/4\u3002<\/p>\n\n\n\n<p>\u4e94\u3001\u4e0e U-Net \u67b6\u6784\u7684\u5bf9\u6bd4\u4e0e SOTA \u7ed3\u679c<br>\u5728 ImageNet 256\u00d7256\uff08\u7c7b\u522b\u6761\u4ef6\uff09\u4e0a\uff0cDiT-XL\/2\uff08\u5728\u91c7\u6837\u65f6\u4f7f\u7528 classifier-free guidance\uff09\u5c06 FID-50K \u63a8\u8fdb\u5230 2.27\uff0c\u4f18\u4e8e\u540c\u4e3a\u6f5c\u7a7a\u95f4\u7684 LDM \u4e0e\u50cf\u7d20\u7a7a\u95f4\u7684 ADM \u7cfb\u5217\uff1b\u5728 512\u00d7512 \u4e0a\u4e5f\u5237\u65b0\u6269\u6563\u6a21\u578b\u6700\u4f73\uff08\u6587\u4e2d\u62a5\u544a 3.04\uff09\uff0c\u540c\u65f6\u5176\u524d\u5411 Gflops \u76f8\u6bd4 ADM\/ADM-U \u66f4\u4f4e\uff0c\u663e\u793a\u51fa\u201c\u9ad8\u8d28\u91cf + \u8ba1\u7b97\u6548\u7387\u201d\u7684\u53cc\u91cd\u4f18\u52bf\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4f5c\u8005\u5f3a\u8c03 DiT \u7684 compute-efficiency\uff1a\u540c\u7b49\u6216\u66f4\u4f4e Gflops \u4e0b\u4f18\u4e8e U-Net \u7cfb\u5217\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u8bad\u7ec3\u4e0e\u8bc4\u6d4b\u8bbe\u7f6e<br>1\uff09\u8bad\u7ec3\u8bbe\u7f6e\uff1a\u5728 JAX + TPU v3-Pods \u4e0a\u8bad\u7ec3\uff0c\u4f18\u5316\u5668 AdamW\uff0c\u5b66\u4e60\u7387 1e-4\uff0c\u6279\u91cf 256\uff0c\u65e0\u6743\u91cd\u8870\u51cf\u4e0e\u9884\u70ed\uff0c\u4ec5\u6c34\u5e73\u7ffb\u8f6c\u589e\u5e7f\uff1b\u5bf9\u6240\u6709\u5c3a\u5bf8\u4e0e p \u53d6\u76f8\u540c\u8d85\u53c2\uff1b\u7ef4\u62a4 0.9999 \u7684 EMA\u3002<br>2\uff09\u6269\u6563\u8d85\u53c2\uff1at_max=1000 \u7684\u7ebf\u6027\u566a\u58f0\u65e5\u7a0b\uff0c\u9884\u6d4b\u566a\u58f0\uff08\u03f5-parametrization\uff09\uff0c\u5e76\u5b66\u4e60\u5bf9\u89d2\u534f\u65b9\u5dee \u03a3\u03b8\u3002<br>3\uff09\u6f5c\u7a7a\u95f4\uff1a\u4f7f\u7528 Stable Diffusion \u7684\u9884\u8bad\u7ec3 VAE\uff08\u7f16\u7801\u4e0b\u91c7\u6837 8 \u500d\uff09\uff0c\u8bba\u6587\u8fd8\u5bf9\u4e0d\u540c VAE \u89e3\u7801\u5668\u505a\u4e86\u5bf9\u7167\uff0c\u6307\u6807\u76f8\u8fd1\u3002<br>4\uff09\u8bc4\u6d4b\uff1a\u4e3b\u6307\u6807\u4e3a FID-50K\uff08\u9ed8\u8ba4 250 DDPM \u91c7\u6837\u6b65\uff09\uff0c\u53e6\u62a5\u544a IS\u3001sFID\u3001Precision\/Recall \u5e76\u5448\u73b0\u4e0e Gflops \u7684\u5173\u8054\u8d8b\u52bf\u66f2\u7ebf\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u91c7\u6837\u8ba1\u7b97 vs \u6a21\u578b\u8ba1\u7b97<br>\u6269\u6563\u6a21\u578b\u53ef\u901a\u8fc7\u589e\u52a0\u91c7\u6837\u6b65\u6570\u5728\u63a8\u7406\u65f6\u201c\u52a0\u7b97\u529b\u201d\u3002\u4f5c\u8005\u5bf9\u6bd4\u4e86\u56fa\u5b9a\u8bad\u7ec3\u6b65\uff08400k\uff09\u4e0b\uff0c\u4e0d\u540c\u6a21\u578b\u5728\u4e0d\u540c\u91c7\u6837\u6b65\uff0816\u21921000\uff09\u65f6\u7684 FID-10K\uff1a\u7ed3\u8bba\u662f\u201c\u5c0f\u6a21\u578b\u5373\u4f7f\u628a\u91c7\u6837\u7b97\u529b\u5806\u5230\u6bd4\u5927\u6a21\u578b\u66f4\u9ad8\uff0c\u4ecd\u96be\u4ee5\u5f25\u5408\u5dee\u8ddd\u201d\u3002\u4f8b\u5982 L\/2 \u7528 1000 \u6b65\u91c7\u6837\u5176\u5355\u56fe\u91c7\u6837\u7b97\u529b\u662f XL\/2\uff08128 \u6b65\uff09\u7684 ~5 \u500d\uff0c\u4f46 XL\/2 \u7684 FID \u4ecd\u66f4\u597d\u3002\u6a21\u578b\u89c4\u6a21\u5e26\u6765\u7684\u201c\u80fd\u529b\u201d\u4e0d\u80fd\u5355\u9760\u589e\u52a0\u91c7\u6837\u6b65\u5728\u6d4b\u8bd5\u65f6\u8865\u9f50\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u4e3a\u4f55 adaLN-Zero \u6709\u6548<br>\u5728 Transformer \u4e2d\u4ee5\u201c\u6761\u4ef6\u9a71\u52a8\u7684\u81ea\u9002\u5e94\u5c42\u5f52\u4e00\u5316\u201d\u66ff\u6362\u6807\u51c6 LN\uff0c\u53ef\u5728\u4e0d\u5f15\u5165\u6602\u8d35 cross-attn \u7684\u524d\u63d0\u4e0b\u5c06\u6761\u4ef6\u5f3a\u8026\u5408\u5230\u6bcf\u4e00\u5c42\u3001\u6bcf\u4e2a\u901a\u9053\u7684\u5c3a\u5ea6\u4e0e\u504f\u7f6e\u4e0a\uff1b\u518d\u52a0\u4e0a\u201c\u96f6\u521d\u59cb\u5316\u7684\u6b8b\u5dee\u7f29\u653e \u03b1\u201d\u8ba9\u6bcf\u4e2a block \u521a\u5f00\u59cb\u66f4\u50cf\u201c\u6052\u7b49\u6620\u5c04\u201d\uff0c\u7f13\u89e3\u6df1\u5c42\u7f51\u7edc\u65e9\u671f\u4f18\u5316\u4e0d\u7a33\u5b9a\uff0c\u5b9e\u8df5\u4e2d\u5728\u7b49\u7b97\u529b\u4e0b\u663e\u8457\u4f18\u4e8e in-context \u4e0e cross-attn\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u65e2\u7a33\u5b9a\u8bad\u7ec3\u3001\u53c8\u7701 Gflops\uff0c\u4f53\u73b0\u4e86\u9488\u5bf9\u6269\u6563\u53bb\u566a\u4efb\u52a1\u7684\u201c\u8584\u800c\u7a33\u201d\u7684\u6761\u4ef6\u6ce8\u5165\u4f18\u52bf\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u5de5\u7a0b\u4e0e\u5b9e\u8df5\u8981\u70b9<br>1\uff09\u4e24\u6761\u6269\u5c55\u8f74\uff1a\u4f18\u5148\u589e\u52a0\u201c\u6709\u6548\u7b97\u529b\u201d\u800c\u975e\u76f2\u76ee\u6269\u53c2\uff1b\u5728\u9884\u7b97\u6709\u9650\u65f6\uff0c\u51cf\u5c0f patch \u589e token \u5f80\u5f80\u66f4\u9ad8\u6548\u3002<br>2\uff09\u6f5c\u7a7a\u95f4\u4f18\u5148\uff1a\u5728\u76f8\u4f3c\u611f\u77e5\u8d28\u91cf\u4e0b\uff0c\u6f5c\u7a7a\u95f4\u6269\u6563\u76f8\u5bf9\u4e8e\u50cf\u7d20\u7a7a\u95f4\u663e\u8457\u7701\u7b97\u3002<br>3\uff09\u6307\u5bfc\u7b56\u7565\uff1aclassifier-free guidance \u4f9d\u7136\u662f\u8d28\u91cf\/\u591a\u6837\u6027\u7684\u5173\u952e\u65cb\u94ae\uff1b\u6587\u4e2d\u8fd8\u89c2\u5bdf\u5230\u5bf9\u6f5c\u901a\u9053\u5b50\u96c6\u65bd\u52a0 guidance \u4e5f\u80fd\u53d6\u5f97\u63a5\u8fd1\u6548\u679c\u3002<br>4\uff09\u7edf\u4e00\u67b6\u6784\u7ea2\u5229\uff1a\u7528\u6807\u51c6\u5316\u7684 ViT \u8bbe\u8ba1\u4e0e\u8bad\u7ec3 recipe\uff08\u65e0\u989d\u5916\u6b63\u5219\/\u9884\u70ed\u4ea6\u7a33\uff09\uff0c\u4fbf\u4e8e\u8de8\u4efb\u52a1\u8fc1\u79fb\u4e0e\u6301\u7eed\u6269\u5bb9\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u5c40\u9650\u4e0e\u5f00\u653e\u95ee\u9898<br>1\uff09\u7b97\u529b\u654f\u611f\uff1a\u5c3d\u7ba1 DiT \u76f8\u5bf9\u9ad8\u6548\uff0c\u4f46\u6700\u4f18\u7ed3\u679c\u4ecd\u4f9d\u8d56\u5927\u91cf\u8bad\u7ec3\u7b97\u529b\u4e0e\u957f\u65f6\u8bad\u7ec3\uff1b\u5982\u4f55\u8fdb\u4e00\u6b65\u63d0\u5347\u6570\u636e\u4e0e\u7b97\u529b\u5229\u7528\u7387\u4ecd\u662f\u95ee\u9898\u3002<br>2\uff09\u6587\u672c\u6761\u4ef6\/\u591a\u6a21\u6001\u6cdb\u5316\uff1a\u672c\u6587\u4ee5\u7c7b\u522b\u6761\u4ef6\u4e0e\u6f5c\u7a7a\u95f4\u4e3a\u4e3b\uff0c\u5982\u4f55\u5728\u6587\u672c\u5230\u56fe\u50cf\u3001\u89c6\u9891\u3001\u591a\u6a21\u6001\u6761\u4ef6\u4e0b\u590d\u7528 DiT \u7684\u53ef\u6269\u5c55\u6027\uff0c\u9700\u8981\u7cfb\u7edf\u9a8c\u8bc1\u3002<br>3\uff09\u63a8\u7406\u901f\u5ea6\uff1a\u5373\u4f7f\u6a21\u578b\u66f4\u5f3a\uff0c\u6269\u6563\u91c7\u6837\u672c\u8eab\u4ecd\u6162\uff0c\u5982\u4f55\u4e0e\u84b8\u998f\u3001\u52a0\u901f\u91c7\u6837\u5668\uff08\u5982 DDIM\u3001EDM \u7cfb\u5217\uff09\u534f\u540c\u4ee5\u4fdd\u6301\u8d28\u91cf\u4e0e\u901f\u5ea6\u5e73\u8861\uff0c\u662f\u5de5\u7a0b\u843d\u5730\u5173\u952e\u3002<\/p>\n\n\n\n<p>\u5341\u4e00\u3001\u7ed3\u8bba\u4e0e\u542f\u793a<br>\u8bba\u6587\u4ee5\u4e25\u683c\u5bf9\u7167\u8bc1\u660e\uff1a\u6269\u6563\u6a21\u578b\u5e76\u4e0d\u4f9d\u8d56 U-Net \u7684\u5377\u79ef\u5f52\u7eb3\u504f\u7f6e\uff0c\u6807\u51c6\u5316\u7684 ViT \u67b6\u6784\u5728\u6f5c\u7a7a\u95f4\u540c\u6837\u80fd\u7a33\u5b9a\u8bad\u7ec3\u5e76\u5177\u5907\u4e00\u6d41\u53ef\u6269\u5c55\u6027\uff1b\u201c\u63d0\u5347\u524d\u5411 Gflops\u201d\u662f\u8d28\u91cf\u63d0\u5347\u7684\u6838\u5fc3\u65cb\u94ae\uff0c\u800c\u4e0d\u662f\u201c\u5355\u7eaf\u6269\u53c2\u6570\u201d\u3002\u7ed3\u5408 adaLN-Zero \u7b49\u8f7b\u91cf\u6761\u4ef6\u6ce8\u5165\u673a\u5236\uff0cDiT \u5728 ImageNet \u57fa\u51c6\u4e0a\u4ee5\u66f4\u4f18\u7b97\u529b\u6548\u7387\u5237\u65b0\u6210\u7ee9\uff0c\u4e3a\u201c\u8de8\u4efb\u52a1\u7edf\u4e00\u4e3b\u5e72\uff08Transformer\uff09\u201d\u63d0\u4f9b\u4e86\u5f3a\u529b\u8bc1\u636e\uff0c\u4e5f\u4e3a\u6587\u672c\u5230\u56fe\u50cf\u7b49\u66f4\u5e7f\u6cdb\u751f\u6210\u4efb\u52a1\u7684\u4e3b\u5e72\u66ff\u6362\u4e0e\u89c4\u6a21\u5316\u8bad\u7ec3\u5960\u5b9a\u4e86\u65b9\u6cd5\u8bba\u57fa\u7840\u3002<\/p>\n\n\n\n<p>\u5341\u4e8c\u3001\u7ed9\u7814\u53d1\u8005\u7684\u843d\u5730\u5efa\u8bae\uff08\u7ed3\u5408\u8bba\u6587\u89c2\u5bdf\uff09<br>1\uff09\u82e5\u5df2\u6709 LDM \u8bad\u7ec3\/\u63a8\u7406\u57fa\u7840\uff0c\u4f18\u5148\u5c1d\u8bd5\u5c06 U-Net \u4e3b\u5e72\u66ff\u6362\u4e3a ViT-style DiT\uff0c\u5e76\u4f7f\u7528 adaLN-Zero\u3002<br>2\uff09\u5728\u56fa\u5b9a\u663e\u5b58\u4e0b\uff0c\u4f18\u5148\u901a\u8fc7\u66f4\u5c0f\u7684 patch\uff08\u66f4\u591a token\uff09\u6765\u589e\u52a0\u201c\u6709\u6548\u7b97\u529b\u201d\uff1b\u82e5\u5e26\u5bbd\/\u663e\u5b58\u5403\u7d27\uff0c\u518d\u6298\u4e2d\u589e\u52a0\u5c42\u5bbd\/\u6df1\u3002<br>3\uff09\u4fdd\u8bc1\u8bc4\u6d4b\u4e00\u81f4\u6027\uff08\u5982\u7edf\u4e00 FID \u8bc4\u4f30\u7ba1\u7ebf\u3001\u56fa\u5b9a\u91c7\u6837\u6b65\u7b49\uff09\uff0c\u5e76\u7528\u201c\u603b\u8bad\u7ec3\u7b97\u529b\u201d\u4e0e\u201c\u5355\u56fe\u91c7\u6837\u7b97\u529b\u201d\u4e24\u7ef4\u6765\u5ea6\u91cf\u201c\u6027\u4ef7\u6bd4\u201d\uff0c\u907f\u514d\u5355\u770b\u53c2\u6570\u6216 FLOPs\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>Diffusion Transformer(DiT) on GitHub: <a href=\"https:\/\/github.com\/facebookresearch\/DiT\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/facebookresearch\/DiT<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6269\u6563\u6a21\u578b\u5728\u56fe\u50cf\u751f\u6210\u4e0a\u957f\u671f\u4ee5\u5377\u79ef\u5f0f U-Net \u4e3a\u4e3b\u5e72\uff0c\u4f46\u8bba\u6587Scalable Diffusion Models [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,95,90,80],"class_list":["post-6239","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-transformer","tag-multimodal","tag-cv"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u6269\u6563\u6a21\u578b\u5728\u56fe\u50cf\u751f\u6210\u4e0a\u957f\u671f\u4ee5\u5377\u79ef\u5f0f U-Net \u4e3a\u4e3b\u5e72\uff0c\u4f46\u8bba\u6587Scalable Diffusion Models&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6239","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6239"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6239\/revisions"}],"predecessor-version":[{"id":6240,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6239\/revisions\/6240"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6239"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6239"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6239"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}