{"id":85918,"date":"2026-07-28T00:24:51","date_gmt":"2026-07-27T16:24:51","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/85918.html"},"modified":"2026-07-28T00:24:51","modified_gmt":"2026-07-27T16:24:51","slug":"sora-dit%e6%9e%b6%e6%9e%84%ef%bc%9a%e6%89%a9%e6%95%a3transformer%e8%a7%86%e9%a2%91%e7%94%9f%e6%88%90%e6%8a%80%e6%9c%af","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/85918.html","title":{"rendered":"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f"},"content":{"rendered":"

Sora\/DiT\u67b6\u6784：\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f<\/h2>\n

\u4e00\u3001\u5f15\u8a00<\/h3>\n

2024\u5e74OpenAI Sora\u6539\u53d8\u4e86\u89c6\u9891\u751f\u6210：1\u5206\u949f\u8fde\u8d2f\u89c6\u9891\u30013D\u4e00\u81f4\u6027\u3001\u6a21\u62df\u7269\u7406\u4e16\u754c\u3002Sora\u80cc\u540e\u662f DiT（Diffusion Transformer） \u67b6\u6784\u2014\u2014\u7528Transformer\u66ff\u4ee3UNet\u505a\u6269\u6563\u6a21\u578b\u7684\u9aa8\u5e72\u7f51\u7edc\u3002\u672c\u6587\u5c06\u6df1\u5165 DiT\/Sora \u7684\u6280\u672f\u539f\u7406\u3002<\/p>\n

\u4e8c\u3001DiT\u67b6\u6784\u8bbe\u8ba1<\/h3>\n

2.1 Patch\u5316 + Transformer<\/h4>\n

# DiT\u5c06\u89c6\u9891\u5e27\u5212\u5206\u4e3aPatches（\u7c7b\u4f3cViT）<\/span>
\n# \u8f93\u5165：\u539f\u59cb\u89c6\u9891 B\u00d7T\u00d7C\u00d7H\u00d7W \u2192 B\u00d7(T\u00b7H\u00b7W\/P\u00b3)\u00d7D<\/span><\/p>\n

class<\/span> PatchEmbed3D<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
\n """3D Patch\u5d4c\u5165：\u65f6\u95f4+\u7a7a\u95f4\u5206\u5757"""<\/span>
\n def<\/span> __init__<\/span>(<\/span>self,<\/span> patch_size=<\/span>(<\/span>2<\/span>,<\/span> 16<\/span>,<\/span> 16<\/span>)<\/span>,<\/span> in_channels=<\/span>3<\/span>,<\/span> embed_dim=<\/span>1152<\/span>)<\/span>:<\/span>
\n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
\n self.<\/span>patch_size =<\/span> patch_size<\/p>\n

# 3D\u5377\u79ef\u5206\u5757<\/span>
\n self.<\/span>proj =<\/span> nn.<\/span>Conv3d(<\/span>
\n in_channels,<\/span> embed_dim,<\/span>
\n kernel_size=<\/span>patch_size,<\/span>
\n stride=<\/span>patch_size
\n )<\/span><\/p>\n

def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
\n # x: [B, C, T, H, W] \u4f8b\u5982 [1, 3, 64, 256, 256]<\/span>
\n x =<\/span> self.<\/span>proj(<\/span>x)<\/span> # \u2192 [1, 1152, 32, 16, 16]<\/span>
\n B,<\/span> C,<\/span> T,<\/span> H,<\/span> W =<\/span> x.<\/span>shape
\n x =<\/span> x.<\/span>flatten(<\/span>2<\/span>)<\/span>.<\/span>transpose(<\/span>1<\/span>,<\/span> 2<\/span>)<\/span> # \u2192 [1, 8192, 1152]<\/span>
\n return<\/span> x<\/p>\n

2.2 \u81ea\u9002\u5e94Layer Norm (adaLN-Zero)<\/h4>\n

class<\/span> DiTBlock<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
\n """DiT\u6838\u5fc3\u5757：adaLN + \u591a\u5934\u6ce8\u610f\u529b + \u70b9\u5f0f\u524d\u9988"""<\/span><\/p>\n

def<\/span> __init__<\/span>(<\/span>self,<\/span> hidden_size=<\/span>1152<\/span>,<\/span> num_heads=<\/span>16<\/span>)<\/span>:<\/span>
\n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
\n self.<\/span>norm1 =<\/span> nn.<\/span>LayerNorm(<\/span>hidden_size,<\/span> elementwise_affine=<\/span>False<\/span>)<\/span>
\n self.<\/span>norm2 =<\/span> nn.<\/span>LayerNorm(<\/span>hidden_size,<\/span> elementwise_affine=<\/span>False<\/span>)<\/span>
\n self.<\/span>attn =<\/span> nn.<\/span>MultiheadAttention(<\/span>hidden_size,<\/span> num_heads,<\/span> batch_first=<\/span>True<\/span>)<\/span><\/p>\n

# MLP (SwiGLU)<\/span>
\n self.<\/span>mlp =<\/span> nn.<\/span>Sequential(<\/span>
\n nn.<\/span>Linear(<\/span>hidden_size,<\/span> 4<\/span> *<\/span> hidden_size)<\/span>,<\/span>
\n nn.<\/span>GELU(<\/span>)<\/span>,<\/span>
\n nn.<\/span>Linear(<\/span>4<\/span> *<\/span> hidden_size,<\/span> hidden_size)<\/span>,<\/span>
\n )<\/span><\/p>\n

# \u2605 adaLN: \u4ece\u65f6\u95f4\u6b65\u5d4c\u5165c\u751f\u6210\u8c03\u5236\u53c2\u6570<\/span>
\n # \u8f93\u51fa: 6\u00d7scale + 6\u00d7shift = 12\u00d7hidden_size<\/span>
\n self.<\/span>adaLN_modulation =<\/span> nn.<\/span>Sequential(<\/span>
\n nn.<\/span>SiLU(<\/span>)<\/span>,<\/span>
\n nn.<\/span>Linear(<\/span>hidden_size,<\/span> 12<\/span> *<\/span> hidden_size)<\/span>
\n )<\/span><\/p>\n

def<\/span> forward<\/span>(<\/span>self,<\/span> x,<\/span> c)<\/span>:<\/span>
\n # x: [B, N, D], c: [B, D] \u65f6\u95f4\u6b65\u6761\u4ef6<\/span>
\n # \u751f\u62106\u7ec4\u8c03\u5236\u53c2\u6570<\/span>
\n params =<\/span> self.<\/span>adaLN_modulation(<\/span>c)<\/span>.<\/span>chunk(<\/span>12<\/span>,<\/span> dim=<\/span>–<\/span>1<\/span>)<\/span>
\n shift_msa,<\/span> scale_msa,<\/span> gate_msa =<\/span> params[<\/span>:<\/span>3<\/span>]<\/span>
\n shift_mlp,<\/span> scale_mlp,<\/span> gate_mlp =<\/span> params[<\/span>3<\/span>:<\/span>6<\/span>]<\/span><\/p>\n

# \u591a\u5934\u6ce8\u610f\u529b（\u5e26\u8c03\u5236）<\/span>
\n x_norm =<\/span> self.<\/span>norm1(<\/span>x)<\/span> *<\/span> (<\/span>1<\/span> +<\/span> scale_msa.<\/span>unsqueeze(<\/span>1<\/span>)<\/span>)<\/span> +<\/span> shift_msa.<\/span>unsqueeze(<\/span>1<\/span>)<\/span>
\n attn_out,<\/span> _ =<\/span> self.<\/span>attn(<\/span>x_norm,<\/span> x_norm,<\/span> x_norm)<\/span>
\n x =<\/span> x +<\/span> gate_msa.<\/span>unsqueeze(<\/span>1<\/span>)<\/span> *<\/span> attn_out<\/p>\n

# MLP（\u5e26\u8c03\u5236）<\/span>
\n x_norm =<\/span> self.<\/span>norm2(<\/span>x)<\/span> *<\/span> (<\/span>1<\/span> +<\/span> scale_mlp.<\/span>unsqueeze(<\/span>1<\/span>)<\/span>)<\/span> +<\/span> shift_mlp.<\/span>unsqueeze(<\/span>1<\/span>)<\/span>
\n mlp_out =<\/span> self.<\/span>mlp(<\/span>x_norm)<\/span>
\n x =<\/span> x +<\/span> gate_mlp.<\/span>unsqueeze(<\/span>1<\/span>)<\/span> *<\/span> mlp_out<\/p>\n

return<\/span> x<\/p>\n

2.3 Full DiT Model<\/h4>\n

class<\/span> DiT<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
\n """\u5b8c\u6574DiT\u6a21\u578b"""<\/span><\/p>\n

def<\/span> __init__<\/span>(<\/span>self,<\/span> input_size=<\/span>32<\/span>,<\/span> patch_size=<\/span>2<\/span>,<\/span> in_channels=<\/span>4<\/span>,<\/span>
\n hidden_size=<\/span>1152<\/span>,<\/span> depth=<\/span>28<\/span>,<\/span> num_heads=<\/span>16<\/span>)<\/span>:<\/span>
\n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span><\/p>\n

# Patch\u5d4c\u5165<\/span>
\n self.<\/span>x_embedder =<\/span> PatchEmbed3D(<\/span>
\n patch_size=<\/span>(<\/span>1<\/span>,<\/span> patch_size,<\/span> patch_size)<\/span>,<\/span>
\n embed_dim=<\/span>hidden_size
\n )<\/span><\/p>\n

# \u65f6\u95f4\u6b65\u5d4c\u5165（\u5085\u91cc\u53f6\u7279\u5f81）<\/span>
\n self.<\/span>t_embedder =<\/span> TimestepEmbedder(<\/span>hidden_size)<\/span><\/p>\n

# \u6761\u4ef6\u5d4c\u5165（\u6587\u672c）<\/span>
\n self.<\/span>y_embedder =<\/span> LabelEmbedder(<\/span>768<\/span>,<\/span> hidden_size)<\/span> # T5-XXL<\/span><\/p>\n

# Transformer\u5757<\/span>
\n self.<\/span>blocks =<\/span> nn.<\/span>ModuleList(<\/span>[<\/span>
\n DiTBlock(<\/span>hidden_size,<\/span> num_heads)<\/span> for<\/span> _ in<\/span> range<\/span>(<\/span>depth)<\/span>
\n ]<\/span>)<\/span><\/p>\n

# \u8f93\u51fa\u6295\u5f71<\/span>
\n self.<\/span>final_layer =<\/span> FinalLayer(<\/span>
\n hidden_size,<\/span> patch_size,<\/span> in_channels
\n )<\/span><\/p>\n

# \u6743\u91cd\u521d\u59cb\u5316（DiT\u4e13\u7528）<\/span>
\n self.<\/span>initialize_weights(<\/span>)<\/span><\/p>\n

def<\/span> forward<\/span>(<\/span>self,<\/span> x,<\/span> t,<\/span> y)<\/span>:<\/span>
\n """x: [B,C,T,H,W], t: [B], y: [B,L,D_text]"""<\/span>
\n # 1. Patch\u5d4c\u5165 + \u4f4d\u7f6e\u7f16\u7801<\/span>
\n x =<\/span> self.<\/span>x_embedder(<\/span>x)<\/span> # [B, N, D]<\/span><\/p>\n

# 2. \u6761\u4ef6\u5d4c\u5165<\/span>
\n c =<\/span> self.<\/span>t_embedder(<\/span>t)<\/span> # \u65f6\u95f4\u6b65 \u2192 [B, D]<\/span>
\n if<\/span> y is<\/span> not<\/span> None<\/span>:<\/span>
\n c =<\/span> c +<\/span> self.<\/span>y_embedder(<\/span>y)<\/span> # \u6587\u672c \u2192 [B, D]<\/span><\/p>\n

# 3. Transformer\u5904\u7406<\/span>
\n for<\/span> block in<\/span> self.<\/span>blocks:<\/span>
\n x =<\/span> block(<\/span>x,<\/span> c)<\/span><\/p>\n

# 4. \u8f93\u51fa：\u9884\u6d4b\u566a\u58f0<\/span>
\n x =<\/span> self.<\/span>final_layer(<\/span>x,<\/span> c)<\/span> # [B, N, patch_size\u00b2*C]<\/span>
\n return<\/span> x<\/p>\n

# DiT-S\/2: depth=12, hidden=384, heads=6, 33M params<\/span>
\n# DiT-B\/2: depth=12, hidden=768, heads=12, 130M<\/span>
\n# DiT-L\/2: depth=24, hidden=1024, heads=16, 458M<\/span>
\n# DiT-XL\/2:depth=28, hidden=1152, heads=16, 675M<\/span><\/p>\n

\u56db\u3001Sora\u4e13\u6709\u7279\u6027<\/h3>\n

# Sora\u7684\u989d\u5916\u8bbe\u8ba1：<\/span><\/p>\n

# 1. \u539f\u751f\u53ef\u53d8\u5206\u8fa8\u7387<\/span>
\n# \u4e0d\u540c\u4e8e\u56fa\u5b9a256\u00d7256，Sora\u5904\u7406\u5404\u79cd\u5206\u8fa8\u73871280\u00d7720\u30011920\u00d71080<\/span>
\n# \u4f7f\u75283D Rotary Position Embedding (3D-RoPE)<\/span><\/p>\n

# 2. \u65f6\u7a7aCompression<\/span>
\n# \u5148\u901a\u8fc7VAE\u538b\u7f29\u5230\u6f5c\u7a7a\u95f4（T\u00d7H\u00d7W \u2192 T\/4\u00d7H\/8\u00d7W\/8）<\/span>
\n# \u964d\u7ef48\u00d78\u00d74=256x\u540e\u518d\u505a\u6269\u6563<\/span><\/p>\n

# 3. LLM\u9a71\u52a8\u7684\u89c6\u9891\u7406\u89e3<\/span>
\n# \u7528GPT-4\/DALL-E 3 re-caption\u6240\u6709\u8bad\u7ec3\u89c6\u9891<\/span>
\n# \u9ad8\u8d28\u91cf\u6587\u672c\u6761\u4ef6(200+ tokens)\u66ff\u4ee3\u7b80\u5355\u6807\u7b7e<\/span><\/p>\n

# 4. \u8054\u5408\u8bad\u7ec3<\/span>
\n# \u56fe\u50cf+\u89c6\u9891\u8054\u5408\u8bad\u7ec3 \u2192 \u56fe\u50cf\u8d28\u91cf\u63d0\u5347\u89c6\u9891\u7ec6\u8282<\/span>
\n# \u8bad\u7ec3\u65f6\u52a8\u6001padding\u5230\u7edf\u4e00patch\u6570<\/span><\/p>\n

\u4e94\u3001\u63a8\u7406\u4f18\u5316<\/h3>\n

# DDIM\u52a0\u901f：50\u6b65 \u2192 20\u6b65<\/span>
\nfrom<\/span> diffusers import<\/span> DPMSolverMultistepScheduler
\nscheduler =<\/span> DPMSolverMultistepScheduler.<\/span>from_config(<\/span>
\n model.<\/span>scheduler.<\/span>config,<\/span>
\n algorithm_type=<\/span>"dpmsolver++"<\/span>,<\/span>
\n solver_order=<\/span>3<\/span>
\n)<\/span><\/p>\n

# \u6a21\u578b\u5e76\u884c（DiT-XL\u5355\u5361\u653e\u4e0d\u4e0b）<\/span>
\n# Tensor Parallelism (TP) + Sequence Parallelism (SP)<\/span>
\n# TP: \u6bcf\u4e2ahead\u5728\u4e0d\u540cGPU<\/span>
\n# SP: \u957f\u5e8f\u5217\u5206\u6563\u5230\u591aGPU\u7684LayerNorm\/Dropout<\/span><\/p>\n

# \u5185\u5b58\u4f18\u5316<\/span>
\npipe.<\/span>enable_model_cpu_offload(<\/span>)<\/span> # \u4e0d\u6d3b\u8dc3\u90e8\u5206\u653eCPU<\/span>
\npipe.<\/span>enable_vae_slicing(<\/span>)<\/span> # VAE\u5206\u5757\u89e3\u7801<\/span>
\npipe.<\/span>enable_vae_tiling(<\/span>)<\/span> # \u50cf\u7d20\u7a7a\u95f4\u5206\u5757<\/span><\/p>\n

# \u7f16\u8bd1\u52a0\u901f<\/span>
\nimport<\/span> torch.<\/span>_dynamo
\npipe.<\/span>unet =<\/span> torch.<\/span>compile<\/span>(<\/span>pipe.<\/span>unet,<\/span> mode=<\/span>"reduce-overhead"<\/span>)<\/span><\/p>\n

\u516d\u3001DiT vs UNet<\/h3>\n\n\u7279\u6027UNet (SD3)DiT (Sora)<\/tr>\n\n\n\n\n\n\n
\u67b6\u6784<\/td>\nCNN+CrossAttn<\/td>\nPure Transformer<\/td>\n<\/tr>\n
\u5206\u8fa8\u7387<\/td>\n\u56fa\u5b9a<\/td>\n\u539f\u751f\u53ef\u53d8<\/td>\n<\/tr>\n
\u89c4\u6a21\u6269\u5c55<\/td>\n\u6b21\u7ebf\u6027<\/td>\n\u7ebf\u6027(DiT scaling law)<\/td>\n<\/tr>\n
\u8bad\u7ec3\u6548\u7387<\/td>\n\u9ad8(\u5c0f\u89c4\u6a21)<\/td>\n\u9ad8(\u5927\u89c4\u6a21)<\/td>\n<\/tr>\n
FID (256\u00d7256)<\/td>\n2.2<\/td>\n2.27<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4e03\u3001\u603b\u7ed3<\/h3>\n

DiT\/Sora\u7684\u6838\u5fc3\u8bbe\u8ba1：<\/p>\n

  • Patch\u5316 \u2192 \u7edf\u4e00\u7684token\u8868\u793a<\/li>\n
  • adaLN-Zero \u2192 \u6761\u4ef6\u6ce8\u5165\u7684\u4f18\u96c5\u65b9\u5f0f<\/li>\n
  • 3D\u538b\u7f29 \u2192 \u65f6\u7a7a\u6f5c\u7a7a\u95f4\u964d\u7ef4<\/li>\n
  • \u539f\u751f\u53ef\u53d8\u5206\u8fa8\u7387 \u2192 \u544a\u522b\u56fa\u5b9a\u5c3a\u5bf8<\/li>\n
  • Scaling Law \u2192 Transformer\u7684\u89c4\u6a21\u7ea2\u5229<\/li>\n","protected":false},"excerpt":{"rendered":"

    Sora\/DiT\u67b6\u6784：\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f
    \n\u4e00\u3001\u5f15\u8a00
    \n2024\u5e74OpenAI Sora\u6539\u53d8\u4e86\u89c6\u9891\u751f\u6210：1\u5206\u949f\u8fde\u8d2f\u89c6\u9891\u30013D\u4e00\u81f4\u6027\u3001\u6a21\u62df\u7269\u7406\u4e16\u754c\u3002Sora\u80cc\u540e\u662f DiT（Diffusion Transformer） \u67b6\u6784\u2014\u2014\u7528Transformer\u66ff\u4ee3UNet\u505a\u6269\u6563\u6a21\u578b\u7684\u9aa8\u5e72\u7f51\u7edc\u3002\u672c\u6587\u5c06\u6df1\u5165 DiT\/Sora \u7684\u6280\u672f\u539f\u7406\u3002
    \n\u4e8c\u3001DiT\u67b6\u6784\u8bbe\u8ba1
    \n2.1 Patch\u5316 Transformer
    \n# DiT\u5c06\u89c6<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[841,86,214],"topic":[],"class_list":["post-85918","post","type-post","status-publish","format-standard","hentry","category-server","tag-transformer","tag-86","tag-214"],"yoast_head":"\nSora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/85918.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"Sora\/DiT\u67b6\u6784：\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f \u4e00\u3001\u5f15\u8a00 2024\u5e74OpenAI Sora\u6539\u53d8\u4e86\u89c6\u9891\u751f\u6210：1\u5206\u949f\u8fde\u8d2f\u89c6\u9891\u30013D\u4e00\u81f4\u6027\u3001\u6a21\u62df\u7269\u7406\u4e16\u754c\u3002Sora\u80cc\u540e\u662f DiT（Diffusion Transformer） \u67b6\u6784\u2014\u2014\u7528Transformer\u66ff\u4ee3UNet\u505a\u6269\u6563\u6a21\u578b\u7684\u9aa8\u5e72\u7f51\u7edc\u3002\u672c\u6587\u5c06\u6df1\u5165 DiT\/Sora \u7684\u6280\u672f\u539f\u7406\u3002 \u4e8c\u3001DiT\u67b6\u6784\u8bbe\u8ba1 2.1 Patch\u5316 Transformer # DiT\u5c06\u89c6\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/85918.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-27T16:24:51+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"3 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/85918.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/85918.html\",\"name\":\"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-27T16:24:51+00:00\",\"dateModified\":\"2026-07-27T16:24:51+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/85918.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/85918.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/85918.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/85918.html","og_locale":"zh_CN","og_type":"article","og_title":"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"Sora\/DiT\u67b6\u6784：\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f \u4e00\u3001\u5f15\u8a00 2024\u5e74OpenAI Sora\u6539\u53d8\u4e86\u89c6\u9891\u751f\u6210：1\u5206\u949f\u8fde\u8d2f\u89c6\u9891\u30013D\u4e00\u81f4\u6027\u3001\u6a21\u62df\u7269\u7406\u4e16\u754c\u3002Sora\u80cc\u540e\u662f DiT（Diffusion Transformer） \u67b6\u6784\u2014\u2014\u7528Transformer\u66ff\u4ee3UNet\u505a\u6269\u6563\u6a21\u578b\u7684\u9aa8\u5e72\u7f51\u7edc\u3002\u672c\u6587\u5c06\u6df1\u5165 DiT\/Sora \u7684\u6280\u672f\u539f\u7406\u3002 \u4e8c\u3001DiT\u67b6\u6784\u8bbe\u8ba1 2.1 Patch\u5316 Transformer # DiT\u5c06\u89c6","og_url":"https:\/\/www.wsisp.com\/helps\/85918.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-27T16:24:51+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"3 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/85918.html","url":"https:\/\/www.wsisp.com\/helps\/85918.html","name":"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-27T16:24:51+00:00","dateModified":"2026-07-27T16:24:51+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/85918.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/85918.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/85918.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Sora\/DiT\u67b6\u6784\uff1a\u6269\u6563Transformer\u89c6\u9891\u751f\u6210\u6280\u672f"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/85918","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=85918"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/85918\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=85918"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=85918"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=85918"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=85918"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}