diffusers/pr_13921/en/_app/immutable/nodes/73.f30d368b.js

import{s as le,n as fe,o as pe}from"../chunks/scheduler.53228c21.js";import{S as he,i as _e,e as d,s as r,c as u,h as ge,a as i,d as t,b as a,f as J,g as l,j as Z,k as j,l as f,m as o,n as p,t as h,o as _,p as g}from"../chunks/index.cac5d66a.js";import{C as ye}from"../chunks/CopyLLMTxtMenu.0ef49226.js";import{D as re}from"../chunks/Docstring.9de32ff4.js";import{C as Te}from"../chunks/CodeBlock.606cbaf4.js";import{H as ae,E as be}from"../chunks/MermaidChart.svelte_svelte_type_style_lang.48d5cb47.js";function Me(se){let m,P,E,V,T,R,b,q,M,de='A Diffusion Transformer model for <a href="https://github.com/Tencent-Hunyuan/HunyuanImage-2.1" rel="nofollow">HunyuanImage2.1</a>.',O,$,ie="The model can be loaded with the following code snippet.",U,v,X,x,S,s,D,ee,z,me='The Transformer model used in <a href="https://github.com/Tencent-Hunyuan/HunyuanImage-2.1" rel="nofollow">HunyuanImage-2.1</a>.',ne,y,H,te,L,ce='The <a href="/docs/diffusers/pr_13921/en/api/models/hunyuanimage_transformer_2d#diffusers.HunyuanImageTransformer2DModel">HunyuanImageTransformer2DModel</a> forward method.',Y,I,G,c,w,oe,N,ue='The output of <a href="/docs/diffusers/pr_13921/en/api/models/transformer2d#diffusers.Transformer2DModel">Transformer2DModel</a>.',A,k,F,W,Q;return T=new ye({props:{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"}}),b=new ae({props:{title:"HunyuanImageTransformer2DModel",local:"hunyuanimagetransformer2dmodel",headingTag:"h1"}}),v=new Te({props:{code:"ZnJvbSUyMGRpZmZ1c2VycyUyMGltcG9ydCUyMEh1bnl1YW5JbWFnZVRyYW5zZm9ybWVyMkRNb2RlbCUwQSUwQXRyYW5zZm9ybWVyJTIwJTNEJTIwSHVueXVhbkltYWdlVHJhbnNmb3JtZXIyRE1vZGVsLmZyb21fcHJldHJhaW5lZCglMjJodW55dWFudmlkZW8tY29tbXVuaXR5JTJGSHVueXVhbkltYWdlLTIuMS1EaWZmdXNlcnMlMjIlMkMlMjBzdWJmb2xkZXIlM0QlMjJ0cmFuc2Zvcm1lciUyMiUyQyUyMHRvcmNoX2R0eXBlJTNEdG9yY2guYmZsb2F0MTYp",highlighted:`<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> HunyuanImageTransformer2DModel

transformer = HunyuanImageTransformer2DModel.from_pretrained(<span class="hljs-string">&quot;hunyuanvideo-community/HunyuanImage-2.1-Diffusers&quot;</span>, subfolder=<span class="hljs-string">&quot;transformer&quot;</span>, torch_dtype=torch.bfloat16)`,lang:"python",wrap:!1}}),x=new ae({props:{title:"HunyuanImageTransformer2DModel",local:"diffusers.HunyuanImageTransformer2DModel",headingTag:"h2"}}),D=new re({props:{name:"class diffusers.HunyuanImageTransformer2DModel",anchor:"diffusers.HunyuanImageTransformer2DModel",parameters:[{name:"in_channels",val:": int = 64"},{name:"out_channels",val:": int = 64"},{name:"num_attention_heads",val:": int = 28"},{name:"attention_head_dim",val:": int = 128"},{name:"num_layers",val:": int = 20"},{name:"num_single_layers",val:": int = 40"},{name:"num_refiner_layers",val:": int = 2"},{name:"mlp_ratio",val:": float = 4.0"},{name:"patch_size",val:": tuple = (1, 1)"},{name:"qk_norm",val:": str = 'rms_norm'"},{name:"guidance_embeds",val:": bool = False"},{name:"text_embed_dim",val:": int = 3584"},{name:"text_embed_2_dim",val:": int | None = None"},{name:"rope_theta",val:": float = 256.0"},{name:"rope_axes_dim",val:": tuple = (64, 64)"},{name:"use_meanflow",val:": bool = False"}],parametersDescription:[{anchor:"diffusers.HunyuanImageTransformer2DModel.in_channels",description:`<strong>in_channels</strong> (<code>int</code>, defaults to <code>16</code>) &#x2014;
The number of channels in the input.`,name:"in_channels"},{anchor:"diffusers.HunyuanImageTransformer2DModel.out_channels",description:`<strong>out_channels</strong> (<code>int</code>, defaults to <code>16</code>) &#x2014;
The number of channels in the output.`,name:"out_channels"},{anchor:"diffusers.HunyuanImageTransformer2DModel.num_attention_heads",description:`<strong>num_attention_heads</strong> (<code>int</code>, defaults to <code>24</code>) &#x2014;
The number of heads to use for multi-head attention.`,name:"num_attention_heads"},{anchor:"diffusers.HunyuanImageTransformer2DModel.attention_head_dim",description:`<strong>attention_head_dim</strong> (<code>int</code>, defaults to <code>128</code>) &#x2014;
The number of channels in each head.`,name:"attention_head_dim"},{anchor:"diffusers.HunyuanImageTransformer2DModel.num_layers",description:`<strong>num_layers</strong> (<code>int</code>, defaults to <code>20</code>) &#x2014;
The number of layers of dual-stream blocks to use.`,name:"num_layers"},{anchor:"diffusers.HunyuanImageTransformer2DModel.num_single_layers",description:`<strong>num_single_layers</strong> (<code>int</code>, defaults to <code>40</code>) &#x2014;
The number of layers of single-stream blocks to use.`,name:"num_single_layers"},{anchor:"diffusers.HunyuanImageTransformer2DModel.num_refiner_layers",description:`<strong>num_refiner_layers</strong> (<code>int</code>, defaults to <code>2</code>) &#x2014;
The number of layers of refiner blocks to use.`,name:"num_refiner_layers"},{anchor:"diffusers.HunyuanImageTransformer2DModel.mlp_ratio",description:`<strong>mlp_ratio</strong> (<code>float</code>, defaults to <code>4.0</code>) &#x2014;
The ratio of the hidden layer size to the input size in the feedforward network.`,name:"mlp_ratio"},{anchor:"diffusers.HunyuanImageTransformer2DModel.patch_size",description:`<strong>patch_size</strong> (<code>int</code>, defaults to <code>2</code>) &#x2014;
The size of the spatial patches to use in the patch embedding layer.`,name:"patch_size"},{anchor:"diffusers.HunyuanImageTransformer2DModel.patch_size_t",description:`<strong>patch_size_t</strong> (<code>int</code>, defaults to <code>1</code>) &#x2014;
The size of the tmeporal patches to use in the patch embedding layer.`,name:"patch_size_t"},{anchor:"diffusers.HunyuanImageTransformer2DModel.qk_norm",description:`<strong>qk_norm</strong> (<code>str</code>, defaults to <code>rms_norm</code>) &#x2014;
The normalization to use for the query and key projections in the attention layers.`,name:"qk_norm"},{anchor:"diffusers.HunyuanImageTransformer2DModel.guidance_embeds",description:`<strong>guidance_embeds</strong> (<code>bool</code>, defaults to <code>True</code>) &#x2014;
Whether to use guidance embeddings in the model.`,name:"guidance_embeds"},{anchor:"diffusers.HunyuanImageTransformer2DModel.text_embed_dim",description:`<strong>text_embed_dim</strong> (<code>int</code>, defaults to <code>4096</code>) &#x2014;
Input dimension of text embeddings from the text encoder.`,name:"text_embed_dim"},{anchor:"diffusers.HunyuanImageTransformer2DModel.pooled_projection_dim",description:`<strong>pooled_projection_dim</strong> (<code>int</code>, defaults to <code>768</code>) &#x2014;
The dimension of the pooled projection of the text embeddings.`,name:"pooled_projection_dim"},{anchor:"diffusers.HunyuanImageTransformer2DModel.rope_theta",description:`<strong>rope_theta</strong> (<code>float</code>, defaults to <code>256.0</code>) &#x2014;
The value of theta to use in the RoPE layer.`,name:"rope_theta"},{anchor:"diffusers.HunyuanImageTransformer2DModel.rope_axes_dim",description:`<strong>rope_axes_dim</strong> (<code>tuple[int]</code>, defaults to <code>(16, 56, 56)</code>) &#x2014;
The dimensions of the axes to use in the RoPE layer.`,name:"rope_axes_dim"},{anchor:"diffusers.HunyuanImageTransformer2DModel.image_condition_type",description:`<strong>image_condition_type</strong> (<code>str</code>, <em>optional</em>, defaults to <code>None</code>) &#x2014;
The type of image conditioning to use. If <code>None</code>, no image conditioning is used. If <code>latent_concat</code>, the
image is concatenated to the latent stream. If <code>token_replace</code>, the image is used to replace first-frame
tokens in the latent stream and apply conditioning.`,name:"image_condition_type"}],source:"https://github.com/huggingface/diffusers/blob/vr_13921/src/diffusers/models/transformers/transformer_hunyuanimage.py#L617"}}),H=new re({props:{name:"forward",anchor:"diffusers.HunyuanImageTransformer2DModel.forward",parameters:[{name:"hidden_states",val:": Tensor"},{name:"timestep",val:": LongTensor"},{name:"encoder_hidden_states",val:": Tensor"},{name:"encoder_attention_mask",val:": Tensor"},{name:"timestep_r",val:": torch.LongTensor | None = None"},{name:"encoder_hidden_states_2",val:": torch.Tensor | None = None"},{name:"encoder_attention_mask_2",val:": torch.Tensor | None = None"},{name:"guidance",val:": torch.Tensor | None = None"},{name:"attention_kwargs",val:": dict[str, typing.Any] | None = None"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.hidden_states",description:`<strong>hidden_states</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, num_channels, height, width)</code> or <code>(batch_size, num_channels, num_frames, height, width)</code>) &#x2014;
Input <code>hidden_states</code>.`,name:"hidden_states"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.timestep",description:`<strong>timestep</strong> (<code>torch.LongTensor</code>) &#x2014;
Used to indicate denoising step.`,name:"timestep"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.encoder_hidden_states",description:`<strong>encoder_hidden_states</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, sequence_len, embed_dims)</code>) &#x2014;
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.`,name:"encoder_hidden_states"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.encoder_attention_mask",description:`<strong>encoder_attention_mask</strong> (<code>torch.Tensor</code>) &#x2014;
Mask applied to <code>encoder_hidden_states</code> during attention.`,name:"encoder_attention_mask"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.timestep_r",description:`<strong>timestep_r</strong> (<code>torch.LongTensor</code>, <em>optional</em>) &#x2014;
Refiner timestep conditioning.`,name:"timestep_r"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.encoder_hidden_states_2",description:`<strong>encoder_hidden_states_2</strong> (<code>torch.Tensor</code>, <em>optional</em>) &#x2014;
Additional conditional embeddings computed from a second text encoder.`,name:"encoder_hidden_states_2"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.encoder_attention_mask_2",description:`<strong>encoder_attention_mask_2</strong> (<code>torch.Tensor</code>, <em>optional</em>) &#x2014;
Mask applied to <code>encoder_hidden_states_2</code> during attention.`,name:"encoder_attention_mask_2"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.guidance",description:`<strong>guidance</strong> (<code>torch.Tensor</code>, <em>optional</em>) &#x2014;
Guidance scale embedding used for guidance-distilled variants of the model.`,name:"guidance"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.attention_kwargs",description:`<strong>attention_kwargs</strong> (<code>dict</code>, <em>optional</em>) &#x2014;
A kwargs dictionary that if specified is passed along to the <code>AttentionProcessor</code> as defined under
<code>self.processor</code> in
<a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py" rel="nofollow">diffusers.models.attention_processor</a>.`,name:"attention_kwargs"},{anchor:"diffusers.HunyuanImageTransformer2DModel.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether or not to return a <code>~models.transformer_2d.Transformer2DModelOutput</code> instead of a plain
tuple.`,name:"return_dict"}],source:"https://github.com/huggingface/diffusers/blob/vr_13921/src/diffusers/models/transformers/transformer_hunyuanimage.py#L743",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>If <code>return_dict</code> is True, an <code>~models.transformer_2d.Transformer2DModelOutput</code> is returned, otherwise a
<code>tuple</code> where the first element is the sample tensor.</p>
`}}),I=new ae({props:{title:"Transformer2DModelOutput",local:"diffusers.models.modeling_outputs.Transformer2DModelOutput",headingTag:"h2"}}),w=new re({props:{name:"class diffusers.models.modeling_outputs.Transformer2DModelOutput",anchor:"diffusers.models.modeling_outputs.Transformer2DModelOutput",parameters:[{name:"sample",val:": torch.Tensor"}],parametersDescription:[{anchor:"diffusers.models.modeling_outputs.Transformer2DModelOutput.sample",description:`<strong>sample</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, num_channels, height, width)</code> or <code>(batch size, num_vector_embeds - 1, num_latent_pixels)</code> if <a href="/docs/diffusers/pr_13921/en/api/models/transformer2d#diffusers.Transformer2DModel">Transformer2DModel</a> is discrete) &#x2014;
The hidden states output conditioned on the <code>encoder_hidden_states</code> input. If discrete, returns probability
distributions for the unnoised latent pixels.`,name:"sample"}],source:"https://github.com/huggingface/diffusers/blob/vr_13921/src/diffusers/models/modeling_outputs.py#L21"}}),k=new be({props:{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/en/api/models/hunyuanimage_transformer_2d.md"}}),{c(){m=d("meta"),P=r(),E=d("p"),V=r(),u(T.$$.fragment),R=r(),u(b.$$.fragment),q=r(),M=d("p"),M.innerHTML=de,O=r(),$=d("p"),$.textContent=ie,U=r(),u(v.$$.fragment),X=r(),u(x.$$.fragment),S=r(),s=d("div"),u(D.$$.fragment),ee=r(),z=d("p"),z.innerHTML=me,ne=r(),y=d("div"),u(H.$$.fragment),te=r(),L=d("p"),L.innerHTML=ce,Y=r(),u(I.$$.fragment),G=r(),c=d("div"),u(w.$$.fragment),oe=r(),N=d("p"),N.innerHTML=ue,A=r(),u(k.$$.fragment),F=r(),W=d("p"),this.h()},l(e){const n=ge("svelte-u9bgzb",document.head);m=i(n,"META",{name:!0,content:!0}),n.forEach(t),P=a(e),E=i(e,"P",{}),J(E).forEach(t),V=a(e),l(T.$$.fragment,e),R=a(e),l(b.$$.fragment,e),q=a(e),M=i(e,"P",{"data-svelte-h":!0}),Z(M)!=="svelte-a51oc8"&&(M.innerHTML=de),O=a(e),$=i(e,"P",{"data-svelte-h":!0}),Z($)!=="svelte-1vuni30"&&($.textContent=ie),U=a(e),l(v.$$.fragment,e),X=a(e),l(x.$$.fragment,e),S=a(e),s=i(e,"DIV",{class:!0});var C=J(s);l(D.$$.fragment,C),ee=a(C),z=i(C,"P",{"data-svelte-h":!0}),Z(z)!=="svelte-1e43fo9"&&(z.innerHTML=me),ne=a(C),y=i(C,"DIV",{class:!0});var B=J(y);l(H.$$.fragment,B),te=a(B),L=i(B,"P",{"data-svelte-h":!0}),Z(L)!=="svelte-25h5v7"&&(L.innerHTML=ce),B.forEach(t),C.forEach(t),Y=a(e),l(I.$$.fragment,e),G=a(e),c=i(e,"DIV",{class:!0});var K=J(c);l(w.$$.fragment,K),oe=a(K),N=i(K,"P",{"data-svelte-h":!0}),Z(N)!=="svelte-2clpd6"&&(N.innerHTML=ue),K.forEach(t),A=a(e),l(k.$$.fragment,e),F=a(e),W=i(e,"P",{}),J(W).forEach(t),this.h()},h(){j(m,"name","hf:doc:metadata"),j(m,"content",$e),j(y,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),j(s,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),j(c,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8")},m(e,n){f(document.head,m),o(e,P,n),o(e,E,n),o(e,V,n),p(T,e,n),o(e,R,n),p(b,e,n),o(e,q,n),o(e,M,n),o(e,O,n),o(e,$,n),o(e,U,n),p(v,e,n),o(e,X,n),p(x,e,n),o(e,S,n),o(e,s,n),p(D,s,null),f(s,ee),f(s,z),f(s,ne),f(s,y),p(H,y,null),f(y,te),f(y,L),o(e,Y,n),p(I,e,n),o(e,G,n),o(e,c,n),p(w,c,null),f(c,oe),f(c,N),o(e,A,n),p(k,e,n),o(e,F,n),o(e,W,n),Q=!0},p:fe,i(e){Q||(h(T.$$.fragment,e),h(b.$$.fragment,e),h(v.$$.fragment,e),h(x.$$.fragment,e),h(D.$$.fragment,e),h(H.$$.fragment,e),h(I.$$.fragment,e),h(w.$$.fragment,e),h(k.$$.fragment,e),Q=!0)},o(e){_(T.$$.fragment,e),_(b.$$.fragment,e),_(v.$$.fragment,e),_(x.$$.fragment,e),_(D.$$.fragment,e),_(H.$$.fragment,e),_(I.$$.fragment,e),_(w.$$.fragment,e),_(k.$$.fragment,e),Q=!1},d(e){e&&(t(P),t(E),t(V),t(R),t(q),t(M),t(O),t($),t(U),t(X),t(S),t(s),t(Y),t(G),t(c),t(A),t(F),t(W)),t(m),g(T,e),g(b,e),g(v,e),g(x,e),g(D),g(H),g(I,e),g(w),g(k,e)}}}const $e='{"title":"HunyuanImageTransformer2DModel","local":"hunyuanimagetransformer2dmodel","sections":[{"title":"HunyuanImageTransformer2DModel","local":"diffusers.HunyuanImageTransformer2DModel","sections":[],"depth":2},{"title":"Transformer2DModelOutput","local":"diffusers.models.modeling_outputs.Transformer2DModelOutput","sections":[],"depth":2}],"depth":1}';function ve(se){return pe(()=>{new URLSearchParams(window.location.search).get("fw")}),[]}class ze extends he{constructor(m){super(),_e(this,m,ve,Me,le,{})}}export{ze as component};