Buckets:

hf-doc-build
/

doc-dev

hf-doc-build/doc-dev / diffusers /pr_14313 /zh /_app /immutable /nodes /7.DPuSJth0.js

HuggingFaceDocBuilder's picture

HuggingFaceDocBuilder

13 days ago

41 kB

	import"../chunks/DsnmJJEf.js";import{i as v,h as A,C as z,H as s,b as e,E as q,s as L}from"../chunks/Dv-EnAnp.js";import{p as x,o as H,s as l,f as P,a as Y,b as K,c as N,n as O}from"../chunks/F-c_tsrl.js";const $='{"title":"快速入门","local":"快速入门","sections":[{"title":"ModularPipelineBlocks","local":"modularpipelineblocks","sections":[],"depth":2},{"title":"管道和块状态","local":"管道和块状态","sections":[],"depth":2},{"title":"自定义块","local":"自定义块","sections":[{"title":"prepare_latents","local":"preparelatents","sections":[],"depth":3},{"title":"去噪","local":"去噪","sections":[],"depth":3}],"depth":2},{"title":"组装块","local":"组装块","sections":[],"depth":2},{"title":"ModularPipeline","local":"modularpipeline","sections":[],"depth":2},{"title":"添加工作流","local":"添加工作流","sections":[{"title":"IP-Adapter","local":"ip-adapter","sections":[],"depth":3},{"title":"ControlNet","local":"controlnet","sections":[],"depth":3},{"title":"AutoPipelineBlocks","local":"autopipelineblocks","sections":[],"depth":3}],"depth":2},{"title":"分享","local":"分享","sections":[],"depth":2}],"depth":1}';var ll=N('<meta name="hf:doc:metadata"/>'),el=N('<p></p> <!> <!> <p>模块化Diffusers是一个快速构建灵活和可定制管道的框架。模块化Diffusers的核心是<code>ModularPipelineBlocks</code>，可以与其他块组合以适应新的工作流程。这些块被转换为<code>ModularPipeline</code>，一个开发者可以使用的友好用户界面。</p> <p>本文档将向您展示如何使用模块化框架实现<a href="https://differential-diffusion.github.io/" rel="nofollow">Differential Diffusion</a>管道。</p> <!> <p><code>ModularPipelineBlocks</code>是<em>定义</em>，指定管道中单个步骤的组件、输入、输出和计算逻辑。有四种类型的块。</p> <ul><li><code>ModularPipelineBlocks</code>是最基本的单一步骤块。</li> <li><code>SequentialPipelineBlocks</code>是一个多块，线性组合其他块。一个块的输出是下一个块的输入。</li> <li><code>LoopSequentialPipelineBlocks</code>是一个多块，迭代运行，专为迭代工作流程设计。</li> <li><code>AutoPipelineBlocks</code>是一个针对不同工作流程的块集合，它根据输入选择运行哪个块。它旨在方便地将多个工作流程打包到单个管道中。</li></ul> <p><a href="https://differential-diffusion.github.io/" rel="nofollow">Differential Diffusion</a>是一个图像到图像的工作流程。从<code>IMAGE2IMAGE_BLOCKS</code>预设开始，这是一个用于图像到图像生成的<code>ModularPipelineBlocks</code>集合。</p> <!> <!> <p>模块化Diffusers使用<em>状态</em>在块之间通信数据。有两种类型的状态。</p> <ul><li><code>PipelineState</code>是一个全局状态，可用于跟踪所有块的所有输入和输出。</li> <li><code>BlockState</code>是<code>PipelineState</code>中相关变量的局部视图，用于单个块。</li></ul> <!> <p><a href="https://differential-diffusion.github.io/" rel="nofollow">Differential Diffusion</a> 与标准的图像到图像转换在其 <code>prepare_latents</code> 和 <code>denoise</code> 块上有所不同。所有其他块都可以重用，但你需要修改这两个。</p> <p>通过复制和修改现有的块，为 <code>prepare_latents</code> 和 <code>denoise</code> 创建占位符 <code>ModularPipelineBlocks</code>。</p> <p>打印 <code>denoise</code> 块，可以看到它由 <code>LoopSequentialPipelineBlocks</code> 组成，包含三个子块，<code>before_denoiser</code>、<code>denoiser</code> 和 <code>after_denoiser</code>。只需要修改 <code>before_denoiser</code> 子块，根据变化图为去噪器准备潜在输入。</p> <!> <p>用新的 <code>SDXLDiffDiffLoopBeforeDenoiser</code> 块替换 <code>StableDiffusionXLLoopBeforeDenoiser</code> 子块。</p> <!> <!> <p><code>prepare_latents</code> 块需要进行以下更改。</p> <ul><li>一个处理器来处理变化图</li> <li>一个新的 <code>inputs</code> 来接受用户提供的变化图，<code>timestep</code> 用于预计算所有潜在变量和 <code>num_inference_steps</code> 来创建更新图像区域的掩码</li> <li>更新 <code>__call__</code> 方法中的计算，用于处理变化图和创建掩码，并将其存储在 <code>BlockState</code> 中</li></ul> <!> <!> <p><code>before_denoiser</code> 子块需要进行以下更改。</p> <ul><li>新的 <code>inputs</code> 以接受 <code>denoising_start</code> 参数，<code>original_latents</code> 和 <code>diffdiff_masks</code> 来自 <code>prepare_latents</code> 块</li> <li>更新 <code>__call__</code> 方法中的计算以应用 Differential Diffusion</li></ul> <!> <!> <p>此时，您应该拥有创建 <code>ModularPipeline</code> 所需的所有块。</p> <p>复制现有的 <code>IMAGE2IMAGE_BLOCKS</code> 预设，对于 <code>set_timesteps</code> 块，使用 <code>TEXT2IMAGE_BLOCKS</code> 中的 <code>set_timesteps</code>，因为 Differential Diffusion 不需要 <code>strength</code> 参数。</p> <p>将 <code>prepare_latents</code> 和 <code>denoise</code> 块设置为您刚刚修改的 <code>SDXLDiffDiffPrepareLatentsStep</code> 和 <code>SDXLDiffDiffDenoiseStep</code> 块。</p> <p>调用 <code>SequentialPipelineBlocks.from_blocks_dict()</code> 在块上创建一个 <code>SequentialPipelineBlocks</code>。</p> <!> <!> <p>将 <code>SequentialPipelineBlocks</code> 转换为 <code>ModularPipeline</code>，使用 <code>ModularPipeline.init_pipeline</code> 方法。这会初始化从 <code>modular_model_index.json</code> 文件加载的预期组件。通过调用 <code>ModularPipeline.load_defau lt_components</code>。</p> <p>初始化<code>ComponentManager</code>时传入pipeline是一个好主意，以帮助管理不同的组件。一旦调用<code>load_components()</code>，组件就会被注册到<code>ComponentManager</code>中，并且可以在工作流之间共享。下面的例子使用<code>collection</code>参数为组件分配了一个<code>"diffdiff"</code>标签，以便更好地组织。</p> <!> <!> <p>可以向<code>ModularPipeline</code>添加其他工作流以支持更多功能，而无需从头重写整个pipeline。</p> <p>本节演示如何添加IP-Adapter或ControlNet。</p> <!> <p>Stable Diffusion XL已经有一个预设的IP-Adapter块，你可以使用，并且不需要对现有的Differential Diffusion pipeline进行任何更改。</p> <!> <p>使用<code>sub_blocks.insert</code>方法将其插入到<code>ModularPipeline</code>中。下面的例子在位置<code>0</code>插入了<code>ip_adapter_block</code>。打印pipeline可以看到<code>ip_adapter_block</code>被添加了，并且它需要一个<code>ip_adapter_image</code>。这也向pipeline添加了两个组件，<code>image_encoder</code>和<code>feature_extractor</code>。</p> <!> <p>调用<code>~ModularPipeline.init_pipeline</code>来初始化一个<code>ModularPipeline</code>，并使用<code>load_components()</code>加载模型组件。加载并设置IP-Adapter以运行pipeline。</p> <!> <!> <p>Stable Diffusion XL 已经预设了一个可以立即使用的 ControlNet 块。</p> <!> <p>然而，它需要修改 <code>denoise</code> 块，因为那是 ControlNet 将控制信息注入到 UNet 的地方。</p> <p>通过将 <code>StableDiffusionXLLoopDenoiser</code> 子块替换为 <code>StableDiffusionXLControlNetLoopDenoiser</code> 来修改 <code>denoise</code> 块。</p> <!> <p>插入 <code>controlnet_input</code> 块并用新的 <code>controlnet_denoise_block</code> 替换 <code>denoise</code> 块。初始化一个 <code>ModularPipeline</code> 并将 <code>load_components()</code> 加载到其中。</p> <!> <!> <p>差分扩散、IP-Adapter 和 ControlNet 工作流可以通过使用 <code>AutoPipelineBlocks</code> 捆绑到一个单一的 <code>ModularPipeline</code> 中。这允许根据输入如 <code>control_image</code> 或 <code>ip_adapter_image</code> 自动选择要运行的子块。如果没有传递这些输入，则默认为差分扩散。</p> <p>使用 <code>block_trigger_inputs</code> 仅在提供 <code>control_image</code> 输入时运行 <code>SDXLDiffDiffControlNetDenoiseStep</code> 块。否则，使用 <code>SDXLDiffDiffDenoiseStep</code>。</p> <!> <p>添加 <code>ip_adapter</code> 和 <code>controlnet_input</code> 块。</p> <!> <p>调用 <code>SequentialPipelineBlocks.from_blocks_dict()</code> 来创建一个 <code>SequentialPipelineBlocks</code> 并创建一个 <code>ModularPipeline</code> 并加载模型组件以运行。</p> <!> <!> <p>使用 <code>save_pretrained()</code> 将您的 <code>ModularPipeline</code> 添加到 Hub，并将 <code>push_to_hub</code> 参数设置为 <code>True</code>。</p> <!> <p>其他用户可以使用 <code>from_pretrained()</code> 加载 <code>ModularPipeline</code>。</p> <!> <!> <p></p>',1);function tl(Q,E){x(E,!1),H(()=>{new URLSearchParams(window.location.search).get("fw")}),v();var a=el();A("1x94atl",S=>{var F=ll();L(F,"content",$),Y(S,F)});var o=l(P(a),2);z(o,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var t=l(o,2);s(t,{title:"快速入门",local:"快速入门",headingTag:"h1"});var n=l(t,6);s(n,{title:"ModularPipelineBlocks",local:"modularpipelineblocks",headingTag:"h2"});var i=l(n,8);e(i,{code:"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",highlighted:`<span class="hljs-keyword">from</span> diffusers.modular_pipelines.stable_diffusion_xl <span class="hljs-keyword">import</span> IMAGE2IMAGE_BLOCKS
	IMAGE2IMAGE_BLOCKS = InsertableDict([
	(<span class="hljs-string">"text_encoder"</span>, StableDiffusionXLTextEncoderStep),
	(<span class="hljs-string">"image_encoder"</span>, StableDiffusionXLVaeEncoderStep),
	(<span class="hljs-string">"input"</span>, StableDiffusionXLInputStep),
	(<span class="hljs-string">"set_timesteps"</span>, StableDiffusionXLImg2ImgSetTimestepsStep),
	(<span class="hljs-string">"prepare_latents"</span>, StableDiffusionXLImg2ImgPrepareLatentsStep),
	(<span class="hljs-string">"prepare_add_cond"</span>, StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep),
	(<span class="hljs-string">"denoise"</span>, StableDiffusionXLDenoiseStep),
	(<span class="hljs-string">"decode"</span>, StableDiffusionXLDecodeStep)
	])`,lang:"py",wrap:!1});var p=l(i,2);s(p,{title:"管道和块状态",local:"管道和块状态",headingTag:"h2"});var d=l(p,6);s(d,{title:"自定义块",local:"自定义块",headingTag:"h2"});var M=l(d,8);e(M,{code:"ZGVub2lzZV9ibG9ja3MlMjAlM0QlMjBJTUFHRTJJTUFHRV9CTE9DS1MlNUIlMjJkZW5vaXNlJTIyJTVEKCklMEFwcmludChkZW5vaXNlX2Jsb2Nrcyk=",highlighted:`denoise_blocks = IMAGE2IMAGE_BLOCKS[<span class="hljs-string">"denoise"</span>]()
	<span class="hljs-built_in">print</span>(denoise_blocks)`,lang:"py",wrap:!1});var c=l(M,4);e(c,{code:"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",highlighted:`<span class="hljs-comment"># 复制现有块作为占位符</span>
	<span class="hljs-keyword">class</span> <span class="hljs-title class_">SDXLDiffDiffPrepareLatentsStep</span>(<span class="hljs-title class_ inherited__">ModularPipelineBlocks</span>):
	<span class="hljs-string">"""Copied from StableDiffusionXLImg2ImgPrepareLatentsStep - will modify later"""</span>
	<span class="hljs-comment"># ... 与 StableDiffusionXLImg2ImgPrepareLatentsStep 相同的实现</span>

	<span class="hljs-keyword">class</span> <span class="hljs-title class_">SDXLDiffDiffDenoiseStep</span>(<span class="hljs-title class_ inherited__">StableDiffusionXLDenoiseLoopWrapper</span>):
	block_classes = [SDXLDiffDiffLoopBeforeDenoiser, StableDiffusionXLLoopDenoiser, StableDiffusionXLLoopAfterDenoiser]
	block_names = [<span class="hljs-string">"before_denoiser"</span>, <span class="hljs-string">"denoiser"</span>, <span class="hljs-string">"after_denoiser"</span>]`,lang:"py",wrap:!1});var J=l(c,2);s(J,{title:"prepare_latents",local:"preparelatents",headingTag:"h3"});var U=l(J,6);e(U,{code:"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",highlighted:`class SDXLDiffDiffPrepareLatentsStep(ModularPipelineBlocks):
	@property
	def expected_components(self) -> List[ComponentSpec]:
	return [
	ComponentSpec("vae", AutoencoderKL),
	ComponentSpec("scheduler", EulerDiscreteScheduler),
	<span class="hljs-addition">+ ComponentSpec("mask_processor", VaeImageProcessor, config=FrozenDict({"do_normalize": False, "do_convert_grayscale": True}))</span>
	]
	@property
	def inputs(self) -> List[Tuple[str, Any]]:
	return [
	InputParam("generator"),
	<span class="hljs-addition">+ InputParam("diffdiff_map", required=True),</span>
	<span class="hljs-deletion">- InputParam("latent_timestep", required=True, type_hint=torch.Tensor),</span>
	<span class="hljs-addition">+ InputParam("timesteps", type_hint=torch.Tensor),</span>
	<span class="hljs-addition">+ InputParam("num_inference_steps", type_hint=int),</span>
	]

	@property
	def intermediate_outputs(self) -> List[OutputParam]:
	return [
	<span class="hljs-addition">+ OutputParam("original_latents", type_hint=torch.Tensor),</span>
	<span class="hljs-addition">+ OutputParam("diffdiff_masks", type_hint=torch.Tensor),</span>
	]
	def __call__(self, components, state: PipelineState):
	# ... existing logic ...
	<span class="hljs-addition">+ # Process change map and create masks</span>
	<span class="hljs-addition">+ diffdiff_map = components.mask_processor.preprocess(block_state.diffdiff_map, height=latent_height, width=latent_width)</span>
	<span class="hljs-addition">+ thresholds = torch.arange(block_state.num_inference_steps, dtype=diffdiff_map.dtype) / block_state.num_inference_steps</span>
	<span class="hljs-addition">+ block_state.diffdiff_masks = diffdiff_map > (thresholds + (block_state.denoising_start or 0))</span>
	<span class="hljs-addition">+ block_state.original_latents = block_state.latents</span>`,lang:"diff",wrap:!1});var y=l(U,2);s(y,{title:"去噪",local:"去噪",headingTag:"h3"});var T=l(y,6);e(T,{code:"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",highlighted:`class SDXLDiffDiffLoopBeforeDenoiser(ModularPipelineBlocks):
	@property
	def description(self) -> str:
	return (
	"Step within the denoising loop for differential diffusion that prepare the latent input for the denoiser"
	)

	@property
	def inputs(self) -> List[str]:
	return [
	InputParam("latents", required=True, type_hint=torch.Tensor),
	<span class="hljs-addition">+ InputParam("denoising_start"),</span>
	<span class="hljs-addition">+ InputParam("original_latents", type_hint=torch.Tensor),</span>
	<span class="hljs-addition">+ InputParam("diffdiff_masks", type_hint=torch.Tensor),</span>
	]

	def __call__(self, components, block_state, i, t):
	<span class="hljs-addition">+ # Apply differential diffusion logic</span>
	<span class="hljs-addition">+ if i == 0 and block_state.denoising_start is None:</span>
	<span class="hljs-addition">+ block_state.latents = block_state.original_latents[:1]</span>
	<span class="hljs-addition">+ else:</span>
	<span class="hljs-addition">+ block_state.mask = block_state.diffdiff_masks[i].unsqueeze(0).unsqueeze(1)</span>
	<span class="hljs-addition">+ block_state.latents = block_state.original_latents[i] * block_state.mask + block_state.latents * (1 - block_state.mask)</span>

	# ... rest of existing logic ...`,lang:"diff",wrap:!1});var r=l(T,2);s(r,{title:"组装块",local:"组装块",headingTag:"h2"});var m=l(r,10);e(m,{code:"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",highlighted:`DIFFDIFF_BLOCKS = IMAGE2IMAGE_BLOCKS.copy()
	DIFFDIFF_BLOCKS[<span class="hljs-string">"set_timesteps"</span>] = TEXT2IMAGE_BLOCKS[<span class="hljs-string">"set_timesteps"</span>]
	DIFFDIFF_BLOCKS[<span class="hljs-string">"prepare_latents"</span>] = SDXLDiffDiffPrepareLatentsStep
	DIFFDIFF_BLOCKS[<span class="hljs-string">"denoise"</span>] = SDXLDiffDiffDenoiseStep

	dd_blocks = SequentialPipelineBlocks.from_blocks_dict(DIFFDIFF_BLOCKS)
	<span class="hljs-built_in">print</span>(dd_blocks)`,lang:"py",wrap:!1});var Z=l(m,2);s(Z,{title:"ModularPipeline",local:"modularpipeline",headingTag:"h2"});var u=l(Z,6);e(u,{code:"ZnJvbSUyMGRpZmZ1c2Vycy5tb2R1bGFyX3BpcGVsaW5lcyUyMGltcG9ydCUyMENvbXBvbmVudHNNYW5hZ2VyJTBBJTBBY29tcG9uZW50cyUyMCUzRCUyMENvbXBvbmVudE1hbmFnZXIoKSUwQSUwQWRkX3BpcGVsaW5lJTIwJTNEJTIwZGRfYmxvY2tzLmluaXRfcGlwZWxpbmUoJTIyWWlZaVh1JTJGbW9kdWxhci1kZW1vLWF1dG8lMjIlMkMlMjBjb21wb25lbnRzX21hbmFnZXIlM0Rjb21wb25lbnRzJTJDJTIwY29sbGVjdGlvbiUzRCUyMmRpZmZkaWZmJTIyKSUwQWRkX3BpcGVsaW5lLmxvYWRfZGVmYXVsdF9jb21wb25lbmV0cyhkdHlwZSUzRHRvcmNoLmZsb2F0MTYpJTBBZGRfcGlwZWxpbmUudG8oJTIyY3VkYSUyMik=",highlighted:`<span class="hljs-keyword">from</span> diffusers.modular_pipelines <span class="hljs-keyword">import</span> ComponentsManager

	components = ComponentManager()

	dd_pipeline = dd_blocks.init_pipeline(<span class="hljs-string">"YiYiXu/modular-demo-auto"</span>, components_manager=components, collection=<span class="hljs-string">"diffdiff"</span>)
	dd_pipeline.load_default_componenets(dtype=torch.float16)
	dd_pipeline.to(<span class="hljs-string">"cuda"</span>)`,lang:"py",wrap:!1});var w=l(u,2);s(w,{title:"添加工作流",local:"添加工作流",headingTag:"h2"});var b=l(w,6);s(b,{title:"IP-Adapter",local:"ip-adapter",headingTag:"h3"});var I=l(b,4);e(I,{code:"ZnJvbSUyMGRpZmZ1c2Vycy5tb2R1bGFyX3BpcGVsaW5lcy5zdGFibGVfZGlmZnVzaW9uX3hsLmVuY29kZXJzJTIwaW1wb3J0JTIwU3RhYmxlRGlmZnVzaW9uWExBdXRvSVBBZGFwdGVyU3RlcCUwQSUwQWlwX2FkYXB0ZXJfYmxvY2slMjAlM0QlMjBTdGFibGVEaWZmdXNpb25YTEF1dG9JUEFkYXB0ZXJTdGVwKCk=",highlighted:`<span class="hljs-keyword">from</span> diffusers.modular_pipelines.stable_diffusion_xl.encoders <span class="hljs-keyword">import</span> StableDiffusionXLAutoIPAdapterStep

	ip_adapter_block = StableDiffusionXLAutoIPAdapterStep()`,lang:"py",wrap:!1});var f=l(I,4);e(f,{code:"ZGRfYmxvY2tzLnN1Yl9ibG9ja3MuaW5zZXJ0KCUyMmlwX2FkYXB0ZXIlMjIlMkMlMjBpcF9hZGFwdGVyX2Jsb2NrJTJDJTIwMCk=",highlighted:'dd_blocks.sub_blocks.insert(<span class="hljs-string">"ip_adapter"</span>, ip_adapter_block, <span class="hljs-number">0</span>)',lang:"py",wrap:!1});var h=l(f,4);e(h,{code:"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",highlighted:`dd_pipeline = dd_blocks.init_pipeline(<span class="hljs-string">"YiYiXu/modular-demo-auto"</span>, collection=<span class="hljs-string">"diffdiff"</span>)
	dd_pipeline.load_components(dtype=torch.float16)
	dd_pipeline.loader.load_ip_adapter(<span class="hljs-string">"h94/IP-Adapter"</span>, subfolder=<span class="hljs-string">"sdxl_models"</span>, weight_name=<span class="hljs-string">"ip-adapter_sdxl.bin"</span>)
	dd_pipeline.loader.set_ip_adapter_scale(<span class="hljs-number">0.6</span>)
	dd_pipeline = dd_pipeline.to(device)

	ip_adapter_image = load_image(<span class="hljs-string">"https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/diffdiff_orange.jpeg"</span>)
	image = load_image(<span class="hljs-string">"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/20240329211129_4024911930.png?download=true"</span>)
	mask = load_image(<span class="hljs-string">"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/gradient_mask.png?download=true"</span>)

	prompt = <span class="hljs-string">"a green pear"</span>
	negative_prompt = <span class="hljs-string">"blurry"</span>
	generator = torch.Generator(device=device).manual_seed(<span class="hljs-number">42</span>)

	image = dd_pipeline(
	prompt=prompt,
	negative_prompt=negative_prompt,
	num_inference_steps=<span class="hljs-number">25</span>,
	generator=generator,
	ip_adapter_image=ip_adapter_image,
	diffdiff_map=mask,
	image=image,

	output=<span class="hljs-string">"images"</span>
	)[<span class="hljs-number">0</span>]`,lang:"py",wrap:!1});var j=l(h,2);s(j,{title:"ControlNet",local:"controlnet",headingTag:"h3"});var G=l(j,4);e(G,{code:"ZnJvbSUyMGRpZmZ1c2Vycy5tb2R1bGFyX3BpcGVsaW5lcy5zdGFibGVfZGlmZnVzaW9uX3hsLm1vZHVsYXJfYmxvY2tzJTIwaW1wb3J0JTIwU3RhYmxlRGlmZnVzaW9uWExBdXRvQ29udHJvbE5ldElucHV0U3RlcCUwQSUwQWNvbnRyb2xfaW5wdXRfYmxvY2slMjAlM0QlMjBTdGFibGVEaWZmdXNpb25YTEF1dG9Db250cm9sTmV0SW5wdXRTdGVwKCk=",highlighted:`<span class="hljs-keyword">from</span> diffusers.modular_pipelines.stable_diffusion_xl.modular_blocks <span class="hljs-keyword">import</span> StableDiffusionXLAutoControlNetInputStep

	control_input_block = StableDiffusionXLAutoControlNetInputStep()`,lang:"py",wrap:!1});var R=l(G,6);e(R,{code:"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",highlighted:`<span class="hljs-keyword">class</span> <span class="hljs-title class_">SDXLDiffDiffControlNetDenoiseStep</span>(<span class="hljs-title class_ inherited__">StableDiffusionXLDenoiseLoopWrapper</span>):
	block_classes = [SDXLDiffDiffLoopBeforeDenoiser, StableDiffusionXLControlNetLoopDenoiser, StableDiffusionXLDenoiseLoopAfterDenoiser]
	block_names = [<span class="hljs-string">"before_denoiser"</span>, <span class="hljs-string">"denoiser"</span>, <span class="hljs-string">"after_denoiser"</span>]

	controlnet_denoise_block = SDXLDiffDiffControlNetDenoiseStep()`,lang:"py",wrap:!1});var C=l(R,4);e(C,{code:"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",highlighted:`dd_blocks.sub_blocks.insert(<span class="hljs-string">"controlnet_input"</span>, control_input_block, <span class="hljs-number">7</span>)
	dd_blocks.sub_blocks[<span class="hljs-string">"denoise"</span>] = controlnet_denoise_block

	dd_pipeline = dd_blocks.init_pipeline(<span class="hljs-string">"YiYiXu/modular-demo-auto"</span>, collection=<span class="hljs-string">"diffdiff"</span>)
	dd_pipeline.load_components(dtype=torch.float16)
	dd_pipeline = dd_pipeline.to(device)

	control_image = load_image(<span class="hljs-string">"https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/diffdiff_tomato_canny.jpeg"</span>)
	image = load_image(<span class="hljs-string">"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/20240329211129_4024911930.png?download=true"</span>)
	mask = load_image(<span class="hljs-string">"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/gradient_mask.png?download=true"</span>)

	prompt = <span class="hljs-string">"a green pear"</span>
	negative_prompt = <span class="hljs-string">"blurry"</span>
	generator = torch.Generator(device=device).manual_seed(<span class="hljs-number">42</span>)

	image = dd_pipeline(
	prompt=prompt,
	negative_prompt=negative_prompt,
	num_inference_steps=<span class="hljs-number">25</span>,
	generator=generator,
	control_image=control_image,
	controlnet_conditioning_scale=<span class="hljs-number">0.5</span>,
	diffdiff_map=mask,
	image=image,
	output=<span class="hljs-string">"images"</span>
	)[<span class="hljs-number">0</span>]`,lang:"py",wrap:!1});var W=l(C,2);s(W,{title:"AutoPipelineBlocks",local:"autopipelineblocks",headingTag:"h3"});var X=l(W,6);e(X,{code:"Y2xhc3MlMjBTRFhMRGlmZkRpZmZBdXRvRGVub2lzZVN0ZXAoQXV0b1BpcGVsaW5lQmxvY2tzKSUzQSUwQSUyMCUyMCUyMCUyMGJsb2NrX2NsYXNzZXMlMjAlM0QlMjAlNUJTRFhMRGlmZkRpZmZDb250cm9sTmV0RGVub2lzZVN0ZXAlMkMlMjBTRFhMRGlmZkRpZmZEZW5vaXNlU3RlcCU1RCUwQSUyMCUyMCUyMCUyMGJsb2NrX25hbWVzJTIwJTNEJTIwJTVCJTIyY29udHIlMEFvbG5ldF9kZW5vaXNlJTIyJTJDJTIwJTIyZGVub2lzZSUyMiU1RCUwQWJsb2NrX3RyaWdnZXJfaW5wdXRzJTIwJTNEJTIwJTVCJTIyY29udHJvbG5ldF9jb25kJTIyJTJDJTIwTm9uZSU1RA==",highlighted:`<span class="hljs-keyword">class</span> <span class="hljs-title class_">SDXLDiffDiffAutoDenoiseStep</span>(<span class="hljs-title class_ inherited__">AutoPipelineBlocks</span>):
	block_classes = [SDXLDiffDiffControlNetDenoiseStep, SDXLDiffDiffDenoiseStep]
	block_names = [<span class="hljs-string">"contr
	olnet_denoise"</span>, <span class="hljs-string">"denoise"</span>]
	block_trigger_inputs = [<span class="hljs-string">"controlnet_cond"</span>, <span class="hljs-literal">None</span>]`,lang:"py",wrap:!1});var _=l(X,4);e(_,{code:"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",highlighted:`DIFFDIFF_AUTO_BLOCKS = IMAGE2IMAGE_BLOCKS.copy()
	DIFFDIFF_AUTO_BLOCKS[<span class="hljs-string">"prepare_latents"</span>] = SDXLDiffDiffPrepareLatentsStep
	DIFFDIFF_AUTO_BLOCKS[<span class="hljs-string">"set_timesteps"</span>] = TEXT2IMAGE_BLOCKS[<span class="hljs-string">"set_timesteps"</span>]
	DIFFDIFF_AUTO_BLOCKS[<span class="hljs-string">"denoise"</span>] = SDXLDiffDiffAutoDenoiseStep
	DIFFDIFF_AUTO_BLOCKS.insert(<span class="hljs-string">"ip_adapter"</span>, StableDiffusionXLAutoIPAdapterStep, <span class="hljs-number">0</span>)
	DIFFDIFF_AUTO_BLOCKS.insert(<span class="hljs-string">"controlnet_input"</span>,StableDiffusionXLControlNetAutoInput, <span class="hljs-number">7</span>)`,lang:"py",wrap:!1});var B=l(_,4);e(B,{code:"ZGRfYXV0b19ibG9ja3MlMjAlM0QlMjBTZXF1ZW50aWFsUGlwZWxpbmVCbG9ja3MuZnJvbV9ibG9ja3NfZGljdChESUZGRElGRl9BVVRPX0JMT0NLUyklMEFkZF9waXBlbGluZSUyMCUzRCUyMGRkX2F1dG9fYmxvY2tzLmluaXRfcGlwZWxpbmUoJTIyWWlZaVh1JTJGbW9kdWxhci1kZW1vLWF1dG8lMjIlMkMlMjBjb2xsZWN0aW9uJTNEJTIyZGlmZmRpZmYlMjIpJTBBZGRfcGlwZWxpbmUubG9hZF9jb21wb25lbnRzKGR0eXBlJTNEdG9yY2guZmxvYXQxNik=",highlighted:`dd_auto_blocks = SequentialPipelineBlocks.from_blocks_dict(DIFFDIFF_AUTO_BLOCKS)
	dd_pipeline = dd_auto_blocks.init_pipeline(<span class="hljs-string">"YiYiXu/modular-demo-auto"</span>, collection=<span class="hljs-string">"diffdiff"</span>)
	dd_pipeline.load_components(dtype=torch.float16)`,lang:"py",wrap:!1});var k=l(B,2);s(k,{title:"分享",local:"分享",headingTag:"h2"});var V=l(k,4);e(V,{code:"ZGRfcGlwZWxpbmUuc2F2ZV9wcmV0cmFpbmVkKCUyMllpWWlYdSUyRnRlc3RfbW9kdWxhcl9kb2MlMjIlMkMlMjBwdXNoX3RvX2h1YiUzRFRydWUp",highlighted:'dd_pipeline.save_pretrained(<span class="hljs-string">"YiYiXu/test_modular_doc"</span>, push_to_hub=<span class="hljs-literal">True</span>)',lang:"py",wrap:!1});var g=l(V,4);e(g,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch
	<span class="hljs-keyword">from</span> diffusers.modular_pipelines <span class="hljs-keyword">import</span> ModularPipeline, ComponentsManager

	components = ComponentsManager()

	diffdiff_pipeline = ModularPipeline.from_pretrained(<span class="hljs-string">"YiYiXu/modular-diffdiff-0704"</span>, trust_remote_code=<span class="hljs-literal">True</span>, components_manager=components, collection=<span class="hljs-string">"diffdiff"</span>)
	diffdiff_pipeline.load_components(dtype=torch.float16)`,lang:"py",wrap:!1});var D=l(g,2);q(D,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/zh/modular_diffusers/quickstart.md"}),O(2),Y(Q,a),K()}export{tl as component};

Xet Storage Details

Size:: 41 kB
Xet hash:: 69f4f1f5804393da073f93507ccd839a8dcd24e98b2bcf5f288142bedb10addf

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.