Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"파이프라인, 모델 및 스케줄러 이해하기","local":"파이프라인-모델-및-스케줄러-이해하기","sections":[{"title":"기본 파이프라인 해체하기","local":"기본-파이프라인-해체하기","sections":[],"depth":2},{"title":"Stable Diffusion 파이프라인 해체하기","local":"stable-diffusion-파이프라인-해체하기","sections":[{"title":"텍스트 임베딩 생성하기","local":"텍스트-임베딩-생성하기","sections":[],"depth":3},{"title":"랜덤 노이즈 생성","local":"랜덤-노이즈-생성","sections":[],"depth":3},{"title":"이미지 노이즈 제거","local":"이미지-노이즈-제거","sections":[],"depth":3},{"title":"이미지 디코딩","local":"이미지-디코딩","sections":[],"depth":3}],"depth":2},{"title":"다음 단계","local":"다음-단계","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/diffusers/main/ko/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/entry/start.df0b2e2f.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/scheduler.94020406.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/singletons.4482ca3c.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/index.8b553f6b.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/paths.1334326c.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/entry/app.4378a265.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/index.a08c8d92.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/nodes/0.96378992.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/nodes/55.7ae382e7.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/Tip.3b0aeee8.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/CodeBlock.b23cf525.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/DocNotebookDropdown.d8a25975.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/globals.7f7f1b26.js"> | |
| <link rel="modulepreload" href="/docs/diffusers/main/ko/_app/immutable/chunks/EditOnGithub.b1bceb47.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"파이프라인, 모델 및 스케줄러 이해하기","local":"파이프라인-모델-및-스케줄러-이해하기","sections":[{"title":"기본 파이프라인 해체하기","local":"기본-파이프라인-해체하기","sections":[],"depth":2},{"title":"Stable Diffusion 파이프라인 해체하기","local":"stable-diffusion-파이프라인-해체하기","sections":[{"title":"텍스트 임베딩 생성하기","local":"텍스트-임베딩-생성하기","sections":[],"depth":3},{"title":"랜덤 노이즈 생성","local":"랜덤-노이즈-생성","sections":[],"depth":3},{"title":"이미지 노이즈 제거","local":"이미지-노이즈-제거","sections":[],"depth":3},{"title":"이미지 디코딩","local":"이미지-디코딩","sections":[],"depth":3}],"depth":2},{"title":"다음 단계","local":"다음-단계","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="파이프라인-모델-및-스케줄러-이해하기" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#파이프라인-모델-및-스케줄러-이해하기"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>파이프라인, 모델 및 스케줄러 이해하기</span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"> </button> </div> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"> </button> </div></div> <p data-svelte-h="svelte-1w08vjm">🧨 Diffusers는 사용자 친화적이며 유연한 도구 상자로, 사용사례에 맞게 diffusion 시스템을 구축 할 수 있도록 설계되었습니다. 이 도구 상자의 핵심은 모델과 스케줄러입니다. <code>DiffusionPipeline</code>은 편의를 위해 이러한 구성 요소를 번들로 제공하지만, 파이프라인을 분리하고 모델과 스케줄러를 개별적으로 사용해 새로운 diffusion 시스템을 만들 수도 있습니다.</p> <p data-svelte-h="svelte-1rh2wvk">이 튜토리얼에서는 기본 파이프라인부터 시작해 Stable Diffusion 파이프라인까지 진행하며 모델과 스케줄러를 사용해 추론을 위한 diffusion 시스템을 조립하는 방법을 배웁니다.</p> <h2 class="relative group"><a id="기본-파이프라인-해체하기" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#기본-파이프라인-해체하기"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>기본 파이프라인 해체하기</span></h2> <p data-svelte-h="svelte-1ejhv0l">파이프라인은 추론을 위해 모델을 실행하는 빠르고 쉬운 방법으로, 이미지를 생성하는 데 코드가 4줄 이상 필요하지 않습니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDPMPipeline | |
| <span class="hljs-meta">>>> </span>ddpm = DDPMPipeline.from_pretrained(<span class="hljs-string">"google/ddpm-cat-256"</span>).to(<span class="hljs-string">"cuda"</span>) | |
| <span class="hljs-meta">>>> </span>image = ddpm(num_inference_steps=<span class="hljs-number">25</span>).images[<span class="hljs-number">0</span>] | |
| <span class="hljs-meta">>>> </span>image<!-- HTML_TAG_END --></pre></div> <div class="flex justify-center" data-svelte-h="svelte-ej6f4c"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ddpm-cat.png" alt="Image of cat created from DDPMPipeline"></div> <p data-svelte-h="svelte-1l346u9">정말 쉽습니다. 그런데 파이프라인은 어떻게 이렇게 할 수 있었을까요? 파이프라인을 세분화하여 내부에서 어떤 일이 일어나고 있는지 살펴보겠습니다.</p> <p data-svelte-h="svelte-abj3gw">위 예시에서 파이프라인에는 <code>UNet2DModel</code> 모델과 <code>DDPMScheduler</code>가 포함되어 있습니다. 파이프라인은 원하는 출력 크기의 랜덤 노이즈를 받아 모델을 여러번 통과시켜 이미지의 노이즈를 제거합니다. 각 timestep에서 모델은 <em>noise residual</em>을 예측하고 스케줄러는 이를 사용하여 노이즈가 적은 이미지를 예측합니다. 파이프라인은 지정된 추론 스텝수에 도달할 때까지 이 과정을 반복합니다.</p> <p data-svelte-h="svelte-dgd067">모델과 스케줄러를 별도로 사용하여 파이프라인을 다시 생성하기 위해 자체적인 노이즈 제거 프로세스를 작성해 보겠습니다.</p> <ol><li><p data-svelte-h="svelte-j1r62h">모델과 스케줄러를 불러옵니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DDPMScheduler, UNet2DModel | |
| <span class="hljs-meta">>>> </span>scheduler = DDPMScheduler.from_pretrained(<span class="hljs-string">"google/ddpm-cat-256"</span>) | |
| <span class="hljs-meta">>>> </span>model = UNet2DModel.from_pretrained(<span class="hljs-string">"google/ddpm-cat-256"</span>).to(<span class="hljs-string">"cuda"</span>)<!-- HTML_TAG_END --></pre></div></li> <li><p data-svelte-h="svelte-khbsw8">노이즈 제거 프로세스를 실행할 timestep 수를 설정합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>scheduler.set_timesteps(<span class="hljs-number">50</span>)<!-- HTML_TAG_END --></pre></div></li> <li><p data-svelte-h="svelte-yj8dtl">스케줄러의 timestep을 설정하면 균등한 간격의 구성 요소를 가진 텐서가 생성됩니다.(이 예시에서는 50개) 각 요소는 모델이 이미지의 노이즈를 제거하는 시간 간격에 해당합니다. 나중에 노이즈 제거 루프를 만들 때 이 텐서를 반복하여 이미지의 노이즈를 제거합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>scheduler.timesteps | |
| tensor([<span class="hljs-number">980</span>, <span class="hljs-number">960</span>, <span class="hljs-number">940</span>, <span class="hljs-number">920</span>, <span class="hljs-number">900</span>, <span class="hljs-number">880</span>, <span class="hljs-number">860</span>, <span class="hljs-number">840</span>, <span class="hljs-number">820</span>, <span class="hljs-number">800</span>, <span class="hljs-number">780</span>, <span class="hljs-number">760</span>, <span class="hljs-number">740</span>, <span class="hljs-number">720</span>, | |
| <span class="hljs-number">700</span>, <span class="hljs-number">680</span>, <span class="hljs-number">660</span>, <span class="hljs-number">640</span>, <span class="hljs-number">620</span>, <span class="hljs-number">600</span>, <span class="hljs-number">580</span>, <span class="hljs-number">560</span>, <span class="hljs-number">540</span>, <span class="hljs-number">520</span>, <span class="hljs-number">500</span>, <span class="hljs-number">480</span>, <span class="hljs-number">460</span>, <span class="hljs-number">440</span>, | |
| <span class="hljs-number">420</span>, <span class="hljs-number">400</span>, <span class="hljs-number">380</span>, <span class="hljs-number">360</span>, <span class="hljs-number">340</span>, <span class="hljs-number">320</span>, <span class="hljs-number">300</span>, <span class="hljs-number">280</span>, <span class="hljs-number">260</span>, <span class="hljs-number">240</span>, <span class="hljs-number">220</span>, <span class="hljs-number">200</span>, <span class="hljs-number">180</span>, <span class="hljs-number">160</span>, | |
| <span class="hljs-number">140</span>, <span class="hljs-number">120</span>, <span class="hljs-number">100</span>, <span class="hljs-number">80</span>, <span class="hljs-number">60</span>, <span class="hljs-number">40</span>, <span class="hljs-number">20</span>, <span class="hljs-number">0</span>])<!-- HTML_TAG_END --></pre></div></li> <li><p data-svelte-h="svelte-1mx5t16">원하는 출력과 같은 모양을 가진 랜덤 노이즈를 생성합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-meta">>>> </span>sample_size = model.config.sample_size | |
| <span class="hljs-meta">>>> </span>noise = torch.randn((<span class="hljs-number">1</span>, <span class="hljs-number">3</span>, sample_size, sample_size), device=<span class="hljs-string">"cuda"</span>)<!-- HTML_TAG_END --></pre></div></li> <li><p data-svelte-h="svelte-1dcn1b9">이제 timestep을 반복하는 루프를 작성합니다. 각 timestep에서 모델은 <code>UNet2DModel.forward()</code>를 통해 noisy residual을 반환합니다. 스케줄러의 <code>step()</code> 메서드는 noisy residual, timestep, 그리고 입력을 받아 이전 timestep에서 이미지를 예측합니다. 이 출력은 노이즈 제거 루프의 모델에 대한 다음 입력이 되며, <code>timesteps</code> 배열의 끝에 도달할 때까지 반복됩니다.</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-built_in">input</span> = noise | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">for</span> t <span class="hljs-keyword">in</span> scheduler.timesteps: | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">with</span> torch.no_grad(): | |
| <span class="hljs-meta">... </span> noisy_residual = model(<span class="hljs-built_in">input</span>, t).sample | |
| <span class="hljs-meta">... </span> previous_noisy_sample = scheduler.step(noisy_residual, t, <span class="hljs-built_in">input</span>).prev_sample | |
| <span class="hljs-meta">... </span> <span class="hljs-built_in">input</span> = previous_noisy_sample<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-x05lv9">이것이 전체 노이즈 제거 프로세스이며, 동일한 패턴을 사용해 모든 diffusion 시스템을 작성할 수 있습니다.</p></li> <li><p data-svelte-h="svelte-1plitwm">마지막 단계는 노이즈가 제거된 출력을 이미지로 변환하는 것입니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np | |
| <span class="hljs-meta">>>> </span>image = (<span class="hljs-built_in">input</span> / <span class="hljs-number">2</span> + <span class="hljs-number">0.5</span>).clamp(<span class="hljs-number">0</span>, <span class="hljs-number">1</span>) | |
| <span class="hljs-meta">>>> </span>image = image.cpu().permute(<span class="hljs-number">0</span>, <span class="hljs-number">2</span>, <span class="hljs-number">3</span>, <span class="hljs-number">1</span>).numpy()[<span class="hljs-number">0</span>] | |
| <span class="hljs-meta">>>> </span>image = Image.fromarray((image * <span class="hljs-number">255</span>).<span class="hljs-built_in">round</span>().astype(<span class="hljs-string">"uint8"</span>)) | |
| <span class="hljs-meta">>>> </span>image<!-- HTML_TAG_END --></pre></div></li></ol> <p data-svelte-h="svelte-1lzspna">다음 섹션에서는 여러분의 기술을 시험해보고 좀 더 복잡한 Stable Diffusion 파이프라인을 분석해 보겠습니다. 방법은 거의 동일합니다. 필요한 구성요소들을 초기화하고 timestep수를 설정하여 <code>timestep</code> 배열을 생성합니다. 노이즈 제거 루프에서 <code>timestep</code> 배열이 사용되며, 이 배열의 각 요소에 대해 모델은 노이즈가 적은 이미지를 예측합니다. 노이즈 제거 루프는 <code>timestep</code>을 반복하고 각 timestep에서 noise residual을 출력하고 스케줄러는 이를 사용하여 이전 timestep에서 노이즈가 덜한 이미지를 예측합니다. 이 프로세스는 <code>timestep</code> 배열의 끝에 도달할 때까지 반복됩니다.</p> <p data-svelte-h="svelte-5z3iwd">한번 사용해 봅시다!</p> <h2 class="relative group"><a id="stable-diffusion-파이프라인-해체하기" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#stable-diffusion-파이프라인-해체하기"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Stable Diffusion 파이프라인 해체하기</span></h2> <p data-svelte-h="svelte-cnatu9">Stable Diffusion 은 text-to-image <em>latent diffusion</em> 모델입니다. latent diffusion 모델이라고 불리는 이유는 실제 픽셀 공간 대신 이미지의 저차원의 표현으로 작업하기 때문이고, 메모리 효율이 더 높습니다. 인코더는 이미지를 더 작은 표현으로 압축하고, 디코더는 압축된 표현을 다시 이미지로 변환합니다. text-to-image 모델의 경우 텍스트 임베딩을 생성하기 위해 tokenizer와 인코더가 필요합니다. 이전 예제에서 이미 UNet 모델과 스케줄러가 필요하다는 것은 알고 계셨을 것입니다.</p> <p data-svelte-h="svelte-1qnf9z2">보시다시피, 이것은 UNet 모델만 포함된 DDPM 파이프라인보다 더 복잡합니다. Stable Diffusion 모델에는 세 개의 개별 사전학습된 모델이 있습니다.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-ra0af6">💡 VAE, UNet 및 텍스트 인코더 모델의 작동방식에 대한 자세한 내용은 <a href="https://huggingface.co/blog/stable_diffusion#how-does-stable-diffusion-work" rel="nofollow">How does Stable Diffusion work?</a> 블로그를 참조하세요.</p></div> <p data-svelte-h="svelte-1v6hzf8">이제 Stable Diffusion 파이프라인에 필요한 구성요소들이 무엇인지 알았으니, <code>from_pretrained()</code> 메서드를 사용해 모든 구성요소를 불러옵니다. 사전학습된 체크포인트 <a href="https://huggingface.co/runwayml/stable-diffusion-v1-5" rel="nofollow"><code>runwayml/stable-diffusion-v1-5</code></a>에서 찾을 수 있으며, 각 구성요소들은 별도의 하위 폴더에 저장되어 있습니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> CLIPTextModel, CLIPTokenizer | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoencoderKL, UNet2DConditionModel, PNDMScheduler | |
| <span class="hljs-meta">>>> </span>vae = AutoencoderKL.from_pretrained(<span class="hljs-string">"CompVis/stable-diffusion-v1-4"</span>, subfolder=<span class="hljs-string">"vae"</span>) | |
| <span class="hljs-meta">>>> </span>tokenizer = CLIPTokenizer.from_pretrained(<span class="hljs-string">"CompVis/stable-diffusion-v1-4"</span>, subfolder=<span class="hljs-string">"tokenizer"</span>) | |
| <span class="hljs-meta">>>> </span>text_encoder = CLIPTextModel.from_pretrained(<span class="hljs-string">"CompVis/stable-diffusion-v1-4"</span>, subfolder=<span class="hljs-string">"text_encoder"</span>) | |
| <span class="hljs-meta">>>> </span>unet = UNet2DConditionModel.from_pretrained(<span class="hljs-string">"CompVis/stable-diffusion-v1-4"</span>, subfolder=<span class="hljs-string">"unet"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-angdpe">기본 <code>PNDMScheduler</code> 대신, <code>UniPCMultistepScheduler</code>로 교체하여 다른 스케줄러를 얼마나 쉽게 연결할 수 있는지 확인합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-meta">>>> </span>scheduler = UniPCMultistepScheduler.from_pretrained(<span class="hljs-string">"CompVis/stable-diffusion-v1-4"</span>, subfolder=<span class="hljs-string">"scheduler"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-xe2wp7">추론 속도를 높이려면 스케줄러와 달리 학습 가능한 가중치가 있으므로 모델을 GPU로 옮기세요:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>torch_device = <span class="hljs-string">"cuda"</span> | |
| <span class="hljs-meta">>>> </span>vae.to(torch_device) | |
| <span class="hljs-meta">>>> </span>text_encoder.to(torch_device) | |
| <span class="hljs-meta">>>> </span>unet.to(torch_device)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="텍스트-임베딩-생성하기" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#텍스트-임베딩-생성하기"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>텍스트 임베딩 생성하기</span></h3> <p data-svelte-h="svelte-nswq27">다음 단계는 임베딩을 생성하기 위해 텍스트를 토큰화하는 것입니다. 이 텍스트는 UNet 모델에서 condition으로 사용되고 입력 프롬프트와 유사한 방향으로 diffusion 프로세스를 조정하는 데 사용됩니다.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-g2ye81">💡 <code>guidance_scale</code> 매개변수는 이미지를 생성할 때 프롬프트에 얼마나 많은 가중치를 부여할지 결정합니다.</p></div> <p data-svelte-h="svelte-193i9if">다른 프롬프트를 생성하고 싶다면 원하는 프롬프트를 자유롭게 선택하세요!</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>prompt = [<span class="hljs-string">"a photograph of an astronaut riding a horse"</span>] | |
| <span class="hljs-meta">>>> </span>height = <span class="hljs-number">512</span> <span class="hljs-comment"># Stable Diffusion의 기본 높이</span> | |
| <span class="hljs-meta">>>> </span>width = <span class="hljs-number">512</span> <span class="hljs-comment"># Stable Diffusion의 기본 너비</span> | |
| <span class="hljs-meta">>>> </span>num_inference_steps = <span class="hljs-number">25</span> <span class="hljs-comment"># 노이즈 제거 스텝 수</span> | |
| <span class="hljs-meta">>>> </span>guidance_scale = <span class="hljs-number">7.5</span> <span class="hljs-comment"># classifier-free guidance를 위한 scale</span> | |
| <span class="hljs-meta">>>> </span>generator = torch.manual_seed(<span class="hljs-number">0</span>) <span class="hljs-comment"># 초기 잠재 노이즈를 생성하는 seed generator</span> | |
| <span class="hljs-meta">>>> </span>batch_size = <span class="hljs-built_in">len</span>(prompt)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-v5k7o5">텍스트를 토큰화하고 프롬프트에서 임베딩을 생성합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>text_input = tokenizer( | |
| <span class="hljs-meta">... </span> prompt, padding=<span class="hljs-string">"max_length"</span>, max_length=tokenizer.model_max_length, truncation=<span class="hljs-literal">True</span>, return_tensors=<span class="hljs-string">"pt"</span> | |
| <span class="hljs-meta">... </span>) | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">with</span> torch.no_grad(): | |
| <span class="hljs-meta">... </span> text_embeddings = text_encoder(text_input.input_ids.to(torch_device))[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1nerc8v">또한 패딩 토큰의 임베딩인 <em>unconditional 텍스트 임베딩</em>을 생성해야 합니다. 이 임베딩은 조건부 <code>text_embeddings</code>과 동일한 shape(<code>batch_size</code> 그리고 <code>seq_length</code>)을 가져야 합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>max_length = text_input.input_ids.shape[-<span class="hljs-number">1</span>] | |
| <span class="hljs-meta">>>> </span>uncond_input = tokenizer([<span class="hljs-string">""</span>] * batch_size, padding=<span class="hljs-string">"max_length"</span>, max_length=max_length, return_tensors=<span class="hljs-string">"pt"</span>) | |
| <span class="hljs-meta">>>> </span>uncond_embeddings = text_encoder(uncond_input.input_ids.to(torch_device))[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-iqpjwl">두번의 forward pass를 피하기 위해 conditional 임베딩과 unconditional 임베딩을 배치(batch)로 연결하겠습니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>text_embeddings = torch.cat([uncond_embeddings, text_embeddings])<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="랜덤-노이즈-생성" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#랜덤-노이즈-생성"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>랜덤 노이즈 생성</span></h3> <p data-svelte-h="svelte-b5xa4e">그다음 diffusion 프로세스의 시작점으로 초기 랜덤 노이즈를 생성합니다. 이것이 이미지의 잠재적 표현이며 점차적으로 노이즈가 제거됩니다. 이 시점에서 <code>latent</code> 이미지는 최종 이미지 크기보다 작지만 나중에 모델이 이를 512x512 이미지 크기로 변환하므로 괜찮습니다.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-132csr5">💡 <code>vae</code> 모델에는 3개의 다운 샘플링 레이어가 있기 때문에 높이와 너비가 8로 나뉩니다. 다음을 실행하여 확인할 수 있습니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-number">2</span> ** (<span class="hljs-built_in">len</span>(vae.config.block_out_channels) - <span class="hljs-number">1</span>) == <span class="hljs-number">8</span><!-- HTML_TAG_END --></pre></div></div> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>latents = torch.randn( | |
| <span class="hljs-meta">... </span> (batch_size, unet.config.in_channels, height // <span class="hljs-number">8</span>, width // <span class="hljs-number">8</span>), | |
| <span class="hljs-meta">... </span> generator=generator, | |
| <span class="hljs-meta">... </span> device=torch_device, | |
| <span class="hljs-meta">... </span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="이미지-노이즈-제거" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#이미지-노이즈-제거"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>이미지 노이즈 제거</span></h3> <p data-svelte-h="svelte-1xr7g73">먼저 <code>UniPCMultistepScheduler</code>와 같은 향상된 스케줄러에 필요한 노이즈 스케일 값인 초기 노이즈 분포 <em>sigma</em> 로 입력을 스케일링 하는 것부터 시작합니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>latents = latents * scheduler.init_noise_sigma<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-4n9fvy">마지막 단계는 <code>latent</code>의 순수한 노이즈를 점진적으로 프롬프트에 설명된 이미지로 변환하는 노이즈 제거 루프를 생성하는 것입니다. 노이즈 제거 루프는 세 가지 작업을 수행해야 한다는 점을 기억하세요:</p> <ol data-svelte-h="svelte-iivnel"><li>노이즈 제거 중에 사용할 스케줄러의 timesteps를 설정합니다.</li> <li>timestep을 따라 반복합니다.</li> <li>각 timestep에서 UNet 모델을 호출하여 noise residual을 예측하고 스케줄러에 전달하여 이전 노이즈 샘플을 계산합니다.</li></ol> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> tqdm.auto <span class="hljs-keyword">import</span> tqdm | |
| <span class="hljs-meta">>>> </span>scheduler.set_timesteps(num_inference_steps) | |
| <span class="hljs-meta">>>> </span><span class="hljs-keyword">for</span> t <span class="hljs-keyword">in</span> tqdm(scheduler.timesteps): | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># classifier-free guidance를 수행하는 경우 두번의 forward pass를 수행하지 않도록 latent를 확장.</span> | |
| <span class="hljs-meta">... </span> latent_model_input = torch.cat([latents] * <span class="hljs-number">2</span>) | |
| <span class="hljs-meta">... </span> latent_model_input = scheduler.scale_model_input(latent_model_input, timestep=t) | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># noise residual 예측</span> | |
| <span class="hljs-meta">... </span> <span class="hljs-keyword">with</span> torch.no_grad(): | |
| <span class="hljs-meta">... </span> noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># guidance 수행</span> | |
| <span class="hljs-meta">... </span> noise_pred_uncond, noise_pred_text = noise_pred.chunk(<span class="hljs-number">2</span>) | |
| <span class="hljs-meta">... </span> noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) | |
| <span class="hljs-meta">... </span> <span class="hljs-comment"># 이전 노이즈 샘플을 계산 x_t -> x_t-1</span> | |
| <span class="hljs-meta">... </span> latents = scheduler.step(noise_pred, t, latents).prev_sample<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="이미지-디코딩" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#이미지-디코딩"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>이미지 디코딩</span></h3> <p data-svelte-h="svelte-12g4hh4">마지막 단계는 <code>vae</code>를 이용하여 잠재 표현을 이미지로 디코딩하고 <code>sample</code>과 함께 디코딩된 출력을 얻는 것입니다:</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-comment"># latent를 스케일링하고 vae로 이미지 디코딩</span> | |
| latents = <span class="hljs-number">1</span> / <span class="hljs-number">0.18215</span> * latents | |
| <span class="hljs-keyword">with</span> torch.no_grad(): | |
| image = vae.decode(latents).sample<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1bjoomr">마지막으로 이미지를 <code>PIL.Image</code>로 변환하면 생성된 이미지를 확인할 수 있습니다!</p> <div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>image = (image / <span class="hljs-number">2</span> + <span class="hljs-number">0.5</span>).clamp(<span class="hljs-number">0</span>, <span class="hljs-number">1</span>) | |
| <span class="hljs-meta">>>> </span>image = image.detach().cpu().permute(<span class="hljs-number">0</span>, <span class="hljs-number">2</span>, <span class="hljs-number">3</span>, <span class="hljs-number">1</span>).numpy() | |
| <span class="hljs-meta">>>> </span>images = (image * <span class="hljs-number">255</span>).<span class="hljs-built_in">round</span>().astype(<span class="hljs-string">"uint8"</span>) | |
| <span class="hljs-meta">>>> </span>pil_images = [Image.fromarray(image) <span class="hljs-keyword">for</span> image <span class="hljs-keyword">in</span> images] | |
| <span class="hljs-meta">>>> </span>pil_images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div> <div class="flex justify-center" data-svelte-h="svelte-1b0w6va"><img src="https://huggingface.co/blog/assets/98_stable_diffusion/stable_diffusion_k_lms.png"></div> <h2 class="relative group"><a id="다음-단계" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#다음-단계"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>다음 단계</span></h2> <p data-svelte-h="svelte-nwx4ql">기본 파이프라인부터 복잡한 파이프라인까지, 자신만의 diffusion 시스템을 작성하는 데 필요한 것은 노이즈 제거 루프뿐이라는 것을 알 수 있었습니다. 이 루프는 스케줄러의 timesteps를 설정하고, 이를 반복하며, UNet 모델을 호출하여 noise residual을 예측하고 스케줄러에 전달하여 이전 노이즈 샘플을 계산하는 과정을 번갈아 가며 수행해야 합니다.</p> <p data-svelte-h="svelte-g0p09q">이것이 바로 🧨 Diffusers가 설계된 목적입니다: 모델과 스케줄러를 사용해 자신만의 diffusion 시스템을 직관적이고 쉽게 작성할 수 있도록 하기 위해서입니다.</p> <p data-svelte-h="svelte-qg73l2">다음 단계를 자유롭게 진행하세요:</p> <ul data-svelte-h="svelte-y3g0al"><li>🧨 Diffusers에 <a href="using-diffusers/#contribute_pipeline">파이프라인 구축 및 기여</a>하는 방법을 알아보세요. 여러분이 어떤 아이디어를 내놓을지 기대됩니다!</li> <li>라이브러리에서 <a href="./api/pipelines/overview">기본 파이프라인</a>을 살펴보고, 모델과 스케줄러를 별도로 사용하여 파이프라인을 처음부터 해체하고 빌드할 수 있는지 확인해 보세요.</li></ul> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/ko/using-diffusers/write_own_pipeline.md" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span data-svelte-h="svelte-x0xyl0">></span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p> | |
| <script> | |
| { | |
| __sveltekit_bmkzir = { | |
| assets: "/docs/diffusers/main/ko", | |
| base: "/docs/diffusers/main/ko", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/diffusers/main/ko/_app/immutable/entry/start.df0b2e2f.js"), | |
| import("/docs/diffusers/main/ko/_app/immutable/entry/app.4378a265.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 55], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 65.1 kB
- Xet hash:
- 47db11d1a18d321308328350437f3e93adb609548c3f3d4f0d85e9b586023207
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.