diffusers/pr_13881/ko/_app/immutable/nodes/51.D-LMD_Cl.js

import"../chunks/DsnmJJEf.js";import{i as w,h as V,C as U,H as s,a as e,E as T,s as v}from"../chunks/CFM6C53a.js";import{p as W,o as k,s as a,f as _,a as b,b as j,c as Z,n as C}from"../chunks/CNc7KuUZ.js";import{D as X}from"../chunks/BK2xlcGK.js";const G='{"title":"Stable Video Diffusion","local":"stable-video-diffusion","sections":[{"title":"torch.compile","local":"torchcompile","sections":[],"depth":2},{"title":"메모리 사용량 줄이기","local":"메모리-사용량-줄이기","sections":[],"depth":2},{"title":"Micro-conditioning","local":"micro-conditioning","sections":[],"depth":2}],"depth":1}';var Q=Z('<meta name="hf:doc:metadata"/>'),R=Z('<p></p> <!> <!> <!> <p><a href="https://huggingface.co/papers/2311.15127" rel="nofollow">Stable Video Diffusion (SVD)</a>은 입력 이미지에 맞춰 2~4초 분량의 고해상도(576x1024) 비디오를 생성할 수 있는 강력한 image-to-video 생성 모델입니다.</p> <p>이 가이드에서는 SVD를 사용하여 이미지에서 짧은 동영상을 생성하는 방법을 설명합니다.</p> <p>시작하기 전에 다음 라이브러리가 설치되어 있는지 확인하세요:</p> <!> <p>이 모델에는 <a href="https://huggingface.co/stabilityai/stable-video-diffusion-img2vid" rel="nofollow">SVD</a>와 <a href="https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt" rel="nofollow">SVD-XT</a> 두 가지 종류가 있습니다. SVD 체크포인트는 14개의 프레임을 생성하도록 학습되었고, SVD-XT 체크포인트는 25개의 프레임을 생성하도록 파인튜닝되었습니다.</p> <p>이 가이드에서는 SVD-XT 체크포인트를 사용합니다.</p> <!> <div class="flex gap-4"><div><img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"/> <figcaption class="mt-2 text-center text-sm text-gray-500">"source image of a rocket"</figcaption></div> <div><img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/output_rocket.gif"/> <figcaption class="mt-2 text-center text-sm text-gray-500">"generated video from source image"</figcaption></div></div> <!> <p>UNet을 <a href="../optimization/torch2.0#torchcompile">컴파일</a>하면 메모리 사용량이 살짝 증가하지만, 20~25%의 속도 향상을 얻을 수 있습니다.</p> <!> <!> <p>비디오 생성은 기본적으로 배치 크기가 큰 text-to-image 생성과 유사하게 ‘num_frames’를 한 번에 생성하기 때문에 메모리 사용량이 매우 높습니다. 메모리 사용량을 줄이기 위해 추론 속도와 메모리 사용량을 절충하는 여러 가지 옵션이 있습니다:</p> <ul><li>모델 오프로링 활성화: 파이프라인의 각 구성 요소가 더 이상 필요하지 않을 때 CPU로 오프로드됩니다.</li> <li>Feed-forward chunking 활성화: feed-forward 레이어가 배치 크기가 큰 단일 feed-forward를 실행하는 대신 루프로 반복해서 실행됩니다.</li> <li><code>decode_chunk_size</code> 감소: VAE가 프레임들을 한꺼번에 디코딩하는 대신 chunk 단위로 디코딩합니다. <code>decode_chunk_size=1</code>을 설정하면 한 번에 한 프레임씩 디코딩하고 최소한의 메모리만 사용하지만(GPU 메모리에 따라 이 값을 조정하는 것이 좋습니다), 동영상에 약간의 깜박임이 발생할 수 있습니다.</li></ul> <!> <p>이러한 모든 방법들을 사용하면 메모리 사용량이 8GAM VRAM보다 적을 것입니다.</p> <!> <p>Stable Diffusion Video는 또한 이미지 conditoning 외에도 micro-conditioning을 허용하므로 생성된 비디오를 더 잘 제어할 수 있습니다:</p> <ul><li><code>fps</code>: 생성된 비디오의 초당 프레임 수입니다.</li> <li><code>motion_bucket_id</code>: 생성된 동영상에 사용할 모션 버킷 아이디입니다. 생성된 동영상의 모션을 제어하는 데 사용할 수 있습니다. 모션 버킷 아이디를 늘리면 생성되는 동영상의 모션이 증가합니다.</li> <li><code>noise_aug_strength</code>: Conditioning 이미지에 추가되는 노이즈의 양입니다. 값이 클수록 비디오가 conditioning 이미지와 덜 유사해집니다. 이 값을 높이면 생성된 비디오의 움직임도 증가합니다.</li></ul> <p>예를 들어, 모션이 더 많은 동영상을 생성하려면 <code>motion_bucket_id</code> 및 <code>noise_aug_strength</code> micro-conditioning 파라미터를 사용합니다:</p> <!> <p><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/output_rocket_with_conditions.gif"/></p> <!> <p></p>',1);function S(J,M){W(M,!1),k(()=>{new URLSearchParams(window.location.search).get("fw")}),w();var l=R();V("1cfs5dz",u=>{var f=Q();v(f,"content",G),b(u,f)});var o=a(_(l),2);U(o,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var i=a(o,2);X(i,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;",options:[{label:"Mixed",value:"https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers_doc/ko/svd.ipynb"},{label:"PyTorch",value:"https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers_doc/ko/pytorch/svd.ipynb"},{label:"TensorFlow",value:"https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers_doc/ko/tensorflow/svd.ipynb"},{label:"Mixed",value:"https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/main/diffusers_doc/ko/svd.ipynb"},{label:"PyTorch",value:"https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/main/diffusers_doc/ko/pytorch/svd.ipynb"},{label:"TensorFlow",value:"https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/main/diffusers_doc/ko/tensorflow/svd.ipynb"}]});var n=a(i,2);s(n,{title:"Stable Video Diffusion",local:"stable-video-diffusion",headingTag:"h1"});var t=a(n,8);e(t,{code:"IXBpcCUyMGluc3RhbGwlMjAtcSUyMC1VJTIwZGlmZnVzZXJzJTIwdHJhbnNmb3JtZXJzJTIwYWNjZWxlcmF0ZQ==",highlighted:"!pip install -q -U diffusers transformers accelerate",lang:"py",wrap:!1});var d=a(t,6);e(d,{code:"aW1wb3J0JTIwdG9yY2glMEElMEFmcm9tJTIwZGlmZnVzZXJzJTIwaW1wb3J0JTIwU3RhYmxlVmlkZW9EaWZmdXNpb25QaXBlbGluZSUwQWZyb20lMjBkaWZmdXNlcnMudXRpbHMlMjBpbXBvcnQlMjBsb2FkX2ltYWdlJTJDJTIwZXhwb3J0X3RvX3ZpZGVvJTBBJTBBcGlwZSUyMCUzRCUyMFN0YWJsZVZpZGVvRGlmZnVzaW9uUGlwZWxpbmUuZnJvbV9wcmV0cmFpbmVkKCUwQSUyMCUyMCUyMCUyMCUyMnN0YWJpbGl0eWFpJTJGc3RhYmxlLXZpZGVvLWRpZmZ1c2lvbi1pbWcydmlkLXh0JTIyJTJDJTIwdG9yY2hfZHR5cGUlM0R0b3JjaC5mbG9hdDE2JTJDJTIwdmFyaWFudCUzRCUyMmZwMTYlMjIlMEEpJTBBcGlwZS5lbmFibGVfbW9kZWxfY3B1X29mZmxvYWQoKSUwQSUwQSUyMyUyMENvbmRpdGlvbmluZyUyMCVFQyU5RCVCNCVFQiVBRiVCOCVFQyVBNyU4MCUyMCVFQiVCNiU4OCVFQiU5RiVBQyVFQyU5OCVBNCVFQSVCOCVCMCUwQWltYWdlJTIwJTNEJTIwbG9hZF9pbWFnZSglMjJodHRwcyUzQSUyRiUyRmh1Z2dpbmdmYWNlLmNvJTJGZGF0YXNldHMlMkZodWdnaW5nZmFjZSUyRmRvY3VtZW50YXRpb24taW1hZ2VzJTJGcmVzb2x2ZSUyRm1haW4lMkZkaWZmdXNlcnMlMkZzdmQlMkZyb2NrZXQucG5nJTIyKSUwQWltYWdlJTIwJTNEJTIwaW1hZ2UucmVzaXplKCgxMDI0JTJDJTIwNTc2KSklMEElMEFnZW5lcmF0b3IlMjAlM0QlMjB0b3JjaC5tYW51YWxfc2VlZCg0MiklMEFmcmFtZXMlMjAlM0QlMjBwaXBlKGltYWdlJTJDJTIwZGVjb2RlX2NodW5rX3NpemUlM0Q4JTJDJTIwZ2VuZXJhdG9yJTNEZ2VuZXJhdG9yKS5mcmFtZXMlNUIwJTVEJTBBJTBBZXhwb3J0X3RvX3ZpZGVvKGZyYW1lcyUyQyUyMCUyMmdlbmVyYXRlZC5tcDQlMjIlMkMlMjBmcHMlM0Q3KQ==",highlighted:`<span class="hljs-keyword">import</span> torch

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableVideoDiffusionPipeline
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> load_image, export_to_video

pipe = StableVideoDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;stabilityai/stable-video-diffusion-img2vid-xt&quot;</span>, torch_dtype=torch.float16, variant=<span class="hljs-string">&quot;fp16&quot;</span>
)
pipe.enable_model_cpu_offload()

<span class="hljs-comment"># Conditioning 이미지 불러오기</span>
image = load_image(<span class="hljs-string">&quot;https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png&quot;</span>)
image = image.resize((<span class="hljs-number">1024</span>, <span class="hljs-number">576</span>))

generator = torch.manual_seed(<span class="hljs-number">42</span>)
frames = pipe(image, decode_chunk_size=<span class="hljs-number">8</span>, generator=generator).frames[<span class="hljs-number">0</span>]

export_to_video(frames, <span class="hljs-string">&quot;generated.mp4&quot;</span>, fps=<span class="hljs-number">7</span>)`,lang:"python",wrap:!1});var p=a(d,4);s(p,{title:"torch.compile",local:"torchcompile",headingTag:"h2"});var c=a(p,4);e(c,{code:"LSUyMHBpcGUuZW5hYmxlX21vZGVsX2NwdV9vZmZsb2FkKCklMEElMkIlMjBwaXBlLnRvKCUyMmN1ZGElMjIpJTBBJTJCJTIwcGlwZS51bmV0JTIwJTNEJTIwdG9yY2guY29tcGlsZShwaXBlLnVuZXQlMkMlMjBtb2RlJTNEJTIycmVkdWNlLW92ZXJoZWFkJTIyJTJDJTIwZnVsbGdyYXBoJTNEVHJ1ZSk=",highlighted:`<span class="hljs-deletion">- pipe.enable_model_cpu_offload()</span>
<span class="hljs-addition">+ pipe.to(&quot;cuda&quot;)</span>
<span class="hljs-addition">+ pipe.unet = torch.compile(pipe.unet, mode=&quot;reduce-overhead&quot;, fullgraph=True)</span>`,lang:"diff",wrap:!1});var m=a(c,2);s(m,{title:"메모리 사용량 줄이기",local:"메모리-사용량-줄이기",headingTag:"h2"});var r=a(m,6);e(r,{code:"LSUyMHBpcGUuZW5hYmxlX21vZGVsX2NwdV9vZmZsb2FkKCklMEEtJTIwZnJhbWVzJTIwJTNEJTIwcGlwZShpbWFnZSUyQyUyMGRlY29kZV9jaHVua19zaXplJTNEOCUyQyUyMGdlbmVyYXRvciUzRGdlbmVyYXRvcikuZnJhbWVzJTVCMCU1RCUwQSUyQiUyMHBpcGUuZW5hYmxlX21vZGVsX2NwdV9vZmZsb2FkKCklMEElMkIlMjBwaXBlLnVuZXQuZW5hYmxlX2ZvcndhcmRfY2h1bmtpbmcoKSUwQSUyQiUyMGZyYW1lcyUyMCUzRCUyMHBpcGUoaW1hZ2UlMkMlMjBkZWNvZGVfY2h1bmtfc2l6ZSUzRDIlMkMlMjBnZW5lcmF0b3IlM0RnZW5lcmF0b3IlMkMlMjBudW1fZnJhbWVzJTNEMjUpLmZyYW1lcyU1QjAlNUQ=",highlighted:`<span class="hljs-deletion">- pipe.enable_model_cpu_offload()</span>
<span class="hljs-deletion">- frames = pipe(image, decode_chunk_size=8, generator=generator).frames[0]</span>
<span class="hljs-addition">+ pipe.enable_model_cpu_offload()</span>
<span class="hljs-addition">+ pipe.unet.enable_forward_chunking()</span>
<span class="hljs-addition">+ frames = pipe(image, decode_chunk_size=2, generator=generator, num_frames=25).frames[0]</span>`,lang:"diff",wrap:!1});var h=a(r,4);s(h,{title:"Micro-conditioning",local:"micro-conditioning",headingTag:"h2"});var g=a(h,8);e(g,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableVideoDiffusionPipeline
<span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> load_image, export_to_video

pipe = StableVideoDiffusionPipeline.from_pretrained(
  <span class="hljs-string">&quot;stabilityai/stable-video-diffusion-img2vid-xt&quot;</span>, torch_dtype=torch.float16, variant=<span class="hljs-string">&quot;fp16&quot;</span>
)
pipe.enable_model_cpu_offload()

<span class="hljs-comment"># Conditioning 이미지 불러오기</span>
image = load_image(<span class="hljs-string">&quot;https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png&quot;</span>)
image = image.resize((<span class="hljs-number">1024</span>, <span class="hljs-number">576</span>))

generator = torch.manual_seed(<span class="hljs-number">42</span>)
frames = pipe(image, decode_chunk_size=<span class="hljs-number">8</span>, generator=generator, motion_bucket_id=<span class="hljs-number">180</span>, noise_aug_strength=<span class="hljs-number">0.1</span>).frames[<span class="hljs-number">0</span>]
export_to_video(frames, <span class="hljs-string">&quot;generated.mp4&quot;</span>, fps=<span class="hljs-number">7</span>)`,lang:"python",wrap:!1});var y=a(g,4);T(y,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/ko/using-diffusers/svd.md"}),C(2),b(J,l),j()}export{S as component};