diffusers/pr_13881/ko/_app/immutable/nodes/24.Cw5FWmXw.js

import"../chunks/DsnmJJEf.js";import{i as Y,h as v,C as z,H as s,a,E as S,s as Q}from"../chunks/CFM6C53a.js";import{p as A,o as k,s as l,f as L,a as V,b as x,c as E,n as H}from"../chunks/CNc7KuUZ.js";const q='{"title":"커스텀 Diffusion 학습 예제","local":"커스텀-diffusion-학습-예제","sections":[{"title":"로컬에서 PyTorch로 실행하기","local":"로컬에서-pytorch로-실행하기","sections":[{"title":"Dependencies 설치하기","local":"dependencies-설치하기","sections":[],"depth":3},{"title":"고양이 예제 😺","local":"고양이-예제-","sections":[],"depth":3},{"title":"멀티 컨셉에 대한 학습 🐱🪵","local":"멀티-컨셉에-대한-학습-","sections":[],"depth":3},{"title":"사람 얼굴에 대한 학습","local":"사람-얼굴에-대한-학습","sections":[],"depth":3}],"depth":2},{"title":"추론","local":"추론","sections":[{"title":"학습된 체크포인트에서 추론하기","local":"학습된-체크포인트에서-추론하기","sections":[],"depth":3}],"depth":2},{"title":"Grads를 None으로 설정","local":"grads를-none으로-설정","sections":[],"depth":2},{"title":"실험 결과","local":"실험-결과","sections":[],"depth":2}],"depth":1}';var D=E('<meta name="hf:doc:metadata"/>'),O=E(`<p></p> <!> <!> <p><a href="https://huggingface.co/papers/2212.04488" rel="nofollow">커스텀 Diffusion</a>은 피사체의 이미지 몇 장(4~5장)만 주어지면 Stable Diffusion처럼 text-to-image 모델을 커스터마이징하는 방법입니다.
‘train_custom_diffusion.py’ 스크립트는 학습 과정을 구현하고 이를 Stable Diffusion에 맞게 조정하는 방법을 보여줍니다.</p> <p>이 교육 사례는 <a href="https://nupurkmr9.github.io/" rel="nofollow">Nupur Kumari</a>가 제공하였습니다. (Custom Diffusion의 저자 중 한명).</p> <!> <!> <p>스크립트를 실행하기 전에 라이브러리의 학습 dependencies를 설치해야 합니다:</p> <p><strong>중요</strong></p> <p>예제 스크립트의 최신 버전을 성공적으로 실행하려면 <strong>소스로부터 설치</strong>하는 것을 매우 권장하며, 예제 스크립트를 자주 업데이트하는 만큼 일부 예제별 요구 사항을 설치하고 설치를 최신 상태로 유지하는 것이 좋습니다. 이를 위해 새 가상 환경에서 다음 단계를 실행하세요:</p> <!> <p><a href="https://github.com/huggingface/diffusers/tree/main/examples/custom_diffusion" rel="nofollow">example folder</a>로 cd하여 이동하세요.</p> <!> <p>이제 실행</p> <!> <p>그리고 <a href="https://github.com/huggingface/accelerate/" rel="nofollow">🤗Accelerate</a> 환경을 초기화:</p> <!> <p>또는 사용자 환경에 대한 질문에 답하지 않고 기본 가속 구성을 사용하려면 다음과 같이 하세요.</p> <!> <p>또는 사용 중인 환경이 대화형 셸을 지원하지 않는 경우(예: jupyter notebook)</p> <!> <!> <p>이제 데이터셋을 가져옵니다. <a href="https://www.cs.cmu.edu/~custom-diffusion/assets/data.zip" rel="nofollow">여기</a>에서 데이터셋을 다운로드하고 압축을 풉니다. 직접 데이터셋을 사용하려면 <a href="create_dataset">학습용 데이터셋 생성하기</a> 가이드를 참고하세요.</p> <p>또한 ‘clip-retrieval’을 사용하여 200개의 실제 이미지를 수집하고, regularization으로서 이를 학습 데이터셋의 타겟 이미지와 결합합니다. 이렇게 하면 주어진 타겟 이미지에 대한 과적합을 방지할 수 있습니다. 다음 플래그를 사용하면 <code>prior_loss_weight=1.</code>로 <code>prior_preservation</code>, <code>real_prior</code> regularization을 활성화할 수 있습니다.
클래스_프롬프트<code>는 대상 이미지와 동일한 카테고리 이름이어야 합니다. 수집된 실제 이미지에는</code>class_prompt<code>와 유사한 텍스트 캡션이 있습니다. 검색된 이미지는</code>class_data_dir<code>에 저장됩니다. 생성된 이미지를 regularization으로 사용하기 위해</code>real_prior\`를 비활성화할 수 있습니다. 실제 이미지를 수집하려면 훈련 전에 이 명령을 먼저 사용하십시오.</p> <!> <p><strong><strong><em>참고: <a href="https://huggingface.co/stabilityai/stable-diffusion-2" rel="nofollow">stable-diffusion-2</a> 768x768 모델을 사용하는 경우 ‘해상도’를 768로 변경하세요.</em></strong></strong></p> <p>스크립트는 모델 체크포인트와 <code>pytorch_custom_diffusion_weights.bin</code> 파일을 생성하여 저장소에 저장합니다.</p> <!> <p><strong>더 낮은 VRAM 요구 사항(GPU당 16GB)으로 더 빠르게 훈련하려면 <code>--enable_xformers_memory_efficient_attention</code>을 사용하세요. 설치 방법은 <a href="https://github.com/facebookresearch/xformers" rel="nofollow">가이드</a>를 따르세요.</strong></p> <p>가중치 및 편향(<code>wandb</code>)을 사용하여 실험을 추적하고 중간 결과를 저장하려면(강력히 권장합니다) 다음 단계를 따르세요:</p> <ul><li><code>wandb</code> 설치: <code>pip install wandb</code>.</li> <li>로그인 : <code>wandb login</code>.</li> <li>그런 다음 트레이닝을 시작하는 동안 <code>validation_prompt</code>를 지정하고 <code>report_to</code>를 <code>wandb</code>로 설정합니다. 다음과 같은 관련 인수를 구성할 수도 있습니다: <ul><li><code>num_validation_images</code></li> <li><code>validation_steps</code></li></ul></li></ul> <!> <p>다음은 <a href="https://wandb.ai/sayakpaul/custom-diffusion/runs/26ghrcau" rel="nofollow">Weights and Biases page</a>의 예시이며, 여러 학습 세부 정보와 함께 중간 결과들을 확인할 수 있습니다.</p> <p><code>--push_to_hub</code>를 지정하면 학습된 파라미터가 허깅 페이스 허브의 리포지토리에 푸시됩니다. 다음은 <a href="https://huggingface.co/sayakpaul/custom-diffusion-cat" rel="nofollow">예제 리포지토리</a>입니다.</p> <!> <p><a href="https://github.com/ShivamShrirao/diffusers/blob/main/examples/dreambooth/train_dreambooth.py" rel="nofollow">this</a>와 유사하게 각 컨셉에 대한 정보가 포함된 <a href="https://github.com/adobe-research/custom-diffusion/blob/main/assets/concept_list.json" rel="nofollow">json</a> 파일을 제공합니다.</p> <p>실제 이미지를 수집하려면 json 파일의 각 컨셉에 대해 이 명령을 실행합니다.</p> <!> <p>그럼 우리는 학습시킬 준비가 되었습니다!</p> <!> <p>다음은 <a href="https://wandb.ai/sayakpaul/custom-diffusion/runs/3990tzkg" rel="nofollow">Weights and Biases page</a>의 예시이며, 다른 학습 세부 정보와 함께 중간 결과들을 확인할 수 있습니다.</p> <!> <p>사람 얼굴에 대한 파인튜닝을 위해 다음과 같은 설정이 더 효과적이라는 것을 확인했습니다: <code>learning_rate=5e-6</code>, <code>max_train_steps=1000 to 2000</code>, <code>freeze_model=crossattn</code>을 최소 15~20개의 이미지로 설정합니다.</p> <p>실제 이미지를 수집하려면 훈련 전에 이 명령을 먼저 사용하십시오.</p> <!> <p>이제 학습을 시작하세요!</p> <!> <!> <p>위 프롬프트를 사용하여 모델을 학습시킨 후에는 아래 프롬프트를 사용하여 추론을 실행할 수 있습니다. 프롬프트에 ‘modifier token’(예: 위 예제에서는 &lt;new1&gt;)을 반드시 포함해야 합니다.</p> <!> <p>허브 리포지토리에서 이러한 매개변수를 직접 로드할 수 있습니다:</p> <!> <p>다음은 여러 컨셉으로 추론을 수행하는 예제입니다:</p> <!> <p>여기서 ‘고양이’와 ‘나무 냄비’는 여러 컨셉을 말합니다.</p> <!> <p><code>--checkpointing_steps</code> 인수를 사용한 경우 학습 과정에서 저장된 전체 체크포인트 중 하나에서 추론을 수행할 수도 있습니다.</p> <!> <p>더 많은 메모리를 절약하려면 스크립트에 <code>--set_grads_to_none</code> 인수를 전달하세요. 이렇게 하면 성적이 0이 아닌 없음으로 설정됩니다. 그러나 특정 동작이 변경되므로 문제가 발생하면 이 인수를 제거하세요.</p> <p>자세한 정보: <a href="https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html" rel="nofollow">https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html</a></p> <!> <p>실험에 대한 자세한 내용은 <a href="https://www.cs.cmu.edu/~custom-diffusion/" rel="nofollow">당사 웹페이지</a>를 참조하세요.</p> <!> <p></p>`,1);function ll(G,F){A(F,!1),k(()=>{new URLSearchParams(window.location.search).get("fw")}),Y();var t=O();v("1gvvfk7",R=>{var C=D();Q(C,"content",q),V(R,C)});var e=l(L(t),2);z(e,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var n=l(e,2);s(n,{title:"커스텀 Diffusion 학습 예제",local:"커스텀-diffusion-학습-예제",headingTag:"h1"});var o=l(n,6);s(o,{title:"로컬에서 PyTorch로 실행하기",local:"로컬에서-pytorch로-실행하기",headingTag:"h2"});var p=l(o,2);s(p,{title:"Dependencies 설치하기",local:"dependencies-설치하기",headingTag:"h3"});var i=l(p,8);a(i,{code:"Z2l0JTIwY2xvbmUlMjBodHRwcyUzQSUyRiUyRmdpdGh1Yi5jb20lMkZodWdnaW5nZmFjZSUyRmRpZmZ1c2VycyUwQWNkJTIwZGlmZnVzZXJzJTBBcGlwJTIwaW5zdGFsbCUyMC1lJTIwLg==",highlighted:`git <span class="hljs-built_in">clone</span> https://github.com/huggingface/diffusers
<span class="hljs-built_in">cd</span> diffusers
pip install -e .`,lang:"bash",wrap:!1});var c=l(i,4);a(c,{code:"Y2QlMjBleGFtcGxlcyUyRmN1c3RvbV9kaWZmdXNpb24=",highlighted:'<span class="hljs-built_in">cd</span> examples/custom_diffusion',lang:"",wrap:!1});var M=l(c,4);a(M,{code:"cGlwJTIwaW5zdGFsbCUyMC1yJTIwcmVxdWlyZW1lbnRzLnR4dCUwQXBpcCUyMGluc3RhbGwlMjBjbGlwLXJldHJpZXZhbA==",highlighted:`pip install -r requirements.txt
pip install clip-retrieval`,lang:"bash",wrap:!1});var d=l(M,4);a(d,{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZw==",highlighted:"accelerate config",lang:"bash",wrap:!1});var r=l(d,4);a(r,{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZyUyMGRlZmF1bHQ=",highlighted:"accelerate config default",lang:"bash",wrap:!1});var h=l(r,4);a(h,{code:"ZnJvbSUyMGFjY2VsZXJhdGUudXRpbHMlMjBpbXBvcnQlMjB3cml0ZV9iYXNpY19jb25maWclMEElMEF3cml0ZV9iYXNpY19jb25maWcoKQ==",highlighted:`<span class="hljs-keyword">from</span> accelerate.utils <span class="hljs-keyword">import</span> write_basic_config

write_basic_config()`,lang:"python",wrap:!1});var U=l(h,2);s(U,{title:"고양이 예제 😺",local:"고양이-예제-",headingTag:"h3"});var y=l(U,6);a(y,{code:"cGlwJTIwaW5zdGFsbCUyMGNsaXAtcmV0cmlldmFsJTBBcHl0aG9uJTIwcmV0cmlldmUucHklMjAtLWNsYXNzX3Byb21wdCUyMGNhdCUyMC0tY2xhc3NfZGF0YV9kaXIlMjByZWFsX3JlZyUyRnNhbXBsZXNfY2F0JTIwLS1udW1fY2xhc3NfaW1hZ2VzJTIwMjAw",highlighted:`pip install clip-retrieval
python retrieve.py --class_prompt <span class="hljs-built_in">cat</span> --class_data_dir real_reg/samples_cat --num_class_images 200`,lang:"bash",wrap:!1});var J=l(y,6);a(J,{code:"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",highlighted:`<span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;path-to-save-model&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;./data/cat&quot;</span>

accelerate launch train_custom_diffusion.py \\
  --pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span>  \\
  --instance_data_dir=<span class="hljs-variable">$INSTANCE_DIR</span> \\
  --output_dir=<span class="hljs-variable">$OUTPUT_DIR</span> \\
  --class_data_dir=./real_reg/samples_cat/ \\
  --with_prior_preservation --real_prior --prior_loss_weight=1.0 \\
  --class_prompt=<span class="hljs-string">&quot;cat&quot;</span> --num_class_images=200 \\
  --instance_prompt=<span class="hljs-string">&quot;photo of a &lt;new1&gt; cat&quot;</span>  \\
  --resolution=512  \\
  --train_batch_size=2  \\
  --learning_rate=1e-5  \\
  --lr_warmup_steps=0 \\
  --max_train_steps=250 \\
  --scale_lr --hflip  \\
  --modifier_token <span class="hljs-string">&quot;&lt;new1&gt;&quot;</span> \\
  --push_to_hub`,lang:"bash",wrap:!1});var m=l(J,8);a(m,{code:"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",highlighted:`accelerate launch train_custom_diffusion.py \\
  --pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span>  \\
  --instance_data_dir=<span class="hljs-variable">$INSTANCE_DIR</span> \\
  --output_dir=<span class="hljs-variable">$OUTPUT_DIR</span> \\
  --class_data_dir=./real_reg/samples_cat/ \\
  --with_prior_preservation --real_prior --prior_loss_weight=1.0 \\
  --class_prompt=<span class="hljs-string">&quot;cat&quot;</span> --num_class_images=200 \\
  --instance_prompt=<span class="hljs-string">&quot;photo of a &lt;new1&gt; cat&quot;</span>  \\
  --resolution=512  \\
  --train_batch_size=2  \\
  --learning_rate=1e-5  \\
  --lr_warmup_steps=0 \\
  --max_train_steps=250 \\
  --scale_lr --hflip  \\
  --modifier_token <span class="hljs-string">&quot;&lt;new1&gt;&quot;</span> \\
  --validation_prompt=<span class="hljs-string">&quot;&lt;new1&gt; cat sitting in a bucket&quot;</span> \\
  --report_to=<span class="hljs-string">&quot;wandb&quot;</span> \\
  --push_to_hub`,lang:"bash",wrap:!1});var u=l(m,6);s(u,{title:"멀티 컨셉에 대한 학습 🐱🪵",local:"멀티-컨셉에-대한-학습-",headingTag:"h3"});var w=l(u,6);a(w,{code:"cGlwJTIwaW5zdGFsbCUyMGNsaXAtcmV0cmlldmFsJTBBcHl0aG9uJTIwcmV0cmlldmUucHklMjAtLWNsYXNzX3Byb21wdCUyMCU3QiU3RCUyMC0tY2xhc3NfZGF0YV9kaXIlMjAlN0IlN0QlMjAtLW51bV9jbGFzc19pbWFnZXMlMjAyMDA=",highlighted:`pip install clip-retrieval
python retrieve.py --class_prompt {} --class_data_dir {} --num_class_images 200`,lang:"bash",wrap:!1});var b=l(w,4);a(b,{code:"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",highlighted:`<span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;path-to-save-model&quot;</span>

accelerate launch train_custom_diffusion.py \\
  --pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span>  \\
  --output_dir=<span class="hljs-variable">$OUTPUT_DIR</span> \\
  --concepts_list=./concept_list.json \\
  --with_prior_preservation --real_prior --prior_loss_weight=1.0 \\
  --resolution=512  \\
  --train_batch_size=2  \\
  --learning_rate=1e-5  \\
  --lr_warmup_steps=0 \\
  --max_train_steps=500 \\
  --num_class_images=200 \\
  --scale_lr --hflip  \\
  --modifier_token <span class="hljs-string">&quot;&lt;new1&gt;+&lt;new2&gt;&quot;</span> \\
  --push_to_hub`,lang:"bash",wrap:!1});var T=l(b,4);s(T,{title:"사람 얼굴에 대한 학습",local:"사람-얼굴에-대한-학습",headingTag:"h3"});var _=l(T,6);a(_,{code:"cGlwJTIwaW5zdGFsbCUyMGNsaXAtcmV0cmlldmFsJTBBcHl0aG9uJTIwcmV0cmlldmUucHklMjAtLWNsYXNzX3Byb21wdCUyMHBlcnNvbiUyMC0tY2xhc3NfZGF0YV9kaXIlMjByZWFsX3JlZyUyRnNhbXBsZXNfcGVyc29uJTIwLS1udW1fY2xhc3NfaW1hZ2VzJTIwMjAw",highlighted:`pip install clip-retrieval
python retrieve.py --class_prompt person --class_data_dir real_reg/samples_person --num_class_images 200`,lang:"bash",wrap:!1});var j=l(_,4);a(j,{code:"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",highlighted:`<span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;path-to-save-model&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;path-to-images&quot;</span>

accelerate launch train_custom_diffusion.py \\
  --pretrained_model_name_or_path=<span class="hljs-variable">$MODEL_NAME</span>  \\
  --instance_data_dir=<span class="hljs-variable">$INSTANCE_DIR</span> \\
  --output_dir=<span class="hljs-variable">$OUTPUT_DIR</span> \\
  --class_data_dir=./real_reg/samples_person/ \\
  --with_prior_preservation --real_prior --prior_loss_weight=1.0 \\
  --class_prompt=<span class="hljs-string">&quot;person&quot;</span> --num_class_images=200 \\
  --instance_prompt=<span class="hljs-string">&quot;photo of a &lt;new1&gt; person&quot;</span>  \\
  --resolution=512  \\
  --train_batch_size=2  \\
  --learning_rate=5e-6  \\
  --lr_warmup_steps=0 \\
  --max_train_steps=1000 \\
  --scale_lr --hflip --noaug \\
  --freeze_model crossattn \\
  --modifier_token <span class="hljs-string">&quot;&lt;new1&gt;&quot;</span> \\
  --enable_xformers_memory_efficient_attention \\
  --push_to_hub`,lang:"bash",wrap:!1});var f=l(j,2);s(f,{title:"추론",local:"추론",headingTag:"h2"});var g=l(f,4);a(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> DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(<span class="hljs-string">&quot;CompVis/stable-diffusion-v1-4&quot;</span>, torch_dtype=torch.float16).to(<span class="hljs-string">&quot;cuda&quot;</span>)
pipe.unet.load_attn_procs(<span class="hljs-string">&quot;path-to-save-model&quot;</span>, weight_name=<span class="hljs-string">&quot;pytorch_custom_diffusion_weights.bin&quot;</span>)
pipe.load_textual_inversion(<span class="hljs-string">&quot;path-to-save-model&quot;</span>, weight_name=<span class="hljs-string">&quot;&lt;new1&gt;.bin&quot;</span>)

image = pipe(
    <span class="hljs-string">&quot;&lt;new1&gt; cat sitting in a bucket&quot;</span>,
    num_inference_steps=<span class="hljs-number">100</span>,
    guidance_scale=<span class="hljs-number">6.0</span>,
    eta=<span class="hljs-number">1.0</span>,
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;cat.png&quot;</span>)`,lang:"python",wrap:!1});var Z=l(g,4);a(Z,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> huggingface_hub.repocard <span class="hljs-keyword">import</span> RepoCard
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

model_id = <span class="hljs-string">&quot;sayakpaul/custom-diffusion-cat&quot;</span>
card = RepoCard.load(model_id)
base_model_id = card.data.to_dict()[<span class="hljs-string">&quot;base_model&quot;</span>]

pipe = DiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16).to(<span class="hljs-string">&quot;cuda&quot;</span>)
pipe.unet.load_attn_procs(model_id, weight_name=<span class="hljs-string">&quot;pytorch_custom_diffusion_weights.bin&quot;</span>)
pipe.load_textual_inversion(model_id, weight_name=<span class="hljs-string">&quot;&lt;new1&gt;.bin&quot;</span>)

image = pipe(
    <span class="hljs-string">&quot;&lt;new1&gt; cat sitting in a bucket&quot;</span>,
    num_inference_steps=<span class="hljs-number">100</span>,
    guidance_scale=<span class="hljs-number">6.0</span>,
    eta=<span class="hljs-number">1.0</span>,
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;cat.png&quot;</span>)`,lang:"python",wrap:!1});var I=l(Z,4);a(I,{code:"aW1wb3J0JTIwdG9yY2glMEFmcm9tJTIwaHVnZ2luZ2ZhY2VfaHViLnJlcG9jYXJkJTIwaW1wb3J0JTIwUmVwb0NhcmQlMEFmcm9tJTIwZGlmZnVzZXJzJTIwaW1wb3J0JTIwRGlmZnVzaW9uUGlwZWxpbmUlMEElMEFtb2RlbF9pZCUyMCUzRCUyMCUyMnNheWFrcGF1bCUyRmN1c3RvbS1kaWZmdXNpb24tY2F0LXdvb2Rlbi1wb3QlMjIlMEFjYXJkJTIwJTNEJTIwUmVwb0NhcmQubG9hZChtb2RlbF9pZCklMEFiYXNlX21vZGVsX2lkJTIwJTNEJTIwY2FyZC5kYXRhLnRvX2RpY3QoKSU1QiUyMmJhc2VfbW9kZWwlMjIlNUQlMEElMEFwaXBlJTIwJTNEJTIwRGlmZnVzaW9uUGlwZWxpbmUuZnJvbV9wcmV0cmFpbmVkKGJhc2VfbW9kZWxfaWQlMkMlMjB0b3JjaF9kdHlwZSUzRHRvcmNoLmZsb2F0MTYpLnRvKCUyMmN1ZGElMjIpJTBBcGlwZS51bmV0LmxvYWRfYXR0bl9wcm9jcyhtb2RlbF9pZCUyQyUyMHdlaWdodF9uYW1lJTNEJTIycHl0b3JjaF9jdXN0b21fZGlmZnVzaW9uX3dlaWdodHMuYmluJTIyKSUwQXBpcGUubG9hZF90ZXh0dWFsX2ludmVyc2lvbihtb2RlbF9pZCUyQyUyMHdlaWdodF9uYW1lJTNEJTIyJTNDbmV3MSUzRS5iaW4lMjIpJTBBcGlwZS5sb2FkX3RleHR1YWxfaW52ZXJzaW9uKG1vZGVsX2lkJTJDJTIwd2VpZ2h0X25hbWUlM0QlMjIlM0NuZXcyJTNFLmJpbiUyMiklMEElMEFpbWFnZSUyMCUzRCUyMHBpcGUoJTBBJTIwJTIwJTIwJTIwJTIydGhlJTIwJTNDbmV3MSUzRSUyMGNhdCUyMHNjdWxwdHVyZSUyMGluJTIwdGhlJTIwc3R5bGUlMjBvZiUyMGElMjAlM0NuZXcyJTNFJTIwd29vZGVuJTIwcG90JTIyJTJDJTBBJTIwJTIwJTIwJTIwbnVtX2luZmVyZW5jZV9zdGVwcyUzRDEwMCUyQyUwQSUyMCUyMCUyMCUyMGd1aWRhbmNlX3NjYWxlJTNENi4wJTJDJTBBJTIwJTIwJTIwJTIwZXRhJTNEMS4wJTJDJTBBKS5pbWFnZXMlNUIwJTVEJTBBaW1hZ2Uuc2F2ZSglMjJtdWx0aS1zdWJqZWN0LnBuZyUyMik=",highlighted:`<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> huggingface_hub.repocard <span class="hljs-keyword">import</span> RepoCard
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

model_id = <span class="hljs-string">&quot;sayakpaul/custom-diffusion-cat-wooden-pot&quot;</span>
card = RepoCard.load(model_id)
base_model_id = card.data.to_dict()[<span class="hljs-string">&quot;base_model&quot;</span>]

pipe = DiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16).to(<span class="hljs-string">&quot;cuda&quot;</span>)
pipe.unet.load_attn_procs(model_id, weight_name=<span class="hljs-string">&quot;pytorch_custom_diffusion_weights.bin&quot;</span>)
pipe.load_textual_inversion(model_id, weight_name=<span class="hljs-string">&quot;&lt;new1&gt;.bin&quot;</span>)
pipe.load_textual_inversion(model_id, weight_name=<span class="hljs-string">&quot;&lt;new2&gt;.bin&quot;</span>)

image = pipe(
    <span class="hljs-string">&quot;the &lt;new1&gt; cat sculpture in the style of a &lt;new2&gt; wooden pot&quot;</span>,
    num_inference_steps=<span class="hljs-number">100</span>,
    guidance_scale=<span class="hljs-number">6.0</span>,
    eta=<span class="hljs-number">1.0</span>,
).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;multi-subject.png&quot;</span>)`,lang:"python",wrap:!1});var N=l(I,4);s(N,{title:"학습된 체크포인트에서 추론하기",local:"학습된-체크포인트에서-추론하기",headingTag:"h3"});var W=l(N,4);s(W,{title:"Grads를 None으로 설정",local:"grads를-none으로-설정",headingTag:"h2"});var X=l(W,6);s(X,{title:"실험 결과",local:"실험-결과",headingTag:"h2"});var B=l(X,4);S(B,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/ko/training/custom_diffusion.md"}),H(2),V(G,t),x()}export{ll as component};