Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /main /ja /training /dreambooth.html
mishig
1 day ago
download
raw
93.9 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;DreamBooth&quot;,&quot;local&quot;:&quot;dreambooth&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;スクリプトのパラメータ&quot;,&quot;local&quot;:&quot;script-parameters&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Min-SNR重み付け&quot;,&quot;local&quot;:&quot;min-snr-weighting&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;事前分布の保存損失&quot;,&quot;local&quot;:&quot;prior-preservation-loss&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;テキストエンコーダーの学習&quot;,&quot;local&quot;:&quot;train-text-encoder&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;学習スクリプト&quot;,&quot;local&quot;:&quot;training-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;スクリプトを実行します&quot;,&quot;local&quot;:&quot;launch-the-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;LoRA&quot;,&quot;local&quot;:&quot;lora&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Stable Diffusion XL&quot;,&quot;local&quot;:&quot;stable-diffusion-xl&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;DeepFloyd IF&quot;,&quot;local&quot;:&quot;deepfloyd-if&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;トレーニングのコツ&quot;,&quot;local&quot;:&quot;training-tips&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;次のステップ&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/diffusers/main/en/_app/immutable/entry/start.D3LnKZ4R.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/BF9UNEM9.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DPnyVmXs.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/entry/app.C0rBoMxR.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSGUT_OW.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/C045MWWx.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DTZsP_3B.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/IaDvMxLQ.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/316.CuftASja.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/bKdwVYpU.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DTd9IGSp.js" rel="modulepreload">
		<!--w5pb01--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;DreamBooth&quot;,&quot;local&quot;:&quot;dreambooth&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Script parameters&quot;,&quot;local&quot;:&quot;script-parameters&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Min-SNR weighting&quot;,&quot;local&quot;:&quot;min-snr-weighting&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Prior preservation loss&quot;,&quot;local&quot;:&quot;prior-preservation-loss&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Train text encoder&quot;,&quot;local&quot;:&quot;train-text-encoder&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Training script&quot;,&quot;local&quot;:&quot;training-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Launch the script&quot;,&quot;local&quot;:&quot;launch-the-script&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;LoRA&quot;,&quot;local&quot;:&quot;lora&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Stable Diffusion XL&quot;,&quot;local&quot;:&quot;stable-diffusion-xl&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;DeepFloyd IF&quot;,&quot;local&quot;:&quot;deepfloyd-if&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Training tips&quot;,&quot;local&quot;:&quot;training-tips&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Next steps&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
		<link href="/docs/diffusers/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"><div dir="ltr" class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" 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><!----></span> <span dir="auto">Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="dreambooth" 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="#dreambooth"><span><svg 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><!---->DreamBooth<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/papers/2208.12242" rel="nofollow">DreamBooth</a>は、対象やスタイルの画像をわずか数枚だけ用いて学習し、拡散モデル全体を更新する学習手法です。プロンプト内の特別な単語を例示画像と関連付けることで機能します。</p> <p>推論用に学習済みチェックポイントを読み込む場合は、<a href="../using-diffusers/dreambooth">推論用のDreamBoothモデルの読み込み</a>を参照してください。</p> <p>vRAMが限られたGPUで学習を行う場合は、学習コマンドで<code>gradient_checkpointing</code>と<code>mixed_precision</code>パラメータを有効にすることを試してください。また、<a href="../optimization/attention_backends">xFormers</a>によるメモリ効率の高いアテンションを使用することで、メモリ使用量を削減することもできます。</p> <p>このガイドでは、<a href="https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth.py" rel="nofollow">train_dreambooth.py</a> スクリプトを詳しく解説し、その仕組みをより深く理解したうえで、ご自身のユースケースに合わせてどのようにカスタマイズできるかについて説明します。</p> <p>スクリプトを実行する前に、ソースからライブラリをインストールしていることを確認してください。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->git <span class="hljs-built_in">clone</span> https://github.com/huggingface/diffusers
<span class="hljs-built_in">cd</span> diffusers
pip install .<!----></pre></div><!----> <p>トレーニングスクリプトを含むサンプルフォルダに移動し、使用しているスクリプトに必要な依存関係をインストールします:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">cd</span> examples/dreambooth
pip install -r requirements.txt<!----></pre></div><!----> <blockquote class="tip"><p>🤗 Accelerate は、複数 GPU/TPU での学習や混合精度学習を支援するためのライブラリです。ハードウェアや環境に基づいて、学習環境を自動的に設定します。詳細については、🤗 Accelerate の <a href="https://huggingface.co/docs/accelerate/quicktour" rel="nofollow">クイックツアー</a> をご覧ください。</p></blockquote> <p>🤗 Accelerate の環境を初期化します:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate config<!----></pre></div><!----> <p>設定を選択せずにデフォルトの🤗 Accelerate環境を設定するには:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate config default<!----></pre></div><!----> <p>または、ノートブックのようにインタラクティブシェルをサポートしていない環境では、以下を使用できます。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> accelerate.utils <span class="hljs-keyword">import</span> write_basic_config

write_basic_config()<!----></pre></div><!----> <p>最後に、独自のデータセットでモデルをトレーニングしたい場合は、<a href="create_dataset">トレーニング用のデータセットを作成する</a>ガイドを参照して、トレーニングスクリプトと互換性のあるデータセットの作成方法を確認してください。</p> <blockquote class="tip"><p>以下のセクションでは、スクリプトの修正方法を理解するために重要な部分に焦点を当てていますが、スクリプトのすべての側面を詳細に説明するものではありません。さらに詳しく知りたい場合は、<a href="https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth.py" rel="nofollow">スクリプト</a>をぜひお読みいただき、ご質問やご懸念事項があればお知らせください。</p></blockquote> <!--[1--><h2 class="relative group"><a id="script-parameters" 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="#script-parameters"><span><svg 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><!---->スクリプトのパラメータ<!----></span></h2><!--]--><!----> <blockquote class="warning"><p>DreamBoothは学習ハイパーパラメータに非常に敏感であり、オーバーフィッティングを起こしやすいです。適切なハイパーパラメータを選択するために、<a href="https://huggingface.co/blog/dreambooth" rel="nofollow">🧨 Diffusers を使用して Stable Diffusion を Dreambooth で学習する</a>というブログ記事で、異なる被写体向けの推奨設定を確認してください。</p></blockquote> <p>トレーニングスクリプトには、トレーニング実行をカスタマイズするための多くのパラメータが用意されています。すべてのパラメータとその説明は、<a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L228" rel="nofollow"><code>parse_args()</code></a>関数に記載されています。これらのパラメータには、そのままでも概ね良好な結果が得られるデフォルト値が設定されていますが、必要に応じてトレーニングコマンド内で独自の値を設定することもできます。</p> <p>例えば、bf16形式で学習を行うには:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_dreambooth.py \
    --mixed_precision=<span class="hljs-string">&quot;bf16&quot;</span><!----></pre></div><!----> <p>知っておくべき、かつ指定すべき基本的で重要なパラメータは以下の通りです。</p> <ul><li><code>--pretrained_model_name_or_path</code>: Hub 上のモデル名、または事前学習済みモデルへのローカルパス</li> <li><code>--instance_data_dir</code>: 学習データセット（例画像）を含むフォルダへのパス</li> <li><code>--instance_prompt</code>: 例示画像に使用される特別な単語を含むテキストプロンプト</li> <li><code>--train_text_encoder</code>: テキストエンコーダーも同時に学習するかどうか</li> <li><code>--output_dir</code>: 学習済みモデルを保存する場所</li> <li><code>--push_to_hub</code>: 学習したモデルを Hub にプッシュするかどうか</li> <li><code>--checkpointing_steps</code>: モデルの学習中にチェックポイントを保存する頻度です。学習が何らかの理由で中断された場合、学習コマンドに<code>--resume_from_checkpoint</code>を追加することで、そのチェックポイントから学習を再開できます。</li></ul> <!--[2--><h3 class="relative group"><a id="min-snr-weighting" 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="#min-snr-weighting"><span><svg 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><!---->Min-SNR重み付け<!----></span></h3><!--]--><!----> <p><a href="https://huggingface.co/papers/2303.09556" rel="nofollow">Min-SNR</a>重み付け戦略は、損失を再バランス化して収束を高速化することで、トレーニングを支援できます。トレーニングスクリプトは<code>epsilon</code>（ノイズ）または<code>v_prediction</code>の予測をサポートしていますが、Min-SNRはどちらの予測タイプにも対応しています。この重み付け戦略はPyTorchでのみサポートされています。</p> <p><code>--snr_gamma</code> パラメータを追加し、推奨値である 5.0 に設定します:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_dreambooth.py \
  --snr_gamma=5.0<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="prior-preservation-loss" 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="#prior-preservation-loss"><span><svg 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><!---->事前分布の保存損失<!----></span></h3><!--]--><!----> <p>Prior preservation loss は、モデル自身が生成したサンプルを用いて、より多様な画像の生成方法を学習させる手法です。これらの生成されたサンプル画像は、ユーザーが提供した画像と同じクラスに属するため、モデルがそのクラスについて学習した内容を保持し、クラスに関する既存の知識を活用して新しい構成を作成するのを助けます。</p> <ul><li><code>--with_prior_preservation</code>: 事前分布保持損失を使用するかどうか</li> <li><code>--prior_loss_weight</code>: モデルに対する事前分布保持損失の影響度を制御します</li> <li><code>--class_data_dir</code>: 生成されたクラスサンプル画像を含むフォルダへのパス</li> <li><code>--class_prompt</code>: 生成されるサンプル画像のクラスを説明するテキストプロンプト</li></ul> <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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_dreambooth.py \
  --with_prior_preservation \
  --prior_loss_weight=1.0 \
  --class_data_dir=<span class="hljs-string">&quot;path/to/class/images&quot;</span> \
  --class_prompt=<span class="hljs-string">&quot;text prompt describing class&quot;</span><!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="train-text-encoder" 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="#train-text-encoder"><span><svg 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><!---->テキストエンコーダーの学習<!----></span></h3><!--]--><!----> <p>生成される出力の品質を向上させるには、UNetに加えてテキストエンコーダーも学習させることができます。これには追加のメモリが必要であり、少なくとも24GBのvRAMを備えたGPUが求められます。必要なハードウェアを所有している場合、テキストエンコーダーの学習により、特に顔の画像生成においてより良い結果が得られます。このオプションを有効にするには、次の手順を実行します。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_dreambooth.py \
  --train_text_encoder<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="training-script" 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="#training-script"><span><svg 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><!---->学習スクリプト<!----></span></h2><!--]--><!----> <p>DreamBoothには独自のデータセットクラスが用意されています:</p> <ul><li><a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L604" rel="nofollow"><code>DreamBoothDataset</code></a>: 学習用の画像とクラス画像の前処理、およびプロンプトのトークン化を行います</li> <li><a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L738" rel="nofollow"><code>PromptDataset</code></a>: クラス画像を生成するためのプロンプト埋め込みを生成します</li></ul> <p><a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L842" rel="nofollow">prior preservation loss</a>を有効にした場合、クラス画像はここで生成されます:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!---->sample_dataset = PromptDataset(args.class_prompt, num_new_images)
sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)

sample_dataloader = accelerator.prepare(sample_dataloader)
pipeline.to(accelerator.device)

<span class="hljs-keyword">for</span> example <span class="hljs-keyword">in</span> tqdm(
    sample_dataloader, desc=<span class="hljs-string">&quot;Generating class images&quot;</span>, disable=<span class="hljs-keyword">not</span> accelerator.is_local_main_process
):
    images = pipeline(example[<span class="hljs-string">&quot;prompt&quot;</span>]).images<!----></pre></div><!----> <p>次に、<a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L799" rel="nofollow"><code>main()</code></a> 関数です。この関数は、学習用のデータセットのセットアップと、学習ループ自体を処理します。スクリプトは、<a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L898" rel="nofollow">tokenizer</a>、<a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L912C1-L912C1" rel="nofollow">schedulerとモデル</a>を読み込みます。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># Load the tokenizer</span>
<span class="hljs-keyword">if</span> args.tokenizer_name:
    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=<span class="hljs-literal">False</span>)
<span class="hljs-keyword">elif</span> args.pretrained_model_name_or_path:
    tokenizer = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder=<span class="hljs-string">&quot;tokenizer&quot;</span>,
        revision=args.revision,
        use_fast=<span class="hljs-literal">False</span>,
    )

<span class="hljs-comment"># Load scheduler and models</span>
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder=<span class="hljs-string">&quot;scheduler&quot;</span>)
text_encoder = text_encoder_cls.from_pretrained(
    args.pretrained_model_name_or_path, subfolder=<span class="hljs-string">&quot;text_encoder&quot;</span>, revision=args.revision
)

<span class="hljs-keyword">if</span> model_has_vae(args):
    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name_or_path, subfolder=<span class="hljs-string">&quot;vae&quot;</span>, revision=args.revision
    )
<span class="hljs-keyword">else</span>:
    vae = <span class="hljs-literal">None</span>

unet = UNet2DConditionModel.from_pretrained(
    args.pretrained_model_name_or_path, subfolder=<span class="hljs-string">&quot;unet&quot;</span>, revision=args.revision
)<!----></pre></div><!----> <p>次に、<code>DreamBoothDataset</code>から<a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L1073" rel="nofollow">学習データセット</a>とDataLoaderを作成します。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!---->train_dataset = DreamBoothDataset(
    instance_data_root=args.instance_data_dir,
    instance_prompt=args.instance_prompt,
    class_data_root=args.class_data_dir <span class="hljs-keyword">if</span> args.with_prior_preservation <span class="hljs-keyword">else</span> <span class="hljs-literal">None</span>,
    class_prompt=args.class_prompt,
    class_num=args.num_class_images,
    tokenizer=tokenizer,
    size=args.resolution,
    center_crop=args.center_crop,
    encoder_hidden_states=pre_computed_encoder_hidden_states,
    class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
    tokenizer_max_length=args.tokenizer_max_length,
)

train_dataloader = torch.utils.data.DataLoader(
    train_dataset,
    batch_size=args.train_batch_size,
    shuffle=<span class="hljs-literal">True</span>,
    collate_fn=<span class="hljs-keyword">lambda</span> examples: collate_fn(examples, args.with_prior_preservation),
    num_workers=args.dataloader_num_workers,
)<!----></pre></div><!----> <p>最後に、<a href="https://github.com/huggingface/diffusers/blob/072e00897a7cf4302c347a63ec917b4b8add16d4/examples/dreambooth/train_dreambooth.py#L1151" rel="nofollow">training loop</a>が、画像を潜在空間に変換すること、入力にノイズを追加すること、ノイズ残差を予測すること、損失を計算することなどの残りのステップを担当します。</p> <p>トレーニングループの仕組みについてさらに詳しく知りたい場合は、デノイジングプロセスの基本的なパターンを解説している<a href="../using-diffusers/write_own_pipeline">パイプライン、モデル、スケジューラーの理解</a>チュートリアルをご覧ください。</p> <!--[1--><h2 class="relative group"><a id="launch-the-script" 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="#launch-the-script"><span><svg 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><!---->スクリプトを実行します<!----></span></h2><!--]--><!----> <p>トレーニングスクリプトの起動準備が整いました! 🚀</p> <p>このガイドでは、<a href="https://huggingface.co/datasets/diffusers/dog-example" rel="nofollow">犬</a>の画像をダウンロードしてディレクトリに保存します。ただし、必要に応じて独自のデータセットを作成して使用することもできます（<a href="create_dataset">学習用データセットの作成</a>ガイドを参照）。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> snapshot_download

local_dir = <span class="hljs-string">&quot;./dog&quot;</span>
snapshot_download(
    <span class="hljs-string">&quot;diffusers/dog-example&quot;</span>,
    local_dir=local_dir,
    repo_type=<span class="hljs-string">&quot;dataset&quot;</span>,
    ignore_patterns=<span class="hljs-string">&quot;.gitattributes&quot;</span>,
)<!----></pre></div><!----> <p>環境変数 <code>MODEL_NAME</code> に Hub 上のモデル ID またはローカルモデルへのパスを、<code>INSTANCE_DIR</code> に先ほど犬の画像をダウンロードしたパスを、<code>OUTPUT_DIR</code> にモデルを保存したい場所を設定します。学習を紐付けるための特別な単語として <code>sks</code> を使用します。</p> <p>トレーニングの進捗を追跡したい場合は、トレーニング中に定期的に生成された画像を保存できます。トレーニングコマンドに次のパラメータを追加してください。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->--validation_prompt=<span class="hljs-string">&quot;a photo of a sks dog&quot;</span>
--num_validation_images=4
--validation_steps=100<!----></pre></div><!----> <p>スクリプトを実行する前に、もう1点だけ確認事項があります。お使いのGPUによっては、DreamBoothの学習のために特定の最適化を有効にする必要がある場合があります。</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">16GB</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">12GB</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">8GB</div><!--]--></div> <div class="language-select"><div><p>16GBのGPUでは、bitsandbytesの8ビットオプティマイザーと勾配チェックポイントを使用することで、DreamBoothモデルの学習を支援できます。bitsandbytesをインストールします:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!---->pip install bitsandbytes<!----></pre></div><!----> <p>次に、トレーニングコマンドに次のパラメータを追加します。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_dreambooth.py \
  --gradient_checkpointing \
  --use_8bit_adam \<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>12GBのGPUでは、メモリ使用量を削減するために、bitsandbytesの8ビットオプティマイザー、勾配チェックポイント、xFormersを使用し、勾配をゼロではなく<code>None</code>に設定する必要があります。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate launch train_dreambooth.py \
  --use_8bit_adam \
  --gradient_checkpointing \
  --enable_xformers_memory_efficient_attention \
  --set_grads_to_none \<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>8GBのGPUでは、vRAM上の一部のテンソルをCPUまたはNVMEにオフロードして、GPUメモリを少なくしてトレーニングできるようにするため、<a href="https://www.deepspeed.ai/" rel="nofollow">DeepSpeed</a>が必要です。</p> <p>🤗 Accelerate 環境を設定するには、次のコマンドを実行してください:</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!---->accelerate config<!----></pre></div><!----> <p>設定時に、DeepSpeed を使用することを確認してください。これにより、DeepSpeed stage 2、fp16 混合精度、およびモデルパラメータとオプティマイザ状態の CPU オフロードを組み合わせることで、8GB 未満の vRAM でトレーニングが可能になります。ただし、これにはより多くのシステム RAM（約 25 GB）が必要です。その他の設定オプションについては、<a href="https://huggingface.co/docs/accelerate/usage_guides/deepspeed" rel="nofollow">DeepSpeed ドキュメント</a>を参照してください。</p> <p>大幅な高速化のため、デフォルトの Adam オプティマイザーを DeepSpeed の最適化版である <a href="https://deepspeed.readthedocs.io/en/latest/optimizers.html#adam-cpu" rel="nofollow"><code>deepspeed.ops.adam.DeepSpeedCPUAdam</code></a> に変更することをお勧めします。<code>DeepSpeedCPUAdam</code> を有効にするには、システムの CUDA ツールチェーンのバージョンが、PyTorch にインストールされたものと一致している必要があります。</p> <p>bitsandbytes の 8-bit オプティマイザーは、現時点では DeepSpeed と互換性がないようです。</p> <p>以上です！トレーニングコマンドに追加のパラメータを指定する必要はありません。</p><!----></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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;stable-diffusion-v1-5/stable-diffusion-v1-5&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;./dog&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;path_to_saved_model&quot;</span>

accelerate launch train_dreambooth.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> \
  --instance_prompt=<span class="hljs-string">&quot;a photo of sks dog&quot;</span> \
  --resolution=512 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=1 \
  --learning_rate=5e-6 \
  --lr_scheduler=<span class="hljs-string">&quot;constant&quot;</span> \
  --lr_warmup_steps=0 \
  --max_train_steps=400 \
  --push_to_hub<!----></pre></div><!----> <p>トレーニングが完了したら、新しく学習したモデルを推論に使用できます!</p> <blockquote class="tip"><p>学習が完了する前に推論でモデルを試してみたいですか? 🤭 🤗 Accelerate の最新バージョンがインストールされていることを確認してください。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, UNet2DConditionModel
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> CLIPTextModel
<span class="hljs-keyword">import</span> torch

unet = UNet2DConditionModel.from_pretrained(<span class="hljs-string">&quot;path/to/model/checkpoint-100/unet&quot;</span>)

<span class="hljs-comment"># if you have trained with `--args.train_text_encoder` make sure to also load the text encoder</span>
text_encoder = CLIPTextModel.from_pretrained(<span class="hljs-string">&quot;path/to/model/checkpoint-100/checkpoint-100/text_encoder&quot;</span>)

pipeline = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;stable-diffusion-v1-5/stable-diffusion-v1-5&quot;</span>, unet=unet, text_encoder=text_encoder, dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)

image = pipeline(<span class="hljs-string">&quot;A photo of sks dog in a bucket&quot;</span>, num_inference_steps=<span class="hljs-number">50</span>, guidance_scale=<span class="hljs-number">7.5</span>).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;dog-bucket.png&quot;</span>)<!----></pre></div><!----></blockquote> <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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
<span class="hljs-keyword">import</span> torch

pipeline = DiffusionPipeline.from_pretrained(<span class="hljs-string">&quot;path_to_saved_model&quot;</span>, dtype=torch.float16, use_safetensors=<span class="hljs-literal">True</span>).to(<span class="hljs-string">&quot;cuda&quot;</span>)  <span class="hljs-comment"># or &quot;mps&quot;, &quot;xpu&quot;, &quot;cpu&quot;</span>
image = pipeline(<span class="hljs-string">&quot;A photo of sks dog in a bucket&quot;</span>, num_inference_steps=<span class="hljs-number">50</span>, guidance_scale=<span class="hljs-number">7.5</span>).images[<span class="hljs-number">0</span>]
image.save(<span class="hljs-string">&quot;dog-bucket.png&quot;</span>)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="lora" 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="#lora"><span><svg 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><!---->LoRA<!----></span></h2><!--]--><!----> <p>LoRAは、学習可能なパラメータ数を大幅に削減するための学習手法です。その結果、学習が高速化され、生成される重みは非常に小さい（約100MB）ため、保存しやすくなります。LoRAを用いた学習には、<a href="https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora.py" rel="nofollow">train_dreambooth_lora.py</a>スクリプトを使用してください。</p> <p>LoRAの学習スクリプトについては、<a href="lora">LoRAの学習</a>ガイドで詳しく説明しています。</p> <!--[1--><h2 class="relative group"><a id="stable-diffusion-xl" 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-xl"><span><svg 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 XL<!----></span></h2><!--]--><!----> <p>Stable Diffusion XL (SDXL) は高解像度の画像を生成する強力なテキストから画像へのモデルであり、そのアーキテクチャに第2のテキストエンコーダーを追加しています。<a href="https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_sdxl.py" rel="nofollow">train_dreambooth_lora_sdxl.py</a> スクリプトを使用して、LoRA で SDXL モデルを学習します。</p> <p>SDXLの学習スクリプトについては、<a href="sdxl">SDXLの学習</a>ガイドで詳しく説明しています。</p> <!--[1--><h2 class="relative group"><a id="deepfloyd-if" 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="#deepfloyd-if"><span><svg 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><!---->DeepFloyd IF<!----></span></h2><!--]--><!----> <p>DeepFloyd IF は、3段階の段階的ピクセル拡散モデルです。第1段階でベース画像を生成し、第2段階と第3段階でベース画像を段階的にアップスケールして、高解像度の1024x1024画像に仕上げます。<a href="https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora.py" rel="nofollow">train_dreambooth_lora.py</a> または <a href="https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth.py" rel="nofollow">train_dreambooth.py</a> スクリプトを使用して、LoRAまたはフルモデルで DeepFloyd IF モデルを学習できます。</p> <p>DeepFloyd IF は予測分散を使用していますが、Diffusers の学習スクリプトは予測誤差を使用しているため、学習済みの DeepFloyd IF モデルは固定分散スケジュールに切り替えられます。学習スクリプトは、完全に学習されたモデルのスケジューラ設定を自動的に更新します。ただし、保存された LoRA 重みをロードする際は、パイプラインのスケジューラ設定も手動で更新する必要があります。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(<span class="hljs-string">&quot;DeepFloyd/IF-I-XL-v1.0&quot;</span>, use_safetensors=<span class="hljs-literal">True</span>)

pipe.load_lora_weights(<span class="hljs-string">&quot;&lt;lora weights path&gt;&quot;</span>)

<span class="hljs-comment"># Update scheduler config to fixed variance schedule</span>
pipe.scheduler = pipe.scheduler.__class__.from_config(pipe.scheduler.config, variance_type=<span class="hljs-string">&quot;fixed_small&quot;</span>)<!----></pre></div><!----> <p>ステージ2のモデルは、アップスケールするために追加の検証画像が必要です。このために、トレーニング画像の縮小版をダウンロードして使用できます。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> snapshot_download

local_dir = <span class="hljs-string">&quot;./dog_downsized&quot;</span>
snapshot_download(
    <span class="hljs-string">&quot;diffusers/dog-example-downsized&quot;</span>,
    local_dir=local_dir,
    repo_type=<span class="hljs-string">&quot;dataset&quot;</span>,
    ignore_patterns=<span class="hljs-string">&quot;.gitattributes&quot;</span>,
)<!----></pre></div><!----> <p>以下のコードサンプルは、DreamBoothとLoRAの組み合わせでDeepFloyd IFモデルを学習する方法を簡潔に示しています。特に注意すべき重要なパラメータは以下の通りです。</p> <ul><li><code>--resolution=64</code>、DeepFloyd IF はピクセル拡散モデルであり、非圧縮ピクセルを扱うため、入力画像はより小さいサイズである必要があります</li> <li><code>--pre_compute_text_embeddings</code>、<a href="https://huggingface.co/docs/transformers/main/en/model_doc/t5#transformers.T5Model" rel="nofollow">T5Model</a> は大量のメモリを消費する可能性があるため、メモリを節約するためにテキスト埋め込みを事前に計算します</li> <li><code>--tokenizer_max_length=77</code> を指定すると、T5 をテキストエンコーダーとして使用する場合に、より長いデフォルトのテキスト長を使用できますが、デフォルトのモデルエンコーディング手順では、より短いテキスト長が使用されます。</li> <li><code>--text_encoder_use_attention_mask</code> を使用して、アテンションマスクをテキストエンコーダーに渡します。</li></ul> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">Stage 1 LoRA DreamBooth</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">ステージ2 LoRA DreamBooth</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">ステージ1 DreamBooth</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">ステージ2 DreamBooth</div><!--]--></div> <div class="language-select"><div><p>LoRAとDreamBoothを用いたDeepFloyd IFのトレーニングステージ1には、約28GBのメモリが必要です。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;DeepFloyd/IF-I-XL-v1.0&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;dog&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;dreambooth_dog_lora&quot;</span>

accelerate launch train_dreambooth_lora.py \
  --report_to wandb \
  --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> \
  --instance_prompt=<span class="hljs-string">&quot;a sks dog&quot;</span> \
  --resolution=64 \
  --train_batch_size=4 \
  --gradient_accumulation_steps=1 \
  --learning_rate=5e-6 \
  --scale_lr \
  --max_train_steps=1200 \
  --validation_prompt=<span class="hljs-string">&quot;a sks dog&quot;</span> \
  --validation_epochs=25 \
  --checkpointing_steps=100 \
  --pre_compute_text_embeddings \
  --tokenizer_max_length=77 \
  --text_encoder_use_attention_mask<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>LoRAとDreamBoothを用いたDeepFloyd IFのステージ2では、以下のパラメータに注意してください。</p> <ul><li><code>--validation_images</code>、検証時にアップスケールする画像</li> <li>ステージ2で必要に応じてUNetに追加の条件付けを行うために、<code>--class_labels_conditioning=timesteps</code>を指定します。</li> <li><code>--learning_rate=1e-6</code>、ステージ1と比較して低い学習率が使用されます</li> <li><code>--resolution=256</code>、アップスケーラーの想定解像度</li></ul> <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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;DeepFloyd/IF-II-L-v1.0&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;dog&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;dreambooth_dog_upscale&quot;</span>
<span class="hljs-built_in">export</span> VALIDATION_IMAGES=<span class="hljs-string">&quot;dog_downsized/image_1.png dog_downsized/image_2.png dog_downsized/image_3.png dog_downsized/image_4.png&quot;</span>

python train_dreambooth_lora.py \
    --report_to wandb \
    --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">$OUTPUT_DIR</span> \
    --instance_prompt=<span class="hljs-string">&quot;a sks dog&quot;</span> \
    --resolution=256 \
    --train_batch_size=4 \
    --gradient_accumulation_steps=1 \
    --learning_rate=1e-6 \
    --max_train_steps=2000 \
    --validation_prompt=<span class="hljs-string">&quot;a sks dog&quot;</span> \
    --validation_epochs=100 \
    --checkpointing_steps=500 \
    --pre_compute_text_embeddings \
    --tokenizer_max_length=77 \
    --text_encoder_use_attention_mask \
    --validation_images <span class="hljs-variable">$VALIDATION_IMAGES</span> \
    --class_labels_conditioning=timesteps<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>DreamBooth を使用した DeepFloyd IF のステージ 1 では、以下のパラメータに注意してください。</p> <ul><li><code>--skip_save_text_encoder</code>、ファインチューニング済みモデルとともに完全な T5 テキストエンコーダーの保存をスキップします</li> <li><code>--use_8bit_adam</code>、フルモデルの学習時にオプティマイザ状態のサイズによるメモリ消費を削減するために、8-bit Adamオプティマイザを使用します。</li> <li><code>--learning_rate=1e-7</code>、フルモデルの学習には非常に低い学習率を使用する必要があります。それ以外の場合、モデルの品質が低下します（より大きなバッチサイズを使用する場合は、より高い学習率を使用できます）。</li></ul> <p>8-bit Adam とバッチサイズ 4 で学習を行う場合、フルモデルは約 48GB のメモリで学習できます。</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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;DeepFloyd/IF-I-XL-v1.0&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;dog&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;dreambooth_if&quot;</span>

accelerate launch train_dreambooth.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> \
  --instance_prompt=<span class="hljs-string">&quot;a photo of sks dog&quot;</span> \
  --resolution=64 \
  --train_batch_size=4 \
  --gradient_accumulation_steps=1 \
  --learning_rate=1e-7 \
  --max_train_steps=150 \
  --validation_prompt <span class="hljs-string">&quot;a photo of sks dog&quot;</span> \
  --validation_steps 25 \
  --text_encoder_use_attention_mask \
  --tokenizer_max_length 77 \
  --pre_compute_text_embeddings \
  --use_8bit_adam \
  --set_grads_to_none \
  --skip_save_text_encoder \
  --push_to_hub<!----></pre></div><!----><!----></div><!----> <div hidden=""><p>DreamBooth を用いた DeepFloyd IF のステージ 2 では、以下のパラメータに注意してください：</p> <ul><li><code>--learning_rate=5e-6</code>、より小さい実効バッチサイズを使用する場合は、より低い学習率を使用します</li> <li><code>--resolution=256</code>、アップスケーラーの想定解像度</li> <li><code>--train_batch_size=2</code> および <code>--gradient_accumulation_steps=6</code> を使用すると、顔を含む画像で効果的に学習するには、より大きなバッチサイズが必要です。</li></ul> <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 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> コピーしました</div><!----></button><!----></div> <pre class="language-bash "><!----><span class="hljs-built_in">export</span> MODEL_NAME=<span class="hljs-string">&quot;DeepFloyd/IF-II-L-v1.0&quot;</span>
<span class="hljs-built_in">export</span> INSTANCE_DIR=<span class="hljs-string">&quot;dog&quot;</span>
<span class="hljs-built_in">export</span> OUTPUT_DIR=<span class="hljs-string">&quot;dreambooth_dog_upscale&quot;</span>
<span class="hljs-built_in">export</span> VALIDATION_IMAGES=<span class="hljs-string">&quot;dog_downsized/image_1.png dog_downsized/image_2.png dog_downsized/image_3.png dog_downsized/image_4.png&quot;</span>

accelerate launch train_dreambooth.py \
  --report_to wandb \
  --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">$OUTPUT_DIR</span> \
  --instance_prompt=<span class="hljs-string">&quot;a sks dog&quot;</span> \
  --resolution=256 \
  --train_batch_size=2 \
  --gradient_accumulation_steps=6 \
  --learning_rate=5e-6 \
  --max_train_steps=2000 \
  --validation_prompt=<span class="hljs-string">&quot;a sks dog&quot;</span> \
  --validation_steps=150 \
  --checkpointing_steps=500 \
  --pre_compute_text_embeddings \
  --tokenizer_max_length=77 \
  --text_encoder_use_attention_mask \
  --validation_images <span class="hljs-variable">$VALIDATION_IMAGES</span> \
  --class_labels_conditioning timesteps \
  --push_to_hub<!----></pre></div><!----><!----></div><!----><!----></div><!----> <!--[2--><h3 class="relative group"><a id="training-tips" 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="#training-tips"><span><svg 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><!---->トレーニングのコツ<!----></span></h3><!--]--><!----> <p>DeepFloyd IFモデルの学習は難しい場合がありますが、以下に有用だと感じたヒントをいくつかご紹介します。</p> <ul><li>モデルの解像度が低いため、細部の表現はどのみち困難であるため、ステージ1モデルの学習にはLoRAで十分です。</li> <li>一般的かつシンプルなオブジェクトの場合、アップスケーラーのファインチューニングは必ずしも必要ありません。アップスケーラーに渡すプロンプトを調整し、インスタンスプロンプトから新しいトークンを削除するようにしてください。例えば、ステージ1のプロンプトが「a sks dog」である場合、ステージ2のプロンプトは「a dog」にする必要があります。</li> <li>顔などの細部をより鮮明にするためには、ステージ2のアップスケーラーを完全に学習させる方が、LoRAを使用してステージ2のモデルを学習させるよりも効果的です。また、より低い学習率とより大きなバッチサイズを使用することも有効です。</li> <li>ステージ2のモデルを学習させるには、より低い学習率を使用する必要があります。</li> <li><a href="/docs/diffusers/main/ja/api/schedulers/ddpm#diffusers.DDPMScheduler">DDPMScheduler</a>は、学習スクリプトで使用されているDPMSolverよりも優れた動作を示します。</li></ul> <!--[1--><h2 class="relative group"><a id="next-steps" 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="#next-steps"><span><svg 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><!---->次のステップ<!----></span></h2><!--]--><!----> <p>DreamBoothモデルの学習が完了しました。おめでとうございます。新しいモデルの使用方法について詳しく知るために、以下のガイドが役立つ可能性があります。</p> <ul><li>推論用に<a href="../using-diffusers/dreambooth">DreamBoothのチェックポイントをロードする方法</a>を学びましょう。</li> <li>LoRAで学習を行った場合は、<a href="../tutorials/using_peft_for_inference">推論のためのLoRA</a>を参照してください。</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/en/training/dreambooth.md" target="_blank"><svg class="mr-1" 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="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->
			
			<script>
				{
					__sveltekit_g37pnb = {
						base: "/docs/diffusers/main/en",
						assets: "/docs/diffusers/main/en"
					};
					const element = document.currentScript.parentElement;
					Promise.all([
						import("/docs/diffusers/main/en/_app/immutable/entry/start.D3LnKZ4R.js"),
						import("/docs/diffusers/main/en/_app/immutable/entry/app.C0rBoMxR.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 316],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 93.9 kB
Xet hash:: 4eacbdcc79215d518848e50e267f212e64b10b0d6a1842d4c902c21963d5f332
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.