Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Cosmos 3","local":"cosmos-3","sections":[{"title":"Cosmos 3の新機能","local":"whats-new-in-cosmos-3","sections":[],"depth":2},{"title":"利用可能なチェックポイント","local":"available-checkpoints","sections":[],"depth":2},{"title":"FP8 混合 W8A8/W8A16 デノイジング","local":"fp8-mixed-w8a8w8a16-denoising","sections":[],"depth":2},{"title":"プロンプトのアップサンプリング","local":"prompt-upsampling","sections":[],"depth":2},{"title":"テキストから動画へ","local":"text-to-video","sections":[],"depth":2},{"title":"テキストから画像へ","local":"text-to-image","sections":[],"depth":2},{"title":"画像から動画へ","local":"image-to-video","sections":[],"depth":2},{"title":"Video-to-video","local":"video-to-video","sections":[],"depth":2},{"title":"音声付きのVideo-to-video","local":"video-to-video-with-sound","sections":[],"depth":2},{"title":"音声付きのテキストから動画へ","local":"text-to-video-with-sound","sections":[],"depth":2},{"title":"アクション条件付き生成","local":"action-conditioned-generation","sections":[{"title":"行動ポリシー","local":"action-policy","sections":[],"depth":3}],"depth":2},{"title":"SeaCache","local":"seacache","sections":[],"depth":2},{"title":"コンテキスト並列","local":"context-parallelism","sections":[{"title":"実行する","local":"run-it","sections":[],"depth":3},{"title":"テンソル並列性を用いた大規模モデルの学習","local":"fitting-large-models-with-tensor-parallelism","sections":[],"depth":3},{"title":"独自のモジュール化パイプラインで使用する","local":"use-it-in-your-own-modular-pipeline","sections":[],"depth":3}],"depth":2},{"title":"メタデータテンプレート","local":"metadata-templates","sections":[],"depth":2},{"title":"セーフティチェッカー","local":"safety-checker","sections":[],"depth":2},{"title":"Cosmos3OmniPipeline","local":"diffusers.Cosmos3OmniPipeline","sections":[],"depth":2},{"title":"Cosmos3OmniModularPipeline","local":"cosmos3omnimodularpipeline","sections":[{"title":"既存のすべてのワークフロー向けのモジュール化された例","local":"modular-examples-for-all-existing-workflows","sections":[],"depth":3},{"title":"モジュール転送（構造制御）","local":"modular-transfer-structural-control","sections":[],"depth":3},{"title":"蒸留(少数ステップ)のテキストから画像へ、画像から動画へ","local":"diffusers.Cosmos3OmniModularPipeline","sections":[],"depth":3}],"depth":2},{"title":"Cosmos3DistilledModularPipeline","local":"diffusers.Cosmos3DistilledModularPipeline","sections":[],"depth":2},{"title":"CosmosActionCondition","local":"diffusers.CosmosActionCondition","sections":[],"depth":2},{"title":"Cosmos3OmniPipelineOutput","local":"diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput","sections":[],"depth":2}],"depth":1}"/> | |
| <link href="/docs/diffusers/main/en/_app/immutable/entry/start.U6TUchwJ.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/chunks/yxSHiGDF.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/chunks/CPfOwUO_.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/entry/app.3qKhPoX-.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/chunks/BYe-f74j.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/chunks/D32hXVzh.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DqpuBF3n.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/chunks/Dyp9xKq8.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/nodes/7.MVmUqGUj.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/chunks/Dk44y1gU.js" rel="modulepreload"> | |
| <link href="/docs/diffusers/main/en/_app/immutable/chunks/CYniLWKs.js" rel="modulepreload"> | |
| <!--if5dit--><meta name="hf:doc:metadata" content="{"title":"Cosmos 3","local":"cosmos-3","sections":[{"title":"What’s new in Cosmos 3","local":"whats-new-in-cosmos-3","sections":[],"depth":2},{"title":"Available checkpoints","local":"available-checkpoints","sections":[],"depth":2},{"title":"FP8 mixed W8A8/W8A16 denoising","local":"fp8-mixed-w8a8w8a16-denoising","sections":[],"depth":2},{"title":"Prompt upsampling","local":"prompt-upsampling","sections":[],"depth":2},{"title":"Text-to-video","local":"text-to-video","sections":[],"depth":2},{"title":"Text-to-image","local":"text-to-image","sections":[],"depth":2},{"title":"Image-to-video","local":"image-to-video","sections":[],"depth":2},{"title":"Video-to-video","local":"video-to-video","sections":[],"depth":2},{"title":"Video-to-video with sound","local":"video-to-video-with-sound","sections":[],"depth":2},{"title":"Text-to-video with sound","local":"text-to-video-with-sound","sections":[],"depth":2},{"title":"Action-conditioned generation","local":"action-conditioned-generation","sections":[{"title":"Action policy","local":"action-policy","sections":[],"depth":3}],"depth":2},{"title":"SeaCache","local":"seacache","sections":[],"depth":2},{"title":"Context parallelism","local":"context-parallelism","sections":[{"title":"Run it","local":"run-it","sections":[],"depth":3},{"title":"Fitting large models with tensor parallelism","local":"fitting-large-models-with-tensor-parallelism","sections":[],"depth":3},{"title":"Use it in your own modular pipeline","local":"use-it-in-your-own-modular-pipeline","sections":[],"depth":3}],"depth":2},{"title":"Metadata templates","local":"metadata-templates","sections":[],"depth":2},{"title":"Safety checker","local":"safety-checker","sections":[],"depth":2},{"title":"Cosmos3OmniPipeline","local":"diffusers.Cosmos3OmniPipeline","sections":[],"depth":2},{"title":"Cosmos3OmniModularPipeline","local":"cosmos3omnimodularpipeline","sections":[{"title":"Modular examples for all existing workflows","local":"modular-examples-for-all-existing-workflows","sections":[],"depth":3},{"title":"Modular transfer (structural control)","local":"modular-transfer-structural-control","sections":[],"depth":3},{"title":"Distilled (few-step) text-to-image and image-to-video","local":"diffusers.Cosmos3OmniModularPipeline","sections":[],"depth":3}],"depth":2},{"title":"Cosmos3DistilledModularPipeline","local":"diffusers.Cosmos3DistilledModularPipeline","sections":[],"depth":2},{"title":"CosmosActionCondition","local":"diffusers.CosmosActionCondition","sections":[],"depth":2},{"title":"Cosmos3OmniPipelineOutput","local":"diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput","sections":[],"depth":2}],"depth":1}"/><!----> | |
| <link href="/docs/diffusers/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <!--[0--><h1 class="relative group"><a id="cosmos-3" 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="#cosmos-3"><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><!---->Cosmos 3<!----></span></h1><!--]--><!----> <p>NVIDIA Cosmos 3 は、Physical AI 向けの統合型ワールドファウンデーションモデル（WFM）です。ワールド生成、物理的推論、アクション生成を組み合わせる単一のオムニモデルとして機能します。以前の Cosmos リリースにおける個別の Predict、Reason、Transfer モデルに代わる存在であり、ロボティクス、自動運転車、スマートスペースのいずれの開発においても、Cosmos 3 は物理的な世界をシミュレートし理解するための単一の基盤を提供します。</p> <p>今回のリリースで提供される内容:</p> <ul><li>モデルカードとライセンスを備えた Hugging Face Hub 上のモデル</li> <li>Cosmos 3 Diffusers による生成パイプライン統合（このページ）</li> <li>Cosmos 3 を独自のデータでファインチューニングするためのポストトレーニングスクリプト</li> <li>Physical AI向けのオープンな合成データ生成(SDG)データセット</li></ul> <!--[1--><h2 class="relative group"><a id="whats-new-in-cosmos-3" 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="#whats-new-in-cosmos-3"><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><!---->Cosmos 3の新機能<!----></span></h2><!--]--><!----> <p>従来の Cosmos リリースからの最大の変更点は、Cosmos 3 が Mixture-of-Transformers (MoT) アーキテクチャを基盤とする<em>オムニモデル</em>であることです。以前は、ワールド生成 (Predict)、制御された生成 (Transfer)、シーン理解 (Reason)、およびアクションポリシー生成のために、それぞれ別々のモデルを使用していました。Cosmos 3 では、これらすべてが単一のモデルに統合され、単一のフォワードパス内で、モダリティを横断して推論と生成を行います。</p> <p>1つのモデルから、以下を行うことができます:</p> <ul><li>テキスト、画像、またはアクション入力から物理的に妥当な動画世界を生成します（画像から動画への変換、テキストから動画への変換、アクション条件付き動画生成）。</li> <li>運動、因果関係、空間的な関係などの物理的性質について推論します。</li> <li>現在の状態から、将来の動画とアクションのシーケンスを予測します。</li> <li>構造的な制御を用いて、視点や条件をまたいでシーンを転送します。<em>(近日公開)</em></li></ul> <p>内部では、単一の<code>Cosmos3OmniTransformer</code>が、Qwenスタイルの言語モデルと拡散生成パスを並行して実行します。テキストトークンは因果的な「理解」ストリームを流れ、動画と音声の潜在変数は双方向に注意を払う「生成」ストリームを流れ、3DマルチモーダルRoPEによって結合されます。アーキテクチャの背景については、<a href="https://huggingface.co/papers/2501.03575" rel="nofollow">Cosmos World Foundation Model Platform論文</a>を参照してください。</p> <!--[1--><h2 class="relative group"><a id="available-checkpoints" 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="#available-checkpoints"><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><!---->利用可能なチェックポイント<!----></span></h2><!--]--><!----> <p>Hubには2つのチェックポイントが公開されています — <a href="https://huggingface.co/nvidia/Cosmos3-Nano" rel="nofollow"><code>nvidia/Cosmos3-Nano</code></a>（小型で高速）と <a href="https://huggingface.co/nvidia/Cosmos3-Super" rel="nofollow"><code>nvidia/Cosmos3-Super</code></a>（大型で高品質）です。同じパイプラインクラスが、テキストから画像へ、テキストから動画へ、画像から動画へ、そして（音声対応のチェックポイントを使用する場合）テキスト+画像から音声付き動画への変換をサポートしています。リポジトリを選択し、以下の各ワークフローでモデルごとのタブを使用してください。</p> <blockquote class="tip"><p>スケジューラーのトレードオフ、すなわち速度と品質のバランスをどのように探求するかを学ぶには、<a href="../../using-diffusers/schedulers">ガイド</a>を必ずご確認ください。また、同じコンポーネントを複数のパイプラインに効率的に読み込む方法については、<a href="../../using-diffusers/loading#reusing-models-in-multiple-pipelines">パイプライン間でコンポーネントを再利用する</a>セクションを参照してください。</p></blockquote> <!--[1--><h2 class="relative group"><a id="fp8-mixed-w8a8w8a16-denoising" 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="#fp8-mixed-w8a8w8a16-denoising"><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><!---->FP8 混合 W8A8/W8A16 デノイジング<!----></span></h2><!--]--><!----> <p>Nano、Super、および Edge 用の ModelOpt FP8 公式チェックポイントは、それぞれ独立した Hub リポジトリとして公開されています: <a href="https://huggingface.co/nvidia/Cosmos3-Nano-FP8" rel="nofollow"><code>nvidia/Cosmos3-Nano-FP8</code></a>、<a href="https://huggingface.co/nvidia/Cosmos3-Super-FP8" rel="nofollow"><code>nvidia/Cosmos3-Super-FP8</code></a>、および <a href="https://huggingface.co/nvidia/Cosmos3-Edge-FP8" rel="nofollow"><code>nvidia/Cosmos3-Edge-FP8</code></a>。FP8 リポジトリの <code>main</code> を読み込みます。<a href="https://huggingface.co/nvidia/Cosmos3-Super-Image2Video" rel="nofollow"><code>nvidia/Cosmos3-Super-Image2Video</code></a> には独立した FP8 リポジトリはありません。 <code>revision="fp8"</code> を指定して読み込みます。</p> <p>これらのチェックポイントはすべて同じ方法で量子化されています。<strong>W8A8</strong>は8ビットの重みと8ビットの活性化値を使用します（復元されたModelOpt GEMM）。<strong>W8A16</strong>は同じ8ビットの重みを再利用しますが、活性化値の量子化をスキップします。FP8の重みはデクォンタイズされ、BF16/FP16/FP32の活性化値に対して標準的な線形演算が実行されます。</p> <p>すべてのステップでW8A8を実行すると、マルチステップの動画生成において目に見えるちらつきが発生する可能性があります。そのため、動画のNano、Super、Super-I2V、およびEdge FP8のチェックポイントは、<code>transformer/config.json</code>にスケジュールを宣言しています。<strong>最初の3ステップと最後の3ステップではW8A16</strong>、<strong>中間ではW8A8</strong>を使用します。Diffusersはこれらのカウントをハードコードするのではなく、チェックポイントから読み取ります。精度はスケジューラーのステップごとに1回選択されるため、classifier-free guidanceのcond/uncond呼び出しが一致します。</p> <p>画像生成および少数ステップ蒸留チェックポイントでは、そのようなフリッカーは観測されません。そのため、Super-T2Iおよび蒸留4ステップFP8リポジトリはスケジュールを宣言せず、すべてのステップでW8A8を維持します。また、このスケジュールは<strong>ModelOpt FP8専用</strong>です。その他の量子化バックエンド（TorchAOなど）は、それぞれのネイティブなフォワード処理を維持します。</p> <p><a href="../../quantization/modelopt">ModelOpt ガイド</a>と同じ復元パスを持つ FP8 リポジトリを読み込みます。リポジトリにはすでに量子化設定が含まれています。混合精度はチェックポイントに従って自動的に設定されます。</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-python "><!----><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano-FP8"</span>, | |
| dtype=torch.bfloat16, | |
| device_map=<span class="hljs-string">"cuda"</span>, | |
| ) | |
| result = pipe(prompt=<span class="hljs-string">"..."</span>, num_inference_steps=<span class="hljs-number">35</span>)<!----></pre></div><!----> <p>生成時の2つの選択肢:</p> <ul><li><strong>デフォルト</strong> (<code>mixed_precision_format=None</code>): チェックポイントが <code>diffusion_step_policy</code> を宣言している場合、最初の N ステップと最後の N ステップでは W8A16 を実行し、中間ステップではネイティブの W8A8 を実行します。これはマルチステップの <strong>video</strong> FP8 に対して意図されたレシピです（全ステップ W8A8 よりもフリッカーが少ない）。蒸留された 4 ステップおよび Super-T2I FP8 はポリシーを省略しているため、デフォルトではすでに全ステップ W8A8 になります。</li> <li><strong><code>mixed_precision_format="none"</code></strong>: すべてのステップをネイティブの W8A8 で実行します。W8A16 はデクアンタ化と <code>torch.nn.functional.linear</code> による処理であり、復元された FP8 GEMM とは異なるため、より高速です。ただし、マルチステップの動画ではちらつきが発生する可能性があります。スケジュールの A/B テストや、完全に量子化されたベースラインとの比較に使用してください。</li></ul> <p>Blackwellワークステーション1台で、720×1280 / 35ステップのCosmos3-Nano-FP8では、デフォルトの混合スケジュールを使用した場合、<code>"none"</code>を使用した場合と比較して、(T2I)で約<strong>27%遅く</strong>、(49フレームのT2V)で約<strong>13%遅く</strong>なりました。これらの数値はスループットを保証するものではありません。ポリシーが設定されていないModelOpt FP8チェックポイントに対して、最初のNステップ/最後のNステップのスケジュールを強制する場合は、<code>"fp8"</code>を渡してください。</p> <!--[1--><h2 class="relative group"><a id="prompt-upsampling" 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="#prompt-upsampling"><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><!---->プロンプトのアップサンプリング<!----></span></h2><!--]--><!----> <p>Cosmos 3 は、長く詳細なキャプションで学習されています。最適な品質を得るためには、短いテキストプロンプトをパイプラインに渡す前に、<strong>特定の JSON 構造にアップサンプリング</strong>する必要があります。アップサンプラーは <a href="https://github.com/NVIDIA/cosmos-framework" rel="nofollow">cosmos-framework</a> パッケージに含まれています。</p> <p>短いプレーンテキストのプロンプトを作成し、<code>assets/prompt.txt</code>に保存します。以下のテキストから動画への例では、元のプロンプトは<em>“A robotic arm is cleaning a plate in a kitchen”</em>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">mkdir</span> -p assets | |
| <span class="hljs-built_in">echo</span> <span class="hljs-string">"A robotic arm is cleaning a plate in a kitchen"</span> > assets/prompt.txt<!----></pre></div><!----> <p>次に、フレームワークをインストールしてアップサマーを実行します。以下の例では、Opus-4.6 を使用してテキストから動画へのアップサンプリングを行います。</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 "><!---->git <span class="hljs-built_in">clone</span> https://github.com/NVIDIA/cosmos-framework.git packages/cosmos-framework | |
| pip install -e packages/cosmos-framework | |
| <span class="hljs-built_in">export</span> PROMPT_UPSAMPLER_ENDPOINT_URL=<span class="hljs-string">"https://api.anthropic.com/v1/"</span> | |
| <span class="hljs-built_in">export</span> PROMPT_UPSAMPLER_MODEL_NAME=<span class="hljs-string">"claude-opus-4-6"</span> | |
| <span class="hljs-built_in">export</span> PROMPT_UPSAMPLER_API_TOKEN=<span class="hljs-string">"<your_token>"</span> | |
| python -m cosmos_framework.inference.prompt_upsampling \ | |
| --input assets/prompt.txt \ | |
| --output assets/example_t2v_prompt.json \ | |
| --mode text2video \ | |
| --endpoint-url <span class="hljs-string">"<span class="hljs-variable">${PROMPT_UPSAMPLER_ENDPOINT_URL}</span>"</span> \ | |
| --model <span class="hljs-string">"<span class="hljs-variable">${PROMPT_UPSAMPLER_MODEL_NAME}</span>"</span> \ | |
| --api-token <span class="hljs-string">"<span class="hljs-variable">${PROMPT_UPSAMPLER_API_TOKEN}</span>"</span> \ | |
| --resolution 720 \ | |
| --aspect-ratio <span class="hljs-string">"16,9"</span><!----></pre></div><!----> <p>ターゲットとするワークフローに合わせて、<code>--mode</code> を切り替えてください（<code>text2image</code>、<code>text2video</code>、<code>image2video</code>）。このコマンドは、アップサンプリングされたプロンプトを JSON 配列として <code>--output</code> ファイルに書き出します（<code>--input</code> の空でない各行につき 1 つのオブジェクト）。1 行につき 1 つの JSON オブジェクトを取得するには、代わりに <code>.jsonl</code> パスを指定してください。<code>image2video</code> の場合、<code>--image-url</code>（URL またはローカルパス）または <code>--image-list</code>（プロンプトごとに 1 枚の画像）を介して条件付け画像を指定する必要があります。</p> <p>利便性のため、事前にアップサンプリングされた正のプロンプト（<code>assets/example_t2v_prompt.json</code>）と負のプロンプト（<code>assets/negative_prompt.json</code>）が用意されており、以下の生成例で使用されています。これらの例では、JSON ファイルを読み込み、<code>json.dumps(...)</code> を介して JSON 文字列としてパイプラインに渡します。</p> <!--[1--><h2 class="relative group"><a id="text-to-video" 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="#text-to-video"><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><!---->テキストから動画へ<!----></span></h2><!--]--><!----> <p>テキストのみを条件とするマルチフレーム生成。<code>num_frames</code> は対象の再生時間に基づいて選択してください。デフォルトの <code>num_frames=189</code> は、24 FPS で約 7.9 秒の動画を生成します。プロンプトとネガティブプロンプトは、<a href="#prompt-upsampling">Prompt upsampling</a> で説明されている JSON アップサンプリング済みファイルから読み込まれます。</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">6.0</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"cosmos3_t2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">6.0</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"cosmos3_t2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="text-to-image" 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="#text-to-image"><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><!---->テキストから画像へ<!----></span></h2><!--]--><!----> <p>単一フレームの生成。モデルはテキストプロンプトのみで条件付けられます。<code>num_frames=1</code>を渡してください。<code>--mode text2image</code>でアップサンプリングして、JSONプロンプトを生成します。</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-comment"># JSON-upsampled prompt (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2i_prompt.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| result = pipe(prompt=json.dumps(json_prompt), num_frames=<span class="hljs-number">1</span>, height=<span class="hljs-number">720</span>, width=<span class="hljs-number">1280</span>) | |
| result.video[<span class="hljs-number">0</span>].save(<span class="hljs-string">"cosmos3_t2i.jpg"</span>, <span class="hljs-built_in">format</span>=<span class="hljs-string">"JPEG"</span>, quality=<span class="hljs-number">85</span>)<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-comment"># JSON-upsampled prompt (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2i_prompt.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| result = pipe(prompt=json.dumps(json_prompt), num_frames=<span class="hljs-number">1</span>, height=<span class="hljs-number">720</span>, width=<span class="hljs-number">1280</span>) | |
| result.video[<span class="hljs-number">0</span>].save(<span class="hljs-string">"cosmos3_t2i.jpg"</span>, <span class="hljs-built_in">format</span>=<span class="hljs-string">"JPEG"</span>, quality=<span class="hljs-number">85</span>)<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="image-to-video" 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="#image-to-video"><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><!---->画像から動画へ<!----></span></h2><!--]--><!----> <p><code>image=</code> を介して条件付け画像を渡します。パイプラインはフレーム 0 を指定された画像に固定し、残りのフレームをデノイズします。画像はアスペクト比を維持したままリサイズされ、要求された出力サイズにセンタークロップされ、uint8 相当の丸めによって正規化されます。<code>--mode image2video</code> を使用してアップサンプリングし、JSON プロンプトを生成します。</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_image | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_i2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| image = load_image( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/releases/download/assets/robot_153.jpg"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| image=image, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"cosmos3_i2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_image | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_i2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| image = load_image( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/releases/download/assets/robot_153.jpg"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| image=image, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"cosmos3_i2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="video-to-video" 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="#video-to-video"><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><!---->Video-to-video<!----></span></h2><!--]--><!----> <p><code>video=</code> を介して条件付け用のクリップを渡します（<code>load_video</code> から取得するなど）。パイプラインは、<code>condition_frame_indexes_vision</code>（デフォルトは <code>[0, 1]</code>）で指定される先頭の潜在フレームをクリップに固定し、残りのフレームをデノイズします。より長いソースクリップから条件付けフレームをどの端から取得するかは、<code>condition_video_keep</code>（<code>"first"</code> または <code>"last"</code>）で選択できます。他のモードと同様に、プロンプトは <a href="#prompt-upsampling">Prompt upsampling</a> で説明されている記述的な JSON 構造に従う必要があります。</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_v2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| video=video, | |
| condition_frame_indexes_vision=[<span class="hljs-number">0</span>, <span class="hljs-number">1</span>], | |
| condition_video_keep=<span class="hljs-string">"first"</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">6.0</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"cosmos3_v2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_v2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| video=video, | |
| condition_frame_indexes_vision=[<span class="hljs-number">0</span>, <span class="hljs-number">1</span>], | |
| condition_video_keep=<span class="hljs-string">"first"</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">6.0</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"cosmos3_v2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="video-to-video-with-sound" 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="#video-to-video-with-sound"><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><!---->音声付きのVideo-to-video<!----></span></h2><!--]--><!----> <p>When the checkpoint carries a <code>sound_tokenizer</code>, add <code>enable_sound=True</code> to the video-to-video call to jointly generate a synchronized audio track. The waveform is returned alongside the video and can be muxed into the MP4 with <a href="/docs/diffusers/main/ja/api/utilities#diffusers.utils.encode_video">encode_video()</a>.</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> encode_video, load_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_v2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| video=video, | |
| condition_frame_indexes_vision=[<span class="hljs-number">0</span>, <span class="hljs-number">1</span>], | |
| condition_video_keep=<span class="hljs-string">"first"</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_sound=<span class="hljs-literal">True</span>, | |
| ) | |
| encode_video( | |
| result.video, | |
| fps=<span class="hljs-number">24</span>, | |
| audio=result.sound, | |
| audio_sample_rate=pipe.sound_tokenizer.config.sampling_rate, | |
| output_path=<span class="hljs-string">"cosmos3_v2v_with_sound.mp4"</span>, | |
| )<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> encode_video, load_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_v2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| video=video, | |
| condition_frame_indexes_vision=[<span class="hljs-number">0</span>, <span class="hljs-number">1</span>], | |
| condition_video_keep=<span class="hljs-string">"first"</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_sound=<span class="hljs-literal">True</span>, | |
| ) | |
| encode_video( | |
| result.video, | |
| fps=<span class="hljs-number">24</span>, | |
| audio=result.sound, | |
| audio_sample_rate=pipe.sound_tokenizer.config.sampling_rate, | |
| output_path=<span class="hljs-string">"cosmos3_v2v_with_sound.mp4"</span>, | |
| )<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="text-to-video-with-sound" 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="#text-to-video-with-sound"><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><!---->音声付きのテキストから動画へ<!----></span></h2><!--]--><!----> <p>チェックポイントに<code>sound_tokenizer</code>が含まれている場合、<code>enable_sound=True</code>を渡すことで、同期したオーディオトラックを同時に生成できます。波形は動画とともに返され、<a href="/docs/diffusers/main/ja/api/utilities#diffusers.utils.encode_video">encode_video()</a>を使用して MP4 にマックスできます。</p> <p>これは、上記のテキストから動画への例と同じ呼び出しに<code>enable_sound=True</code>を追加したものです。</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> encode_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2v_sound_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_sound=<span class="hljs-literal">True</span>, | |
| ) | |
| encode_video( | |
| result.video, | |
| fps=<span class="hljs-number">24</span>, | |
| audio=result.sound, | |
| audio_sample_rate=pipe.sound_tokenizer.config.sampling_rate, | |
| output_path=<span class="hljs-string">"cosmos3_with_sound.mp4"</span>, | |
| )<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> encode_video | |
| <span class="hljs-comment"># JSON-upsampled positive and negative prompts (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2v_sound_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| result = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_sound=<span class="hljs-literal">True</span>, | |
| ) | |
| encode_video( | |
| result.video, | |
| fps=<span class="hljs-number">24</span>, | |
| audio=result.sound, | |
| audio_sample_rate=pipe.sound_tokenizer.config.sampling_rate, | |
| output_path=<span class="hljs-string">"cosmos3_with_sound.mp4"</span>, | |
| )<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="action-conditioned-generation" 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="#action-conditioned-generation"><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><!---->アクション条件付き生成<!----></span></h2><!--]--><!----> <p>Action runs は、各アクション固有の入力を <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.CosmosActionCondition">CosmosActionCondition</a> にまとめ、トップレベルの <code>image</code> / <code>video</code> / <code>height</code> / <code>width</code> 引数の代わりに <code>action</code> 引数として渡します。<code>resolution_tier</code> (<code>256</code>/<code>480</code>/<code>704</code>/<code>720</code>) は、入力動画のネイティブ解像度に近い値に設定してください。これにより、条件付け用のキャンバスが選択されます。Cosmos 3 は 3 つのアクションモードをサポートしています — <code>policy</code>、<code>forward_dynamics</code>、および <code>inverse_dynamics</code>。<code>policy</code> と <code>forward_dynamics</code> は最初のフレームのみを条件付けの対象とします（そのため <code>image</code> でも <code>video</code> でも動作します）。一方、<code>inverse_dynamics</code> には <code>video</code> が必要です。アクション実行用の条件付け動画は、パイプラインのトップレベルの <code>video</code> 引数ではなく、<code>action.video</code>（または <code>action.image</code>）に設定します。</p> <p><code>prompt</code> にプレーンなタスク説明を渡し、<code>action.view_point</code> でカメラを選択します（デフォルトは <code>"ego_view"</code>。また <code>"third_person_view"</code>、<code>"wrist_view"</code>、<code>"concat_view"</code> も利用可能）。このパイプラインはこれらをモデルが学習に使用した構造化 JSON キャプションに変換するため、アクションのプロンプトは LLM によるアップサンプリングを行わないでください。</p> <!--[2--><h3 class="relative group"><a id="action-policy" 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="#action-policy"><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><!---->行動ポリシー<!----></span></h3><!--]--><!----> <p>Action policy generation は、最初の観測フレーム、テキストプロンプト、およびアクションドメインのメタデータから、将来の動画トークンとアクショントークンを予測します。以下の例では、Bridge ロボットドメインを使用し、予測されたアクションチャンクをモデル正規化アクション空間で JSON に書き出します。</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">Nano</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">最高</div><!--]--></div> <div class="language-select"><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline, CosmosActionCondition | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_video | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| prompt = <span class="hljs-string">"Put the pot to the left of the purple item."</span> | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/action/bridge_20260501_0.mp4"</span> | |
| ) | |
| result = pipe( | |
| prompt=prompt, | |
| action=CosmosActionCondition( | |
| mode=<span class="hljs-string">"policy"</span>, | |
| chunk_size=<span class="hljs-number">16</span>, | |
| domain_name=<span class="hljs-string">"bridge_orig_lerobot"</span>, | |
| resolution_tier=<span class="hljs-number">480</span>, | |
| video=video, | |
| view_point=<span class="hljs-string">"ego_view"</span>, | |
| ), | |
| fps=<span class="hljs-number">5</span>, | |
| num_inference_steps=<span class="hljs-number">30</span>, | |
| guidance_scale=<span class="hljs-number">1.0</span>, | |
| use_system_prompt=<span class="hljs-literal">False</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"sample.mp4"</span>, fps=<span class="hljs-number">5</span>, macro_block_size=<span class="hljs-number">1</span>) | |
| <span class="hljs-keyword">if</span> result.action <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>: | |
| <span class="hljs-keyword">with</span> <span class="hljs-built_in">open</span>(<span class="hljs-string">"sample_action.json"</span>, <span class="hljs-string">"w"</span>) <span class="hljs-keyword">as</span> f: | |
| json.dump(result.action[<span class="hljs-number">0</span>].tolist(), f)<!----></pre></div><!----></div><!----> <div hidden=""><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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline, CosmosActionCondition | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_video | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Super"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| ) | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| prompt = <span class="hljs-string">"Put the pot to the left of the purple item."</span> | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/action/bridge_20260501_0.mp4"</span> | |
| ) | |
| result = pipe( | |
| prompt=prompt, | |
| action=CosmosActionCondition( | |
| mode=<span class="hljs-string">"policy"</span>, | |
| chunk_size=<span class="hljs-number">16</span>, | |
| domain_name=<span class="hljs-string">"bridge_orig_lerobot"</span>, | |
| resolution_tier=<span class="hljs-number">480</span>, | |
| video=video, | |
| view_point=<span class="hljs-string">"ego_view"</span>, | |
| ), | |
| fps=<span class="hljs-number">5</span>, | |
| num_inference_steps=<span class="hljs-number">30</span>, | |
| guidance_scale=<span class="hljs-number">1.0</span>, | |
| use_system_prompt=<span class="hljs-literal">False</span>, | |
| ) | |
| <span class="hljs-comment"># macro_block_size=1 allows arbitrary frame sizes (Cosmos3 outputs are not always divisible by 16).</span> | |
| export_to_video(result.video, <span class="hljs-string">"sample.mp4"</span>, fps=<span class="hljs-number">5</span>, macro_block_size=<span class="hljs-number">1</span>) | |
| <span class="hljs-keyword">if</span> result.action <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>: | |
| <span class="hljs-keyword">with</span> <span class="hljs-built_in">open</span>(<span class="hljs-string">"sample_action.json"</span>, <span class="hljs-string">"w"</span>) <span class="hljs-keyword">as</span> f: | |
| json.dump(result.action[<span class="hljs-number">0</span>].tolist(), f)<!----></pre></div><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="seacache" 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="#seacache"><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><!---->SeaCache<!----></span></h2><!--]--><!----> <p>SeaCache はデフォルトで無効になっています。Cosmos 3 では、<a href="/docs/diffusers/main/ja/api/cache#diffusers.SeaCacheConfig">SeaCacheConfig</a> を使用して明示的に有効化できます。SeaCache は、Spectral-Evolution-Aware 指標がゆっくりと変化するときに transformer の残差を再利用し、完全な transformer 実行の回数を削減します。この指標は、画像から動画への生成におけるクリーンな条件付けフレームを含む、生のビジョン潜在変数から計算されます。パイプラインからのスケジューラメタデータコールバックを使用して、transformer でこれを有効にします:</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-python "><!----><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline, SeaCacheConfig | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16, device_map=<span class="hljs-string">"cuda"</span> | |
| ) | |
| pipe.transformer.enable_cache(SeaCacheConfig(threshold=<span class="hljs-number">0.2</span>, max_consecutive_cached=<span class="hljs-number">2</span>))<!----></pre></div><!----> <p>同じモデルレベルのAPIは、<a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniPipeline">Cosmos3OmniPipeline</a>、<a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniModularPipeline">Cosmos3OmniModularPipeline</a>、および <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3DistilledModularPipeline">Cosmos3DistilledModularPipeline</a> でも動作します。SeaCache は近似であり、生成される出力が変わる可能性があります。すべてのデノイジングステップでトランスフォーマーを完全に実行する必要がある場合は、<code>pipe.transformer.disable_cache()</code> を使用して無効にしてください。キャッシュの状態は、パイプラインの呼び出しごとにリセットされ、条件付きと非条件付きのガイダンス分岐はそれぞれ独立した履歴を保持します。</p> <p>Cosmos 3 は SeaCache ゲートを反復デコーダー層の外側に配置しているため、リージョンコンパイルと互換性があります。 | |
| キャッシュを有効化した後に、各層をコンパイルします。</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-python "><!---->pipe.transformer.compile_repeated_blocks(fullgraph=<span class="hljs-literal">True</span>)<!----></pre></div><!----> <p>SeaCache は、以下に説明されている Cosmos 3 の Ulysses コンテキスト並列および DTensor ベースのテンソル並列ヘルパーもサポートしています。キャッシュの判断はランク間で同期されます。フルモデルのコンパイルやその他のモデルシャーディング戦略については、サポートを主張していません。</p> <!--[1--><h2 class="relative group"><a id="context-parallelism" 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="#context-parallelism"><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><!---->コンテキスト並列<!----></span></h2><!--]--><!----> <p>長時間の動画や高解像度の場合、単一のフォワードパスで1つのGPUのメモリおよびレイテンシの予算を超える可能性があります。Cosmos 3は<strong>コンテキスト並列性 (CP)</strong>をサポートしており、シーケンス次元を複数のGPUにシャード化して分割し、アテンションの計算を分割することで、各デバイスがトークンのスライスのみを保持するようにします。</p> <p>Cosmos 3 は <strong>Ulysses</strong> コンテキスト並列性（all-to-all によるシーケンス/ヘッドの交換）をサポートしています。リングアテンションはサポートされていません。</p> <p>ほとんどの diffusers モデルとは異なり、Cosmos 3 は CP を transformer や宣言的な <code>enable_parallelism()</code> パスに<strong>組み込んでいません</strong>。そのグループクエリアテンション、独立した理解/生成ストリーム（生成ストリームは両方にアテンションを向ける）、ストリームごとの不揃いな長さは、<code>_cp_plan</code> として表現できないためです。代わりに、モデルは小さな no-op の shard/gather シームを公開し、実装は <a href="https://github.com/huggingface/diffusers/blob/main/examples/cosmos3/cosmos_parallel.py" rel="nofollow"><code>examples/cosmos3/cosmos_parallel.py</code></a> にあります。これは、最初から最後まで読み、適応できる自己完結型モジュールです。2つの直交し、組み合わせ可能なシャーディング軸を提供します：</p> <table><thead><tr><th>ヘルパー</th><th>シャード</th><th>用途</th></tr></thead><tbody><tr><td><code>enable_cosmos3_context_parallel(transformer, cp_mesh)</code></td><td>シーケンス(CP / Ulysses)</td><td>1つのGPUに収まるモデル(<code>Nano</code>)のレイテンシ</td></tr><tr><td><code>enable_cosmos3_tensor_parallel(transformer, tp_mesh)</code></td><td>重み (TP)</td><td>1つのGPUに収まらないモデルの学習(<code>Super</code>)</td></tr></tbody></table> <p>2次元の<code>(tp, cp)</code>メッシュ(<a href="#fitting-large-models-with-tensor-parallelism">テンソル並列による大規模モデルの適合</a>を参照)で、どちらか一方を単独で使用するか、両方を組み合わせて使用できます。</p> <p>Cosmos 3 に固有の要件は2つあります:</p> <ul><li><code>native</code> アテンションバックエンドを使用します。Cosmos 3 はグループクエリアテンション (GQA) を使用しており、<code>enable_gqa</code> を受け入れるのはネイティブ SDPA バックエンドのみです (cuDNN と flash はこれを拒否します)。ヘルパーは KV ヘッドをクエリヘッド数に展開し、<code>enable_gqa=False</code> で SDPA を呼び出すことで、引き続き flash カーネルにディスパッチされます (math フォールバックでは完全な <code>[S, S]</code> スコアがマテリアライズされ、長いシーケンスで OOM になります)。</li> <li>CP (Ulysses) の度合いは、クエリヘッド数（<code>Nano</code> では 32、<code>Super</code> では 64）で割り切れる必要があります。TP の場合、度合いは KV ヘッド数（8）で割り切れる必要があります。理解（テキスト）と生成（動画/音声）のストリームは、シーケンス方向に独立してシャード化され、不揃いな長さは内部でワールドサイズの倍数にゼロパディングされます。</li></ul> <!--[2--><h3 class="relative group"><a id="run-it" 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="#run-it"><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><!---->実行する<!----></span></h3><!--]--><!----> <p>完全な CLI <a href="https://github.com/huggingface/diffusers/blob/main/examples/cosmos3/inference_cosmos3.py" rel="nofollow"><code>examples/cosmos3/inference_cosmos3.py</code></a> は <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniModularPipeline">Cosmos3OmniModularPipeline</a> を使用し、これらのヘルパーを再利用するため、<strong>任意のモダリティ</strong>（テキストから画像/動画へ、画像から動画へ、サウンド、アクションモード）が <code>--tp-degree</code> / <code>--cp-degree</code> を通じてマルチ GPU で実行できます。<a href="https://docs.pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a> で起動してください。<code>--tp-degree * --cp-degree</code> は <code>--nproc_per_node</code> と等しくなければなりません。すべてのランクが同じ出力を生成し、ランク 0 がそれを書き出します。</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-comment"># CP only — Nano (fits one GPU); CP degree must divide 32 query heads.</span> | |
| torchrun --nproc_per_node=4 examples/cosmos3/inference_cosmos3.py --model nano --cp-degree 4 --prompt <span class="hljs-string">"..."</span> | |
| <span class="hljs-comment"># TP only — Super; TP degree must divide 64 query heads and 8 KV heads.</span> | |
| torchrun --nproc_per_node=4 examples/cosmos3/inference_cosmos3.py --model super --tp-degree 4 --prompt <span class="hljs-string">"..."</span> | |
| <span class="hljs-comment"># TP + CP — Super, with sound (TP=2 x CP=2 across 4 GPUs).</span> | |
| torchrun --nproc_per_node=4 examples/cosmos3/inference_cosmos3.py \ | |
| --model super --tp-degree 2 --cp-degree 2 --enable-sound --prompt <span class="hljs-string">"..."</span><!----></pre></div><!----> <p><code>Super</code>の重みは約120 GBであり、1つの96 GB GPUには収まらないため、TPが必要です。<code>Nano</code>は1つのGPUに収まるため、CPは純粋なレイテンシ最適化です。（両方のフラグを省略すると、シングルGPUで実行されます。）</p> <!--[2--><h3 class="relative group"><a id="fitting-large-models-with-tensor-parallelism" 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="#fitting-large-models-with-tensor-parallelism"><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><!---->テンソル並列性を用いた大規模モデルの学習<!----></span></h3><!--]--><!----> <p>CP は <em>アクティベーション</em> をシャード化しますが、すべてのランクにすべての重みを複製するため、モデルの重みフットプリントを削減することはありません。1 つの GPU に収まらないモデルは、CP だけでは依然として収まりません。<strong>重み</strong> をシャード化するには、<code>enable_cosmos3_tensor_parallel(transformer, tp_mesh)</code> が、第 2 の直交メッシュ軸上で Megatron スタイルのテンソル並列性を適用します:</p> <ul><li>Attention と MLP の射影は TP グループ内でカラム/行シャード化されています（<code>to_q/to_k/to_v</code> と <code>add_q/k/v</code>、および MLP の <code>gate/up</code> はカラム並列、<code>to_out/to_add_out</code> と MLP の <code>down</code> は all-reduce を伴う行並列です）。各ランクは <code>query_heads / tp</code> 個のクエリヘッドと <code>kv_heads / tp</code> 個の KV ヘッドを保有することになります。</li> <li>TP は 2 次元の <code>(tp, cp)</code> デバイスメッシュ上で CP と組み合わせられます。TP はヘッド/重みを永続的に分割し、CP はその上にシーケンスをシャードします。制約は、<code>tp</code> が KV ヘッド数（8）を割り、<code>tp * cp</code> がクエリヘッド数（<code>Nano</code> では 32、<code>Super</code> では 64）を割りることです。</li> <li>重みはCPUに読み込まれ、レイヤーごとにGPUにシャード化されるため、完全なモデルが単一のデバイス上にマテリアライズされることはありません。</li></ul> <blockquote class="tip"><p>TP はすべての attention ブロックと MLP ブロックで all-reduce を発行するため、帯域幅の負荷が重くなります。NVLink を備えていないホストでは、これが支配的なコストとなります。重みが収まる最小の TP 度を優先し、残りの GPU は CP に割り当ててください。</p></blockquote> <!--[2--><h3 class="relative group"><a id="use-it-in-your-own-modular-pipeline" 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="#use-it-in-your-own-modular-pipeline"><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><!---->独自のモジュール化パイプラインで使用する<!----></span></h3><!--]--><!----> <p>CLIフラグは便利ですが、<a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniModularPipeline">Cosmos3OmniModularPipeline</a>でヘルパーを直接呼び出すこともできます。パイプラインの設定とコンポーネントをCPUに読み込み、パイプラインをランクローカルGPUに移動する<em>前に</em>TPを適用し、<code>native</code>バックエンドに切り替えた後、CPを有効にします。このフローでは<code>device_map</code>を使用しないでください。</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-python "><!----><span class="hljs-keyword">import</span> os | |
| <span class="hljs-keyword">import</span> sys | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">import</span> torch.distributed <span class="hljs-keyword">as</span> dist | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniModularPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> torch.distributed.device_mesh <span class="hljs-keyword">import</span> init_device_mesh | |
| <span class="hljs-comment"># Make the helper module importable.</span> | |
| sys.path.insert(<span class="hljs-number">0</span>, <span class="hljs-string">"examples/cosmos3"</span>) | |
| <span class="hljs-keyword">from</span> cosmos_parallel <span class="hljs-keyword">import</span> ( | |
| enable_cosmos3_context_parallel, | |
| enable_cosmos3_flash_attention, | |
| enable_cosmos3_tensor_parallel, | |
| ) | |
| <span class="hljs-comment"># torchrun sets RANK / WORLD_SIZE / LOCAL_RANK. Pick tp_degree * cp_degree == world size.</span> | |
| local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">"LOCAL_RANK"</span>]) | |
| torch.cuda.set_device(local_rank) | |
| dist.init_process_group(<span class="hljs-string">"nccl"</span>) | |
| mesh = init_device_mesh(<span class="hljs-string">"cuda"</span>, (tp_degree, cp_degree), mesh_dim_names=(<span class="hljs-string">"tp"</span>, <span class="hljs-string">"cp"</span>)) | |
| <span class="hljs-comment"># Load components on CPU first; a TP-sharded model may not fit one GPU.</span> | |
| pipe = Cosmos3OmniModularPipeline.from_pretrained(model_id) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.enable_safety_checker() | |
| <span class="hljs-keyword">if</span> tp_degree > <span class="hljs-number">1</span>: | |
| enable_cosmos3_tensor_parallel(pipe.transformer, mesh[<span class="hljs-string">"tp"</span>]) <span class="hljs-comment"># shard weights -> GPUs</span> | |
| pipe.to(<span class="hljs-string">f"cuda:<span class="hljs-subst">{local_rank}</span>"</span>) <span class="hljs-comment"># move the replicated remainder</span> | |
| pipe.transformer.set_attention_backend(<span class="hljs-string">"native"</span>) | |
| <span class="hljs-keyword">if</span> cp_degree > <span class="hljs-number">1</span>: | |
| enable_cosmos3_context_parallel(pipe.transformer, mesh[<span class="hljs-string">"cp"</span>]) <span class="hljs-comment"># shard the sequence</span> | |
| <span class="hljs-keyword">elif</span> tp_degree > <span class="hljs-number">1</span>: | |
| enable_cosmos3_flash_attention(pipe.transformer) <span class="hljs-comment"># GQA-safe dense attention</span> | |
| <span class="hljs-comment"># Modular pipelines replace components through update_components().</span> | |
| scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| pipe.update_components(scheduler=scheduler) | |
| <span class="hljs-comment"># A single output name returns that value; a list returns a dictionary.</span> | |
| outputs = pipe( | |
| prompt=<span class="hljs-string">'{"scene":"A robot arm in a kitchen"}'</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| output=[<span class="hljs-string">"videos"</span>, <span class="hljs-string">"sound"</span>, <span class="hljs-string">"sampling_rate"</span>, <span class="hljs-string">"action"</span>], | |
| ) | |
| videos = outputs[<span class="hljs-string">"videos"</span>] | |
| sound = outputs[<span class="hljs-string">"sound"</span>] <span class="hljs-comment"># None unless sound generation was requested.</span> | |
| action = outputs[<span class="hljs-string">"action"</span>] <span class="hljs-comment"># None unless an action workflow produced actions.</span><!----></pre></div><!----> <p>CP のみ（ウェイトシャーディングなし）の場合は、1 次元メッシュを使用します: <code>init_device_mesh("cuda", (world_size,), mesh_dim_names=("cp",))</code> を呼び出し、<code>enable_cosmos3_context_parallel</code> を有効にするだけです。</p> <p><code>enable_safety_checker()</code> はデフォルトのチェッカーを読み込んで有効化し、<code>disable_safety_checker()</code> は明示的に無効化します。タスクパイプラインの構築引数である <code>enable_safety_checker=</code> や呼び出し引数である <code>enable_safety_check=</code> の代わりに、これらのパイプラインメソッドを使用してください。モジュラーパイプラインは <code>Cosmos3OmniPipelineOutput</code> を返しません。フレームのみを必要とする場合は <code>output="videos"</code> を使用するか、上記に示したように出力リストと返された辞書を使用し、<code>result.video</code>、<code>result.sound</code>、または <code>result.action</code> を使用しないでください。</p> <blockquote class="tip"><p>On some multi-GPU topologies the first NCCL all-to-all can hang. If a CP run stalls at the start of the first denoising step, set <code>NCCL_P2P_DISABLE=1</code> in the environment before launching <code>torchrun</code>.</p></blockquote> <p>CP と TP は、上記のすべてのワークフロー（テキストから動画へ、画像から動画へ、音声付きのテキストから動画へ、アクション条件付き生成）および <code>Nano</code> と <code>Super</code> の両方のチェックポイントと組み合わせることができます。変更されるのは、パイプラインの構築と並列化の設定を行う行のみです。</p> <!--[1--><h2 class="relative group"><a id="metadata-templates" 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="#metadata-templates"><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><!---->メタデータテンプレート<!----></span></h2><!--]--><!----> <p><code>tokenize_prompt</code> appends short metadata sentences inside the user message so the LLM sees the conditioning the model was trained with. The positive prompt gets sentences like <em>“The video is 7.9 seconds long and is of 24 FPS.”</em> and <em>“This video is of 720x1280 resolution.”</em>; the negative prompt gets the inverse (<em>”… is not …”</em>).</p> <p>どちらもデフォルトで有効になっています。<code>__call__</code> を通じて、いずれかのペアを無効にできます。</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-python "><!---->result = pipe( | |
| prompt=prompt, | |
| negative_prompt=negative_prompt, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| add_duration_template=<span class="hljs-literal">False</span>, <span class="hljs-comment"># skip the duration sentence on both prompts</span> | |
| add_resolution_template=<span class="hljs-literal">False</span>, <span class="hljs-comment"># skip the resolution sentence on both prompts</span> | |
| )<!----></pre></div><!----> <p><code>add_duration_template</code> は、<code>num_frames == 1</code>（画像モード）の場合には効果がなく、その場合は解像度の文のみが追加されます。</p> <!--[1--><h2 class="relative group"><a id="safety-checker" 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="#safety-checker"><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><!---->セーフティチェッカー<!----></span></h2><!--]--><!----> <p>Cosmos3 は <a href="https://pypi.org/project/cosmos-guardrail/" rel="nofollow"><code>cosmos_guardrail</code></a> の <code>CosmosSafetyChecker</code> を組み込み、<strong>デフォルトで</strong>実行します。テキストガードレールは生成前に不適切なプロンプトを拒否します（<code>ValueError</code>）。動画ガードレールはデコードされたフレームに対して実行され、検出された顔をピクセル化するか、出力を拒否します。音声出力にはガードレールは適用されません。</p> <p>デフォルトのチェッカーを有効にするために、オプションの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=" "><!---->pip <span class="hljs-keyword">install</span> cosmos_guardrail<!----></pre></div><!----> <p>NVIDIA Open Model License Agreement では、チェッカーの使用が必須です。以下の2つのフラグは、ガードレールが冗長になるテストや開発ワークフロー（例：入力がすでにクリア済みである場合、または意図的にエッジケースの入力に対してパイプラインを実行している場合）のために用意されています。</p> <p><strong>構築時に無効化</strong>（チェッカーがインスタンス化されないため、ガードレールモデルはダウンロードされず、メモリにも読み込まれません）：</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-python "><!----><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniPipeline | |
| pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, | |
| dtype=torch.bfloat16, | |
| device_map=<span class="hljs-string">"cuda"</span>, <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| enable_safety_checker=<span class="hljs-literal">False</span>, | |
| )<!----></pre></div><!----> <p><strong>単一の呼び出しに対して無効にする</strong>（チェッカーは読み込み状態を維持します。デフォルトを有効にしたまま、一時的にバイパスする場合に便利です）：</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-python "><!---->result = pipe( | |
| prompt=prompt, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_safety_check=<span class="hljs-literal">False</span>, | |
| )<!----></pre></div><!----> <p>カスタムチェッカー(例:高速テスト用の no-op サブクラス)を指定するには、<code>safety_checker=</code>として渡します:</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-python "><!---->pipe = Cosmos3OmniPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, | |
| dtype=torch.bfloat16, | |
| device_map=<span class="hljs-string">"cuda"</span>, <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| safety_checker=MyCustomSafetyChecker(), | |
| )<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="diffusers.Cosmos3OmniPipeline" 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="#diffusers.Cosmos3OmniPipeline"><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><!---->Cosmos3OmniPipeline<!----></span></h2><!--]--><!----> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.Cosmos3OmniPipeline"><!----><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">class</span> <span class="font-medium">diffusers.</span><span class="font-semibold">Cosmos3OmniPipeline</span></span></h3><!----> <a id="diffusers.Cosmos3OmniPipeline" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.Cosmos3OmniPipeline"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L413" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">transformer<span class="opacity-60">: Cosmos3OmniTransformer</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">text_tokenizer<span class="opacity-60">: AutoTokenizer</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vae<span class="opacity-60">: AutoencoderKLWan</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">scheduler<span class="opacity-60">: UniPCMultistepScheduler</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">sound_tokenizer<span class="opacity-60">: diffusers.models.autoencoders.autoencoder_cosmos3_audio.Cosmos3AVAEAudioTokenizer | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">safety_checker<span class="opacity-60">: diffusers.pipelines.cosmos.pipeline_cosmos3_omni.CosmosSafetyChecker | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">enable_safety_checker<span class="opacity-60">: bool = True</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">default_use_system_prompt<span class="opacity-60">: bool = True</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_native_flow_schedule<span class="opacity-60">: bool = False</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.Cosmos3OmniPipeline.decode_sound"><!----><h4 class="!m-0"><span class="flex-1 rounded-xl py-0.5 break-all bg-gradient-to-r from-blue-50/60 to-white dark:from-gray-900 dark:to-gray-950 text-blue-700 dark:text-blue-300 font-medium px-2"><svg width="1em" height="1em" viewBox="0 0 32 33" class="mr-1 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg"><path d="M5.80566 18.3545C4.90766 17.4565 4.90766 16.0005 5.80566 15.1025L14.3768 6.53142C15.2748 5.63342 16.7307 5.63342 17.6287 6.53142L26.1999 15.1025C27.0979 16.0005 27.0979 17.4565 26.1999 18.3545L17.6287 26.9256C16.7307 27.8236 15.2748 27.8236 14.3768 26.9256L5.80566 18.3545Z" fill="currentColor" fill-opacity="0.25"/><path fill-rule="evenodd" clip-rule="evenodd" d="M16.4801 13.9619C16.4801 12.9761 16.7467 12.5436 16.9443 12.3296C17.1764 12.078 17.5731 11.8517 18.2275 11.707C18.8821 11.5623 19.638 11.5342 20.4038 11.5582C20.7804 11.57 21.1341 11.5932 21.4719 11.6156L21.5263 11.6193C21.8195 11.6389 22.1626 11.6618 22.4429 11.6618V7.40825C22.3209 7.40825 22.1219 7.39596 21.7544 7.37149C21.4202 7.34925 20.9976 7.32115 20.5371 7.30672C19.6286 7.27824 18.4672 7.29779 17.3093 7.55377C16.1512 7.8098 14.8404 8.33724 13.8181 9.4452C12.7612 10.5907 12.2266 12.1236 12.2266 13.9619V15.0127H10.6836V19.2662H12.2266V26.6332H16.4801V19.2662H20.3394V15.0127H16.4801V13.9619Z" fill="currentColor"/></svg>decode_sound</span></h4><!----> <a id="diffusers.Cosmos3OmniPipeline.decode_sound" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.Cosmos3OmniPipeline.decode_sound"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L517" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">latent<span class="opacity-60">: Tensor</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>音声潜在変数<code>[C, T]</code>を波形<code>[audio_ch, N]</code>にデコードします。</p> <p>サウンドトークナイザのデコーダが期待するバッチ次元を追加/削除します。</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.Cosmos3OmniPipeline.prepare_latents"><!----><h4 class="!m-0"><span class="flex-1 rounded-xl py-0.5 break-all bg-gradient-to-r from-blue-50/60 to-white dark:from-gray-900 dark:to-gray-950 text-blue-700 dark:text-blue-300 font-medium px-2"><svg width="1em" height="1em" viewBox="0 0 32 33" class="mr-1 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg"><path d="M5.80566 18.3545C4.90766 17.4565 4.90766 16.0005 5.80566 15.1025L14.3768 6.53142C15.2748 5.63342 16.7307 5.63342 17.6287 6.53142L26.1999 15.1025C27.0979 16.0005 27.0979 17.4565 26.1999 18.3545L17.6287 26.9256C16.7307 27.8236 15.2748 27.8236 14.3768 26.9256L5.80566 18.3545Z" fill="currentColor" fill-opacity="0.25"/><path fill-rule="evenodd" clip-rule="evenodd" d="M16.4801 13.9619C16.4801 12.9761 16.7467 12.5436 16.9443 12.3296C17.1764 12.078 17.5731 11.8517 18.2275 11.707C18.8821 11.5623 19.638 11.5342 20.4038 11.5582C20.7804 11.57 21.1341 11.5932 21.4719 11.6156L21.5263 11.6193C21.8195 11.6389 22.1626 11.6618 22.4429 11.6618V7.40825C22.3209 7.40825 22.1219 7.39596 21.7544 7.37149C21.4202 7.34925 20.9976 7.32115 20.5371 7.30672C19.6286 7.27824 18.4672 7.29779 17.3093 7.55377C16.1512 7.8098 14.8404 8.33724 13.8181 9.4452C12.7612 10.5907 12.2266 12.1236 12.2266 13.9619V15.0127H10.6836V19.2662H12.2266V26.6332H16.4801V19.2662H20.3394V15.0127H16.4801V13.9619Z" fill="currentColor"/></svg>prepare_latents</span></h4><!----> <a id="diffusers.Cosmos3OmniPipeline.prepare_latents" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.Cosmos3OmniPipeline.prepare_latents"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L763" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">image<span class="opacity-60">: typing.Union[PIL.Image.Image, numpy.ndarray, torch.Tensor, NoneType] = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">video<span class="opacity-60">: typing.Union[list[PIL.Image.Image], torch.Tensor, numpy.ndarray, NoneType] = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">condition_frame_indexes_vision<span class="opacity-60">: collections.abc.Iterable[int] = (0, 1)</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">condition_video_keep<span class="opacity-60">: typing.Literal['first', 'last'] = 'first'</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">num_frames<span class="opacity-60">: int | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">height<span class="opacity-60">: int | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">width<span class="opacity-60">: int | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fps<span class="opacity-60">: float = 24.0</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">latents<span class="opacity-60">: typing.Optional[torch.Tensor] = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">sound_latents<span class="opacity-60">: typing.Optional[torch.Tensor] = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">action_latents<span class="opacity-60">: typing.Optional[torch.Tensor] = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">generator<span class="opacity-60">: typing.Optional[torch.Generator] = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">device<span class="opacity-60">: str = 'cuda'</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">dtype<span class="opacity-60">: dtype = torch.bfloat16</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">enable_sound<span class="opacity-60">: bool = False</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">action<span class="opacity-60">: CosmosActionCondition | None = None</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>単一サンプル用の条件付けと初期ノイズを構築します。</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.Cosmos3OmniPipeline.tokenize_prompt"><!----><h4 class="!m-0"><span class="flex-1 rounded-xl py-0.5 break-all bg-gradient-to-r from-blue-50/60 to-white dark:from-gray-900 dark:to-gray-950 text-blue-700 dark:text-blue-300 font-medium px-2"><svg width="1em" height="1em" viewBox="0 0 32 33" class="mr-1 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg"><path d="M5.80566 18.3545C4.90766 17.4565 4.90766 16.0005 5.80566 15.1025L14.3768 6.53142C15.2748 5.63342 16.7307 5.63342 17.6287 6.53142L26.1999 15.1025C27.0979 16.0005 27.0979 17.4565 26.1999 18.3545L17.6287 26.9256C16.7307 27.8236 15.2748 27.8236 14.3768 26.9256L5.80566 18.3545Z" fill="currentColor" fill-opacity="0.25"/><path fill-rule="evenodd" clip-rule="evenodd" d="M16.4801 13.9619C16.4801 12.9761 16.7467 12.5436 16.9443 12.3296C17.1764 12.078 17.5731 11.8517 18.2275 11.707C18.8821 11.5623 19.638 11.5342 20.4038 11.5582C20.7804 11.57 21.1341 11.5932 21.4719 11.6156L21.5263 11.6193C21.8195 11.6389 22.1626 11.6618 22.4429 11.6618V7.40825C22.3209 7.40825 22.1219 7.39596 21.7544 7.37149C21.4202 7.34925 20.9976 7.32115 20.5371 7.30672C19.6286 7.27824 18.4672 7.29779 17.3093 7.55377C16.1512 7.8098 14.8404 8.33724 13.8181 9.4452C12.7612 10.5907 12.2266 12.1236 12.2266 13.9619V15.0127H10.6836V19.2662H12.2266V26.6332H16.4801V19.2662H20.3394V15.0127H16.4801V13.9619Z" fill="currentColor"/></svg>tokenize_prompt</span></h4><!----> <a id="diffusers.Cosmos3OmniPipeline.tokenize_prompt" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.Cosmos3OmniPipeline.tokenize_prompt"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L1132" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">prompt<span class="opacity-60">: str</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">negative_prompt<span class="opacity-60">: str | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">num_frames<span class="opacity-60">: int = 189</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">height<span class="opacity-60">: int = 720</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">width<span class="opacity-60">: int = 1280</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fps<span class="opacity-60">: float = 24.0</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_system_prompt<span class="opacity-60">: bool | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">add_resolution_template<span class="opacity-60">: bool = True</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">add_duration_template<span class="opacity-60">: bool = True</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">action_mode<span class="opacity-60">: str | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">action_view_point<span class="opacity-60">: str | None = None</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>設定済みのチャットテンプレートを使用して、プロンプト拡張テンプレートを適用し、cond/uncondプロンプトをトークン化します。</p> <p>このパイプラインでは、個別のテキストエンコーダーを実行しません。ジョイントのCosmos3トランスフォーマーは、ビジョン（および任意でサウンド）トークンと並んで、生のトークンIDを直接消費します。</p> <p><code>negative_prompt</code> が <code>None</code> の場合、空の文字列が使用されます。Cosmos3 のドキュメントページでは、text2video / image2video に明示的に渡すことを推奨する品質管理用のネガティブプロンプトが記載されています。有効な場合、duration と resolution のテンプレートはプロンプトに、逆のテンプレートはネガティブプロンプトに付加されます。</p> <p><code>action_mode</code>が設定されている場合、プロンプトはフラットなメタデータテンプレートの代わりに、モデルが学習に使用した構造化アクションJSONキャプションに変換されます（<code>_build_action_json_prompt</code>を参照）。この際、フレーミングフィールドには<code>action_view_point</code>が使用されます。JSONにはduration/fps/resolution/aspect_ratioが既に含まれているため、フラットなメタデータテンプレートはスキップされます。</p></div></div> <ul><li>すべて</li> <li><strong>call</strong></li></ul> <!--[1--><h2 class="relative group"><a id="cosmos3omnimodularpipeline" 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="#cosmos3omnimodularpipeline"><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><!---->Cosmos3OmniModularPipeline<!----></span></h2><!--]--><!----> <p>Cosmos 3 は、Modular Diffusers パイプラインとしても利用可能です。タスクベースの <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniPipeline">Cosmos3OmniPipeline</a> は引き続き利用可能であり、モジュラーパイプラインはこれと併存し、同じモード（<code>text2image</code>、<code>text2video</code>、<code>image2video</code>、<code>video2video</code>、アクション条件付き生成、および <code>transfer</code>（構造制御）、チェックポイントがサポートしている場合はオプションのサウンド）をカバーします。</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-python "><!----><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniModularPipeline | |
| pipe = Cosmos3OmniModularPipeline.from_pretrained( | |
| <span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16 | |
| ) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.enable_safety_checker() | |
| videos = pipe( | |
| prompt=<span class="hljs-string">'{"scene":"A robot arm in a kitchen"}'</span>, | |
| num_frames=<span class="hljs-number">1</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| <span class="hljs-comment"># Modular pipelines expose declared outputs directly instead of using the task pipeline's</span> | |
| <span class="hljs-comment"># `return_dict`/`Cosmos3OmniPipelineOutput` API.</span> | |
| image = videos[<span class="hljs-number">0</span>]<!----></pre></div><!----> <p><a href="/docs/diffusers/main/ja/api/modular_diffusers/pipeline#diffusers.ModularPipeline">ModularPipeline</a> を通して読み込み、リポジトリの構成でブロッククラスを選択させることもできます。</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-python "><!----><span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> ModularPipeline | |
| pipe = ModularPipeline.from_pretrained(<span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.enable_safety_checker() | |
| videos = pipe( | |
| prompt=<span class="hljs-string">'{"scene":"A robot arm in a kitchen"}'</span>, num_frames=<span class="hljs-number">1</span>, height=<span class="hljs-number">720</span>, width=<span class="hljs-number">1280</span>, output=<span class="hljs-string">"videos"</span> | |
| )<!----></pre></div><!----> <p>特定の Cosmos モジュラーワークフローを検査またはカスタマイズするには、<code>available_workflows</code> + <code>get_workflow()</code> を使用します:</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-python "><!---->available = pipe.blocks.available_workflows | |
| image2video_blocks = pipe.blocks.get_workflow(<span class="hljs-string">"image2video"</span>)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="modular-examples-for-all-existing-workflows" 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="#modular-examples-for-all-existing-workflows"><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><!---->既存のすべてのワークフロー向けのモジュール化された例<!----></span></h3><!--]--><!----> <p>モジュラーパイプラインは、タスクパイプラインと同じ呼び出しシグネチャをサポートしています。以下のスニペットは、上記で示されたすべての生成例（<code>text2video</code>、<code>text2image</code>、<code>image2video</code>、<code>video2video</code>、<code>video2video_sound</code>、<code>text2video_sound</code>、および<code>action_policy</code>）を反映しています。Transfer（構造制御）は独自の入力を持ち、以下の<a href="#modular-transfer-structural-control">Modular transfer</a>で個別に示されています。</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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniModularPipeline, CosmosActionCondition | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> encode_video, export_to_video, load_image, load_video | |
| pipe = Cosmos3OmniModularPipeline.from_pretrained(<span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.enable_safety_checker() | |
| pipe.to(<span class="hljs-string">"cuda"</span>) <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| <span class="hljs-comment"># text2video</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| videos = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">6.0</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| export_to_video(videos, <span class="hljs-string">"cosmos3_modular_t2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>) | |
| <span class="hljs-comment"># text2image</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2i_prompt.json"</span>)) | |
| videos = pipe(prompt=json.dumps(json_prompt), num_frames=<span class="hljs-number">1</span>, height=<span class="hljs-number">720</span>, width=<span class="hljs-number">1280</span>, output=<span class="hljs-string">"videos"</span>) | |
| videos[<span class="hljs-number">0</span>].save(<span class="hljs-string">"cosmos3_modular_t2i.jpg"</span>, <span class="hljs-built_in">format</span>=<span class="hljs-string">"JPEG"</span>, quality=<span class="hljs-number">85</span>) | |
| <span class="hljs-comment"># image2video</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_i2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| image = load_image(<span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/releases/download/assets/robot_153.jpg"</span>) | |
| videos = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| image=image, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| export_to_video(videos, <span class="hljs-string">"cosmos3_modular_i2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>) | |
| <span class="hljs-comment"># video2video</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_v2v_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt_i2v.json"</span>)) | |
| video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4"</span> | |
| ) | |
| videos = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| video=video, | |
| condition_frame_indexes_vision=[<span class="hljs-number">0</span>, <span class="hljs-number">1</span>], | |
| condition_video_keep=<span class="hljs-string">"first"</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">6.0</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| export_to_video(videos, <span class="hljs-string">"cosmos3_modular_v2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>) | |
| <span class="hljs-comment"># video2video_sound</span> | |
| outputs = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| video=video, | |
| condition_frame_indexes_vision=[<span class="hljs-number">0</span>, <span class="hljs-number">1</span>], | |
| condition_video_keep=<span class="hljs-string">"first"</span>, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_sound=<span class="hljs-literal">True</span>, | |
| output=[<span class="hljs-string">"videos"</span>, <span class="hljs-string">"sound"</span>, <span class="hljs-string">"sampling_rate"</span>], | |
| ) | |
| encode_video( | |
| outputs[<span class="hljs-string">"videos"</span>], | |
| fps=<span class="hljs-number">24</span>, | |
| audio=outputs[<span class="hljs-string">"sound"</span>], | |
| audio_sample_rate=outputs[<span class="hljs-string">"sampling_rate"</span>], | |
| output_path=<span class="hljs-string">"cosmos3_modular_v2v_with_sound.mp4"</span>, | |
| ) | |
| <span class="hljs-comment"># text2video_sound</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2v_sound_prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| outputs = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">24.0</span>, | |
| enable_sound=<span class="hljs-literal">True</span>, | |
| output=[<span class="hljs-string">"videos"</span>, <span class="hljs-string">"sound"</span>, <span class="hljs-string">"sampling_rate"</span>], | |
| ) | |
| encode_video( | |
| outputs[<span class="hljs-string">"videos"</span>], | |
| fps=<span class="hljs-number">24</span>, | |
| audio=outputs[<span class="hljs-string">"sound"</span>], | |
| audio_sample_rate=outputs[<span class="hljs-string">"sampling_rate"</span>], | |
| output_path=<span class="hljs-string">"cosmos3_modular_t2v_with_sound.mp4"</span>, | |
| ) | |
| <span class="hljs-comment"># action_policy</span> | |
| prompt = <span class="hljs-string">"Put the pot to the left of the purple item."</span> | |
| action_video = load_video( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/action/bridge_20260501_0.mp4"</span> | |
| ) | |
| outputs = pipe( | |
| prompt=prompt, | |
| action=CosmosActionCondition( | |
| mode=<span class="hljs-string">"policy"</span>, | |
| chunk_size=<span class="hljs-number">16</span>, | |
| domain_name=<span class="hljs-string">"bridge_orig_lerobot"</span>, | |
| resolution_tier=<span class="hljs-number">480</span>, | |
| video=action_video, | |
| view_point=<span class="hljs-string">"ego_view"</span>, | |
| ), | |
| fps=<span class="hljs-number">5</span>, | |
| num_inference_steps=<span class="hljs-number">30</span>, | |
| guidance_scale=<span class="hljs-number">1.0</span>, | |
| use_system_prompt=<span class="hljs-literal">False</span>, | |
| output=[<span class="hljs-string">"videos"</span>, <span class="hljs-string">"action"</span>], | |
| ) | |
| export_to_video(outputs[<span class="hljs-string">"videos"</span>], <span class="hljs-string">"cosmos3_modular_action_policy.mp4"</span>, fps=<span class="hljs-number">5</span>, macro_block_size=<span class="hljs-number">1</span>) | |
| <span class="hljs-keyword">if</span> outputs[<span class="hljs-string">"action"</span>] <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>: | |
| <span class="hljs-keyword">with</span> <span class="hljs-built_in">open</span>(<span class="hljs-string">"cosmos3_modular_action_policy.json"</span>, <span class="hljs-string">"w"</span>) <span class="hljs-keyword">as</span> f: | |
| json.dump(outputs[<span class="hljs-string">"action"</span>][<span class="hljs-number">0</span>].tolist(), f)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="modular-transfer-structural-control" 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="#modular-transfer-structural-control"><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><!---->モジュール転送（構造制御）<!----></span></h3><!--]--><!----> <p>Transfer は、<code>control_videos=</code> に <code>{hint: video}</code> マッピングとして渡される、<strong>事前計算された制御動画</strong>（エッジ、ぼかし、深度、セグメンテーション、またはワールドシナリオマップ）に従って行われます。これは動画専用であり、<code>image</code> / <code>video</code> / <code>action</code> / <code>enable_sound</code> はサポートされていません。プロンプトは事前アップサンプリング済みの JSON キャプションです（<a href="#prompt-upsampling">Prompt upsampling</a> を参照）。また、長いクリップは <code>num_video_frames_per_chunk</code> フレームごとのチャンクで自己回帰的に生成され、自動的に結合されます。<code>guidance_scale</code> は通常のテキスト CFG です。<code>control_guidance</code>（<code>!= 1.0</code>）は、制御シグナルをさらに増幅します。各ヒントごとの推奨初期値は以下の通りです。</p> <table><thead><tr><th>ヒント</th><th><code>guidance_scale</code></th><th><code>control_guidance</code></th><th><code>flow_shift</code></th><th>幾何学</th></tr></thead><tbody><tr><td>エッジ / ブラー / 深度</td><td>3.0</td><td>1.5</td><td>10.0</td><td>30 FPSで121フレーム</td></tr><tr><td>セグメンテーション</td><td>3.0</td><td>2.0</td><td>10.0</td><td>30 FPSで121フレーム</td></tr><tr><td>ワールドシナリオ(WSM)</td><td>1.0</td><td>3.0</td><td>10.0</td><td>101フレーム @ 10 FPS</td></tr></tbody></table> <p>Diffusersにはコントロールアセットは同梱されていません。完成済みのアセット（各ヒントに対応するコントロール動画と<code>prompt.json</code>、および共有の<code>negative_prompt.json</code>）は、<a href="https://github.com/NVIDIA/cosmos/tree/6e2c69afafd6444c1ca14a021aba93fcc6b7996a/cookbooks/cosmos3/generator/transfer/assets" rel="nofollow">Cosmos cookbook</a>に格納されています。以下のエッジの例では、これらをローカルの<code>assets/</code>フォルダにダウンロードしてください。</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 "><!---->base=https://github.com/NVIDIA/cosmos/raw/6e2c69afafd6444c1ca14a021aba93fcc6b7996a/cookbooks/cosmos3/generator/transfer/assets | |
| <span class="hljs-built_in">mkdir</span> -p assets/edge | |
| curl -sL <span class="hljs-string">"<span class="hljs-variable">$base</span>/edge/control_edge.mp4"</span> -o assets/edge/control_edge.mp4 | |
| curl -sL <span class="hljs-string">"<span class="hljs-variable">$base</span>/edge/prompt.json"</span> -o assets/edge/prompt.json | |
| curl -sL <span class="hljs-string">"<span class="hljs-variable">$base</span>/negative_prompt.json"</span> -o assets/negative_prompt.json<!----></pre></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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3OmniModularPipeline | |
| <span class="hljs-keyword">from</span> diffusers.schedulers.scheduling_unipc_multistep <span class="hljs-keyword">import</span> UniPCMultistepScheduler | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_video | |
| pipe = Cosmos3OmniModularPipeline.from_pretrained(<span class="hljs-string">"nvidia/Cosmos3-Nano"</span>, dtype=torch.bfloat16) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.to(<span class="hljs-string">"cuda"</span>) <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| pipe.scheduler = UniPCMultistepScheduler.from_config( | |
| pipe.scheduler.config, flow_shift=<span class="hljs-number">10.0</span>, use_karras_sigmas=<span class="hljs-literal">False</span> | |
| ) | |
| <span class="hljs-comment"># Downloaded into assets/ from the Cosmos cookbook (see the curl snippet above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/edge/prompt.json"</span>)) | |
| negative_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/negative_prompt.json"</span>)) | |
| control_edge = load_video(<span class="hljs-string">"assets/edge/control_edge.mp4"</span>) | |
| videos = pipe( | |
| prompt=json.dumps(json_prompt), | |
| negative_prompt=json.dumps(negative_prompt), | |
| control_videos={<span class="hljs-string">"edge"</span>: control_edge}, | |
| num_frames=<span class="hljs-number">121</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| fps=<span class="hljs-number">30.0</span>, | |
| num_inference_steps=<span class="hljs-number">35</span>, | |
| guidance_scale=<span class="hljs-number">3.0</span>, | |
| control_guidance=<span class="hljs-number">1.5</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| export_to_video(videos, <span class="hljs-string">"cosmos3_modular_transfer_edge.mp4"</span>, fps=<span class="hljs-number">30</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----> <!--[2--><h3 class="relative group"><a id="diffusers.Cosmos3OmniModularPipeline" 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="#diffusers.Cosmos3OmniModularPipeline"><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><!---->蒸留(少数ステップ)のテキストから画像へ、画像から動画へ<!----></span></h3><!--]--><!----> <p>Few-step distilled checkpoints は <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3DistilledModularPipeline">Cosmos3DistilledModularPipeline</a>（ブロック: <code>Cosmos3DistilledBlocks</code>）によって提供されます。ベースの <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniModularPipeline">Cosmos3OmniModularPipeline</a> および <a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniPipeline">Cosmos3OmniPipeline</a> はこれらをサポートしていません。<code>num_inference_steps</code> は、チェックポイントの <code>modular_model_index.json</code> にある <code>distilled_sigmas</code> パイプライン設定の長さに固定され、guidance が重みに組み込まれているため <code>guidance_scale</code> は 1.0 に強制されます。どちらかの値を他の値に設定するとエラーが発生し、<code>negative_prompt</code> は警告が表示された上で無視されます。</p> <p>FP8蒸留チェックポイント（<code>revision="fp8"</code>）は混合精度ポリシーを宣言しないため、すべてのステップがネイティブのW8A8のままになります。</p> <p>プロンプトは、蒸留されていないモデルと同じ記述的な JSON 構造に従うため、短いテキストはまずアップサンプリングする必要があります。<a href="#prompt-upsampling">Prompt upsampling</a> に記載されているように、<code>--mode text2image</code> (T2I) または <code>--mode image2video</code> (I2V) を使用してアップサンプリングを行い、その後 <code>json.dumps(...)</code> を介して JSON を渡します。</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-python "><!----><span class="hljs-keyword">import</span> json | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Cosmos3DistilledModularPipeline | |
| <span class="hljs-keyword">from</span> diffusers.utils <span class="hljs-keyword">import</span> export_to_video, load_image | |
| <span class="hljs-comment"># JSON-upsampled prompt (see "Prompt upsampling" above).</span> | |
| json_prompt = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_t2i_prompt.json"</span>)) | |
| repo = <span class="hljs-string">"nvidia/Cosmos3-Super-Text2Image-4Step"</span> | |
| pipe = Cosmos3DistilledModularPipeline.from_pretrained(repo, dtype=torch.bfloat16) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.to(<span class="hljs-string">"cuda"</span>) <span class="hljs-comment"># or "mps", "xpu", "cpu"</span> | |
| <span class="hljs-comment"># text-to-image (distilled)</span> | |
| videos = pipe( | |
| prompt=json.dumps(json_prompt), | |
| num_frames=<span class="hljs-number">1</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| videos[<span class="hljs-number">0</span>].save(<span class="hljs-string">"cosmos3_distilled_t2i.jpg"</span>, <span class="hljs-built_in">format</span>=<span class="hljs-string">"JPEG"</span>, quality=<span class="hljs-number">85</span>) | |
| <span class="hljs-comment"># image-to-video (distilled) — load the I2V repo instead</span> | |
| <span class="hljs-comment"># JSON-upsampled prompt (see "Prompt upsampling" above); upsampled from the source prompt</span> | |
| <span class="hljs-comment"># "The right robotic hand picks up the red sphere on the shelf."</span> | |
| json_prompt_i2v = json.load(<span class="hljs-built_in">open</span>(<span class="hljs-string">"assets/example_i2v_prompt.json"</span>)) | |
| repo_i2v = <span class="hljs-string">"nvidia/Cosmos3-Super-Image2Video-4Step"</span> | |
| pipe = Cosmos3DistilledModularPipeline.from_pretrained(repo_i2v, dtype=torch.bfloat16) | |
| pipe.load_components(dtype=torch.bfloat16) | |
| pipe.to(<span class="hljs-string">"cuda"</span>) | |
| image = load_image( | |
| <span class="hljs-string">"https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_153.jpg"</span> | |
| ) | |
| videos = pipe( | |
| prompt=json.dumps(json_prompt_i2v), | |
| image=image, | |
| num_frames=<span class="hljs-number">189</span>, | |
| height=<span class="hljs-number">720</span>, | |
| width=<span class="hljs-number">1280</span>, | |
| output=<span class="hljs-string">"videos"</span>, | |
| ) | |
| export_to_video(videos, <span class="hljs-string">"cosmos3_distilled_i2v.mp4"</span>, fps=<span class="hljs-number">24</span>, macro_block_size=<span class="hljs-number">1</span>)<!----></pre></div><!----> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.Cosmos3OmniModularPipeline"><!----><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">クラス</span> <span class="font-medium">diffusers.</span><span class="font-semibold">Cosmos3OmniModularPipeline</span></span></h3><!----> <a id="diffusers.Cosmos3OmniModularPipeline" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.Cosmos3OmniModularPipeline"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/modular_pipelines/cosmos/modular_pipeline.py#L7" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">blocks<span class="opacity-60">: diffusers.modular_pipelines.modular_pipeline.ModularPipelineBlocks | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">pretrained_model_name_or_path<span class="opacity-60">: str | os.PathLike | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">components_manager<span class="opacity-60">: diffusers.modular_pipelines.components_manager.ComponentsManager | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">collection<span class="opacity-60">: str | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">workflow<span class="opacity-60">: str | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">modular_config_dict<span class="opacity-60">: dict[str, typing.Any] | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">config_dict<span class="opacity-60">: dict[str, typing.Any] | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">**kwargs<span class="opacity-60"></span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>Cosmos 3 のオムニ生成用の ModularPipeline。</p></div> <ul><li>すべて</li> <li><strong>call</strong></li></ul> <!--[1--><h2 class="relative group"><a id="diffusers.Cosmos3DistilledModularPipeline" 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="#diffusers.Cosmos3DistilledModularPipeline"><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><!---->Cosmos3DistilledModularPipeline<!----></span></h2><!--]--><!----> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.Cosmos3DistilledModularPipeline"><!----><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">class</span> <span class="font-medium">diffusers.</span><span class="font-semibold">Cosmos3DistilledModularPipeline</span></span></h3><!----> <a id="diffusers.Cosmos3DistilledModularPipeline" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.Cosmos3DistilledModularPipeline"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/modular_pipelines/cosmos/modular_pipeline.py#L121" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">blocks<span class="opacity-60">: diffusers.modular_pipelines.modular_pipeline.ModularPipelineBlocks | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">pretrained_model_name_or_path<span class="opacity-60">: str | os.PathLike | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">components_manager<span class="opacity-60">: diffusers.modular_pipelines.components_manager.ComponentsManager | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">collection<span class="opacity-60">: str | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">workflow<span class="opacity-60">: str | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">modular_config_dict<span class="opacity-60">: dict[str, typing.Any] | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">config_dict<span class="opacity-60">: dict[str, typing.Any] | None = None</span></span></span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">**kwargs<span class="opacity-60"></span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>蒸留(少数ステップ)Cosmos 3 omni 生成用の ModularPipeline です。</p> <p>蒸留済みチェックポイントは、classifier-free guidance を重みに焼き込み、パイプラインの <code>distilled_sigmas</code> 設定（<code>modular_model_index.json</code> から設定されます）から読み取られる固定スケジュールでサンプリングを行うため、<code>guidance_scale</code> と <code>num_inference_steps</code> は固定されており、<code>negative_prompt</code> はサポートされていません。</p></div> <ul><li>すべて</li> <li><strong>call</strong></li></ul> <!--[1--><h2 class="relative group"><a id="diffusers.CosmosActionCondition" 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="#diffusers.CosmosActionCondition"><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><!---->CosmosActionCondition<!----></span></h2><!--]--><!----> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.CosmosActionCondition"><!----><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">class</span> <span class="font-medium">diffusers.</span><span class="font-semibold">CosmosActionCondition</span></span></h3><!----> <a id="diffusers.CosmosActionCondition" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.CosmosActionCondition"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L302" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">mode<span class="opacity-60">: typing.Literal['policy', 'forward_dynamics', 'inverse_dynamics']</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">chunk_size<span class="opacity-60">: int</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">domain_name<span class="opacity-60">: str</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">resolution_tier<span class="opacity-60">: int = 480</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">raw_actions<span class="opacity-60">: typing.Optional[torch.Tensor] = None</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">image<span class="opacity-60">: typing.Union[PIL.Image.Image, numpy.ndarray, torch.Tensor, NoneType] = None</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">video<span class="opacity-60">: typing.Union[list, numpy.ndarray, torch.Tensor, NoneType] = None</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">view_point<span class="opacity-60">: str = 'ego_view'</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[0--><p class="flex items-center font-semibold !mt-2 !mb-2 text-gray-800">パラメータ <span class="flex-auto border-t-2 border-gray-100 dark:border-gray-700 ml-3"></span></p> <ul class="px-2"><!--[--><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.mode" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.mode"><span><svg class="text-smd" 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><!----><strong>mode</strong> (<code>str</code>) — | |
| アクションタスク。 <code>"forward_dynamics"</code>（最初のフレームと指定された<code>raw_actions</code>シーケンスから将来の動画をロールアウトする）、<code>"inverse_dynamics"</code>（条件付けフレームを結ぶアクションを推論する）、または<code>"policy"</code>（最初のフレームから将来の動画とアクションを同時にロールアウトする）のいずれかです。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.chunk_size" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.chunk_size"><span><svg class="text-smd" 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><!----><strong>chunk_size</strong> (<code>int</code>) — | |
| チャンク内のアクション遷移ステップ数です。対応する条件付け動画は<code>chunk_size + 1</code>フレームにまたがります。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.domain_name" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.domain_name"><span><svg class="text-smd" 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><!----><strong>domain_name</strong> (<code>str</code>) — | |
| ドメイン認識アクション投影の重みを選択するためのエンボディメントドメインです。登録済みの Cosmos 3 エンボディメントドメインのいずれかである必要があります。また、予測アクションをスライスするために使用されるパディングなしのアクション幅も、この名前から内部で解決されるため固定されます（<code>_EMBODIMENT_TO_RAW_ACTION_DIM</code> を参照）。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.resolution_tier" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.resolution_tier"><span><svg class="text-smd" 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><!----><strong>resolution_tier</strong> (<code>int</code>、デフォルトは <code>480</code>) — | |
| アクション条件付けの解像度 <em>tier</em>（<code>256</code>、<code>480</code>、<code>704</code>、<code>720</code> のいずれか）。tier は、入力に最も近いアスペクト比を持つ事前定義済みのキャンバスを選択します。入力はダウンスケール（アップスケールは行われません）され、条件付けのためにそのキャンバスにパディングされます。これは出力フレームサイズではなく、出力は入力のコンテンツに追従します。tier は入力のネイティブ解像度に合わせる必要があります。低い tier を選ぶと詳細が失われ、高い tier を選ぶと解像度は向上しませんが（アップスケールされないため）、パディングに計算リソースが浪費され、学習と推論の不一致が生じて品質が低下する可能性があります。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.raw_actions" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.raw_actions"><span><svg class="text-smd" 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><!----><strong>raw_actions</strong> (<code>torch.Tensor</code>, <em>任意</em>) — | |
| <code>"forward_dynamics"</code> を駆動する、形状が <code>[T, raw_action_dim]</code> の生ドメインのアクションベクトルです。<code>chunk_size</code> より短いシーケンスでは最後のアクションが繰り返し使用され、長いシーケンスは切り詰められます。モデルの <code>action_dim</code> を超えるチャネルは拒否され、幅が狭い入力は <code>action_dim</code> までゼロパディングされます。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.image" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.image"><span><svg class="text-smd" 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><!----><strong>image</strong> (<code>PIL.Image.Image</code>、<code>np.ndarray</code>、または <code>torch.Tensor</code>、<em>省略可</em>) — | |
| <code>"policy"</code> / <code>"forward_dynamics"</code> の条件付けフレームです。<code>video</code> とは排他的です。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.video" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.video"><span><svg class="text-smd" 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><!----><strong>video</strong>（<code>list</code>、<code>np.ndarray</code>、または <code>torch.Tensor</code>、<em>任意</em>）— | |
| 条件付け用の動画です。<code>"inverse_dynamics"</code> では必須です。<code>"policy"</code> / <code>"forward_dynamics"</code> の場合は、最初のフレームのみが使用されます。<code>image</code> とは排他です。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.CosmosActionCondition.view_point" class="header-link block pr-0.5 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="#diffusers.CosmosActionCondition.view_point"><span><svg class="text-smd" 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><!----><strong>view_point</strong> (<code>str</code>、デフォルトは <code>"ego_view"</code>) — | |
| アクションキャプションの <code>cinematography.framing</code> フィールドに設定するために使用されるカメラの視点ラベル。 | |
| <code>"ego_view"</code>、<code>"third_person_view"</code>、<code>"wrist_view"</code>、または <code>"concat_view"</code> のいずれかです。アクションモデルは、この視点の文を含む構造化 JSON キャプションで学習されています。認識されないラベルが指定された場合、framing フィールドは削除されます（警告が表示されます）。<!----></span></span></li><!--]--></ul><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>Cosmos 3 のアクション条件付き生成タスクに必要なすべての入力をグループ化します。</p> <p>トップレベルの <code>image</code> / <code>height</code> / <code>width</code> 引数ではなく、<code>action</code> 引数を通じて <code>Cosmos3OmniPipeline.__call__()</code> に渡してください。これらの引数は t2v、i2v の実行に予約されています。</p></div> <!--[1--><h2 class="relative group"><a id="diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput" 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="#diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput"><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><!---->Cosmos3OmniPipelineOutput<!----></span></h2><!--]--><!----> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput"><!----><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">class</span> <span class="font-medium">diffusers.pipelines.cosmos.pipeline_cosmos3_omni.</span><span class="font-semibold">Cosmos3OmniPipelineOutput</span></span></h3><!----> <a id="diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput" class="header-link invisible with-hover:group-hover:visible pr-2" href="#diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput"><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><!----></a> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L283" target="_blank"><span><</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">video<span class="opacity-60">: Any</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">sound<span class="opacity-60">: typing.Optional[torch.Tensor] = None</span></span></span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">action<span class="opacity-60">: list[torch.Tensor] | None = None</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[0--><p class="flex items-center font-semibold !mt-2 !mb-2 text-gray-800">パラメータ <span class="flex-auto border-t-2 border-gray-100 dark:border-gray-700 ml-3"></span></p> <ul class="px-2"><!--[--><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput.video" class="header-link block pr-0.5 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="#diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput.video"><span><svg class="text-smd" 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><!----><strong>video</strong> — 生成された動画。正確な型は、パイプラインに渡される<code>output_type</code>によって異なります。<code>"pil"</code>（デフォルト）の場合はPILフレームのリスト、<code>"np"</code>の場合は<code>[T, H, W, C]</code>の形状を持つ<code>np.ndarray</code>、<code>"pt"</code>の場合は<code>[T, C, H, W]</code>の形状を持つ<code>torch.Tensor</code>、<code>output_type="latent"</code>の場合は生の潜在テンソルです。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput.sound" class="header-link block pr-0.5 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="#diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput.sound"><span><svg class="text-smd" 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><!----><strong>sound</strong> — 形状が<code>[C, N]</code>のデコード済み音声波形。<code>enable_sound=False</code>の場合、<code>None</code>になります。<!----></span></span></li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput.action" class="header-link block pr-0.5 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="#diffusers.pipelines.cosmos.pipeline_cosmos3_omni.Cosmos3OmniPipelineOutput.action"><span><svg class="text-smd" 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><!----><strong>action</strong> — 予測されたアクショントークン。アクションモードでアクションを予測する場合を除き、<code>None</code> です。<!----></span></span></li><!--]--></ul><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p><a href="/docs/diffusers/main/ja/api/pipelines/cosmos3#diffusers.Cosmos3OmniPipeline">Cosmos3OmniPipeline</a>の出力データクラスです。</p></div> <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/api/pipelines/cosmos3.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_gxjg5b = { | |
| 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.U6TUchwJ.js"), | |
| import("/docs/diffusers/main/en/_app/immutable/entry/app.3qKhPoX-.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 7], | |
| data: [null,null], | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 241 kB
- Xet hash:
- 34e6756ed4565509054bac6e26d17e8fe9208f7b71dcd86d775b1aa8cbfb84e9
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.