{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"CalibAtt：无需训练的稀疏注意力方法，将文生视频速度提升至 1.58 倍","description":"Apple 与特拉维夫大学联合提出 CalibAtt，一种无需训练的校准稀疏注意力方法，通过离线识别 token 间可跳过的低分连接并编译为优化操作，在推理时跳过无关计算。在 Wan 2.1 14B、Mochi 1 及少步蒸馏模型上，CalibAtt 实现最高 1.58 倍端到端加速，在保持视频质量和文本-视频对齐的同时优于现有免训练方法。","url":"https://www.aioga.com/news/cmruw1v74005vbiymxxxmhf22/","mainEntityOfPage":"https://www.aioga.com/news/cmruw1v74005vbiymxxxmhf22/","datePublished":"2026-07-21T00:00:00.000Z","dateModified":"2026-07-21T00:00:00.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://machinelearning.apple.com/research/calibrated-sparse-attention","https://aihot.virxact.com/items/cmruw1v74005vbiymxxxmhf22"],"canonicalUrl":"https://www.aioga.com/news/cmruw1v74005vbiymxxxmhf22/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：Apple 与特拉维夫大学联合提出 CalibAtt，一种无需训练的校准稀疏注意力方法，通过离线识别 token 间可跳过的低分连接并编译为优化操作，在推理时跳过无关计算。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmruw1v74005vbiymxxxmhf22/","dateCreated":"2026-07-21T00:00:00.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"machinelearning.apple.com source article","url":"https://machinelearning.apple.com/research/calibrated-sparse-attention","datePublished":"2026-07-21T00:00:00.000Z","provider":{"@type":"Organization","name":"machinelearning.apple.com","url":"https://machinelearning.apple.com/research/calibrated-sparse-attention"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmruw1v74005vbiymxxxmhf22","datePublished":"2026-07-21T00:00:00.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmruw1v74005vbiymxxxmhf22"}}],"aggregationSource":"Apple Machine Learning Research（RSS）","originalPublisher":{"name":"machinelearning.apple.com","url":"https://machinelearning.apple.com/research/calibrated-sparse-attention"},"article":{"id":"cmruw1v74005vbiymxxxmhf22","slug":"cmruw1v74005vbiymxxxmhf22","url":"https://www.aioga.com/news/cmruw1v74005vbiymxxxmhf22/","title":"CalibAtt：无需训练的稀疏注意力方法，将文生视频速度提升至 1.58 倍","title_en":"Accelerating Text-to-Video Generation with Calibrated Sparse Attention","summary":"Apple 与特拉维夫大学联合提出 CalibAtt，一种无需训练的校准稀疏注意力方法，通过离线识别 token 间可跳过的低分连接并编译为优化操作，在推理时跳过无关计算。在 Wan 2.1 14B、Mochi 1 及少步蒸馏模型上，CalibAtt 实现最高 1.58 倍端到端加速，在保持视频质量和文本-视频对齐的同时优于现有免训练方法。","source":"Apple Machine Learning Research（RSS）","sourceUrl":"https://machinelearning.apple.com/research/calibrated-sparse-attention","aiHotUrl":"https://aihot.virxact.com/items/cmruw1v74005vbiymxxxmhf22","publishedAt":"2026-07-21T00:00:00.000Z","category":"论文研究","score":55,"selected":false,"articleBody":["Accelerating Text-to-Video Generation with Calibrated Sparse Attention","Authors Shai Yehezkel†**, Shahar Yadin, Noam Elata, Yaron Ostrovsky-Berman, Bahjat Kawar","View publication：https://arxiv.org/abs/2603.05503","VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization","July 2, 2026 research area Computer Vision：/research/?domain=Computer%20Vision conference ICML：/research/?event=ICML","Visual tokenizers map high-dimensional raw pixels into a compressed representation for downstream modeling. Beyond compression, tokenizers dictate what information is preserved and how it is organized. A de facto standard approach to video tokenization is to represent a video as a spatiotemporal 3D grid of tokens, each capturing the corresponding local information in the original signal. This requires the downstream model that consumes the…","STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows","April 30, 2026 research area Computer Vision：/research/?domain=Computer%20Vision, research area Methods and Algorithms：/research/?domain=Methods%20and%20Algorithms conference CVPR：/research/?event=CVPR","Normalizing flows (NFs) are end-to-end likelihood-based generative models for continuous data, and have recently regained attention with encouraging progress on image generation. Yet in the video generation domain, where spatiotemporal complexity and computational cost are substantially higher, state-of-the-art systems almost exclusively rely on diffusion-based models. In this work, we revisit this design space by presenting STARFlow-V, a…","Our research in machine learning breaks new ground every day."],"articleImages":[{"sourceUrl":"https://mlr.cdn-apple.com/media/Discover_1440x420_2x_9c465d585e.jpg","alt":"Bottom banner","afterParagraph":8,"url":"/media/articles/cmruw1v74005vbiymxxxmhf22/64c2324784f2cf67.jpg"}],"mediaStatus":"ok","articleBodyZh":["加速基于校准稀疏注意力的文本到视频生成","作者 Shai Yehezkel†**, Shahar Yadin, Noam Elata, Yaron Ostrovsky-Berman, Bahjat Kawar","查看出版物：https://arxiv.org/abs/2603.05503","VideoFlexTok：灵活长度的粗到细视频标记化","2026年7月2日 研究领域 计算机视觉：/research/?domain=Computer%20Vision 会议 ICML：/research/?event=ICML","视觉标记器将高维原始像素映射到下游建模的压缩表示。除了压缩之外，标记器还决定了哪些信息被保留以及如何组织。视频标记化的事实标准方法是将视频表示为一个时空 3D 令牌网格，每个令牌捕捉原始信号中相应的局部信息。这要求下游模型必须处理...","STARFlow-V：基于归一化流的端到端视频生成建模","2026年4月30日 研究领域 计算机视觉：/research/?domain=Computer%20Vision, 研究领域 方法与算法：/research/?domain=Methods%20and%20Algorithms 会议 CVPR：/research/?event=CVPR","归一化流（NFs）是面向连续数据的端到端基于似然的生成模型，并且最近在图像生成方面取得了令人鼓舞的进展。但在视频生成领域，由于时空复杂性和计算成本显著增加，最先进的系统几乎完全集中依赖基于扩散的模型。在这项工作中，我们通过提出 STARFlow-V 重新审视了这一设计空间，该模型...","我们的机器学习研究每天都在开辟新领域。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：Apple 与特拉维夫大学联合提出 CalibAtt，一种无需训练的校准稀疏注意力方法，通过离线识别 token 间可跳过的低分连接并编译为优化操作，在推理时跳过无关计算。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：模型与研究类动态需要结合能力边界、开放方式、成本、可用性和真实任务表现判断，单项指标领先不等于已经形成稳定采用。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文档、实际可用性、价格变化、开发者反馈和竞品回应。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-07-23T06:49:19.093Z","sourceHash":"a19da4343dfe6bbf","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["论文研究","Apple Machine Learning Research（RSS）"],"translations":{"zh-CN":{"title":"CalibAtt：无需训练的稀疏注意力方法，将文生视频速度提升至 1.58 倍","summary":"Apple 与特拉维夫大学联合提出 CalibAtt，一种无需训练的校准稀疏注意力方法，通过离线识别 token 间可跳过的低分连接并编译为优化操作，在推理时跳过无关计算。在 Wan 2.1 14B、Mochi 1 及少步蒸馏模型上，CalibAtt 实现最高 1.58 倍端到端加速，在保持视频质量和文本-视频对齐的同时优于现有免训练方法。","category":"论文研究","source":"machinelearning.apple.com","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt：无需训练的稀疏注意力方法，将文生视频速度提升至 1.58 倍 - Aioga AI资讯","description":"Apple 与特拉维夫大学联合提出 CalibAtt，一种无需训练的校准稀疏注意力方法，通过离线识别 token 间可跳过的低分连接并编译为优化操作，在推理时跳过无关计算。在 Wan 2.1 14B、Mochi 1 及少步蒸馏模型上，CalibAtt 实现最高 1.58 倍端到端加速，在保持视频质量和文本-视频对齐的同时优于现有免训练方法。","url":"https://www.aioga.com/news/cmruw1v74005vbiymxxxmhf22/"},"en":{"title":"CalibAtt: A training-free sparse attention method that increases the speed of text-to-video generation to 1.58 times","summary":"Apple and Tel Aviv University jointly proposed CalibAtt, a training-free calibrated sparse attention method that skips irrelevant computations during inference by offline identifying low-score connections between tokens that can be skipped and compiling them into optimized operations. On Wan 2.1 14B, Mochi 1, and few-step distilled models, CalibAtt achieves up to 1.58× end-to-end acceleration, outperforming existing training-free methods while maintaining video quality and text-video alignment.","category":"Research","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: A training-free sparse attention method that increases the speed of text-to-video generation to 1.58 times - Aioga AI News","description":"Apple and Tel Aviv University jointly proposed CalibAtt, a training-free calibrated sparse attention method that skips irrelevant computations during inference by offline identifyi...","url":"https://www.aioga.com/en/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:02:24.096Z"},"ja":{"title":"CalibAtt：トレーニング不要のスパースアテンション手法により、テキストから動画への生成速度を1.58倍に向上","summary":"Appleはテルアビブ大学と共同でCalibAttを提案しました。これは、訓練を必要としないキャリブレーション型のスパースアテンション手法で、オフラインでトークン間のスキップ可能な低スコア接続を識別し、最適化操作としてコンパイルすることで、推論時に無関係な計算をスキップします。Wan 2.1 14B、Mochi 1、および少ステップ蒸留モデル上で、CalibAttはエンドツーエンドで最大1.58倍の高速化を実現し、動画品質とテキスト-ビデオの整合性を維持しつつ、既存の訓練不要手法よりも優れています。","category":"論文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt：トレーニング不要のスパースアテンション手法により、テキストから動画への生成速度を1.58倍に向上 - Aioga AIニュース","description":"Appleはテルアビブ大学と共同でCalibAttを提案しました。これは、訓練を必要としないキャリブレーション型のスパースアテンション手法で、オフラインでトークン間のスキップ可能な低スコア接続を識別し、最適化操作としてコンパイルすることで、推論時に無関係な計算をスキップします。Wan 2.1 14B、Mochi 1、および少ステップ蒸留モデル上で、Calib...","url":"https://www.aioga.com/ja/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:02:53.331Z"},"ko":{"title":"CalibAtt: 훈련이 필요 없는 희소 주의 방법으로, 텍스트-비디오 생성 속도를 1.58배로 향상","summary":"Apple은 텔아비브 대학교와 공동으로 CalibAtt를 제안했는데, 이는 훈련이 필요 없는 보정 희소 어텐션 방법으로, 오프라인에서 토큰 간 건너뛸 수 있는 낮은 점수 연결을 식별하고 최적화 연산으로 컴파일하여 추론 시 관련 없는 연산을 건너뛸 수 있도록 한다. Wan 2.1 14B, Mochi 1 및 소규모 단계 증류 모델에서 CalibAtt는 최대 1.58배의 엔드투엔드 가속을 달성했으며, 비디오 품질과 텍스트-비디오 정렬을 유지하면서 기존의 비훈련 방법보다 우수하다.","category":"연구","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: 훈련이 필요 없는 희소 주의 방법으로, 텍스트-비디오 생성 속도를 1.58배로 향상 - Aioga AI 뉴스","description":"Apple은 텔아비브 대학교와 공동으로 CalibAtt를 제안했는데, 이는 훈련이 필요 없는 보정 희소 어텐션 방법으로, 오프라인에서 토큰 간 건너뛸 수 있는 낮은 점수 연결을 식별하고 최적화 연산으로 컴파일하여 추론 시 관련 없는 연산을 건너뛸 수 있도록 한다. Wan 2.1 14B, Mochi 1 및 소규모 단계 증류...","url":"https://www.aioga.com/ko/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:03:39.859Z"},"es":{"title":"CalibAtt: Método de atención dispersa que no requiere entrenamiento, que aumenta la velocidad de VideoText a 1,58 veces","summary":"Apple y la Universidad de Tel Aviv propusieron conjuntamente CalibAtt, un método de atención dispersa calibrada sin necesidad de entrenamiento, que identifica de forma offline las conexiones de baja puntuación que se pueden saltar entre tokens y las compila como operaciones optimizadas, omitiendo cálculos irrelevantes durante la inferencia. En los modelos Wan 2.1 14B, Mochi 1 y modelos de destilación de pocos pasos, CalibAtt logra una aceleración de extremo a extremo de hasta 1,58 veces, superando a los métodos existentes sin entrenamiento mientras mantiene la calidad de video y la alineación texto-video.","category":"Investigación","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Método de atención dispersa que no requiere entrenamiento, que aumenta la velocidad de VideoText a 1,58 veces - Aioga Noticias de IA","description":"Apple y la Universidad de Tel Aviv propusieron conjuntamente CalibAtt, un método de atención dispersa calibrada sin necesidad de entrenamiento, que identifica de forma offline las...","url":"https://www.aioga.com/es/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:03:36.449Z"},"fr":{"title":"CalibAtt : une méthode d'attention éparse sans entraînement, qui augmente la vitesse de Vidéo à Texte à 1,58 fois","summary":"Apple et l'Université de Tel Aviv ont conjointement proposé CalibAtt, une méthode d'attention éparse calibrée sans entraînement, qui identifie hors ligne les connexions à faible score entre les tokens pouvant être sautées et les compile en opérations optimisées, permettant d'éviter les calculs non pertinents lors de l'inférence. Sur les modèles Wan 2.1 14B, Mochi 1 et les modèles de distillation en quelques étapes, CalibAtt réalise jusqu'à 1,58 fois l'accélération de bout en bout, surpassant les méthodes sans entraînement existantes tout en maintenant la qualité vidéo et l'alignement texte-vidéo.","category":"Recherche","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt : une méthode d'attention éparse sans entraînement, qui augmente la vitesse de Vidéo à Texte à 1,58 fois - Aioga Actualités IA","description":"Apple et l'Université de Tel Aviv ont conjointement proposé CalibAtt, une méthode d'attention éparse calibrée sans entraînement, qui identifie hors ligne les connexions à faible sc...","url":"https://www.aioga.com/fr/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:04:26.774Z"},"de":{"title":"CalibAtt: Eine trainingsfreie Methode der sparsamen Aufmerksamkeit, die die Geschwindigkeit von textbasierten Videos auf das 1,58-fache erhöht","summary":"Apple und die Universität Tel Aviv haben gemeinsam CalibAtt vorgestellt, eine trainingsfreie Methode zur Kalibrierung sparsamer Aufmerksamkeit, die durch das offline Erkennen von Tokens mit überspringbaren niedrigen Verbindungen und deren Kompilierung zu optimierten Operationen irrelevante Berechnungen während der Inferenz überspringt. Bei Wan 2.1 14B, Mochi 1 und einigen distillierten Modellen mit wenigen Schritten erreicht CalibAtt eine End-to-End-Beschleunigung von bis zu 1,58-fach und übertrifft dabei bestehende trainingsfreie Methoden, während die Videoqualität und die Text-Video-Ausrichtung erhalten bleiben.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Eine trainingsfreie Methode der sparsamen Aufmerksamkeit, die die Geschwindigkeit von textbasierten Videos auf das 1,58-fache erhöht - Aioga KI-News","description":"Apple und die Universität Tel Aviv haben gemeinsam CalibAtt vorgestellt, eine trainingsfreie Methode zur Kalibrierung sparsamer Aufmerksamkeit, die durch das offline Erkennen von T...","url":"https://www.aioga.com/de/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:04:25.791Z"},"pt-BR":{"title":"CalibAtt: um método de atenção esparsa sem necessidade de treinamento, que aumenta a velocidade de vídeos gerados por texto em até 1,58 vezes","summary":"A Apple, em conjunto com a Universidade de Tel Aviv, apresentou o CalibAtt, um método de atenção esparsa calibrada sem treinamento, que identifica offline as conexões de baixa pontuação entre tokens que podem ser ignoradas e as compila em operações otimizadas, pulando cálculos irrelevantes durante a inferência. Nos modelos Wan 2.1 14B, Mochi 1 e em modelos de destilação de poucos passos, o CalibAtt alcançou uma aceleração de ponta a ponta de até 1,58 vezes, superando os métodos existentes sem treinamento, mantendo a qualidade do vídeo e o alinhamento texto-vídeo.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: um método de atenção esparsa sem necessidade de treinamento, que aumenta a velocidade de vídeos gerados por texto em até 1,58 vezes - Aioga Notícias de IA","description":"A Apple, em conjunto com a Universidade de Tel Aviv, apresentou o CalibAtt, um método de atenção esparsa calibrada sem treinamento, que identifica offline as conexões de baixa pont...","url":"https://www.aioga.com/pt-BR/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:05:08.290Z"},"ru":{"title":"CalibAtt: Метод разреженного внимания без обучения, ускоряющий генерацию видео с текста до 1,58 раза","summary":"Apple совместно с Университетом Тель-Авива предложили CalibAtt, метод калиброванного разреженного внимания без обучения, который через оффлайн-выявление мало значимых соединений между токенами, которые можно пропустить, и компиляцию их в оптимизированные операции, позволяет пропускать несущественные вычисления во время инференса. На моделях Wan 2.1 14B, Mochi 1 и моделях с малым числом шагов дистилляции, CalibAtt достигает максимального ускорения сквозного выполнения в 1,58 раза, превосходя существующие методы без обучения при сохранении качества видео и соответствия текста и видео.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Метод разреженного внимания без обучения, ускоряющий генерацию видео с текста до 1,58 раза - Aioga Новости ИИ","description":"Apple совместно с Университетом Тель-Авива предложили CalibAtt, метод калиброванного разреженного внимания без обучения, который через оффлайн-выявление мало значимых соединений ме...","url":"https://www.aioga.com/ru/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:05:11.246Z"},"ar":{"title":"CalibAtt: طريقة الانتباه النادر التي لا تتطلب تدريبًا، ترفع سرعة تحويل النص إلى فيديو بمقدار 1.58 مرة","summary":"قدمت شركة آبل وجامعة تل أبيب معًا CalibAtt، وهي طريقة لمواءمة الانتباه النادر دون تدريب، تقوم من خلال التعرف خارج الخط على الروابط منخفضة الدرجة التي يمكن تخطيها بين الرموز وتجميعها في عمليات محسّنة، لتتخطى الحسابات غير ذات الصلة أثناء الاستدلال. على نماذج Wan 2.1 14B وMochi 1 ونماذج التقطير قليلة الخطوات، تحقق CalibAtt تسريعًا أقصى بمقدار 1.58 مرة من البداية للنهاية، متفوقة على الطرق الحالية التي لا تتطلب تدريبًا مع الحفاظ على جودة الفيديو وتوافق النص مع الفيديو.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: طريقة الانتباه النادر التي لا تتطلب تدريبًا، ترفع سرعة تحويل النص إلى فيديو بمقدار 1.58 مرة - Aioga أخبار الذكاء الاصطناعي","description":"قدمت شركة آبل وجامعة تل أبيب معًا CalibAtt، وهي طريقة لمواءمة الانتباه النادر دون تدريب، تقوم من خلال التعرف خارج الخط على الروابط منخفضة الدرجة التي يمكن تخطيها بين الرموز وتجميعه...","url":"https://www.aioga.com/ar/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:06:03.668Z"},"hi":{"title":"CalibAtt: बिना प्रशिक्षण का दुर्बल ध्यान तरीका, जीवित वीडियो की गति को 1.58 गुना तक बढ़ाता है","summary":"Apple ने तेल अवीव विश्वविद्यालय के साथ मिलकर CalibAtt प्रस्तुत किया, एक ऐसी प्रशिक्षण-रहित कैलिब्रेटेड स्पर्स एटेंशन विधि, जो ऑफ़लाइन यह पहचानती है कि किस टोकन के बीच के निम्न स्कोर कनेक्शन को छोड़ सकते हैं और इसे ऑप्टिमाइज्ड ऑपरेशन में कंपाइल करती है, ताकि इन्फ़रेंस के दौरान अप्रासंगिक गणनाओं को छोड़ दिया जा सके। Wan 2.1 14B, Mochi 1 और शॉर्ट-स्टेप डिस्टिलेशन मॉडल पर, CalibAtt ने अधिकतम 1.58 गुना एंड-टू-एंड गति हासिल की, वीडियो गुणवत्ता और टेक्स्ट-वीडियो समकालिकता बनाए रखते हुए मौजूदा प्रशिक्षण-रहित विधियों से बेहतर प्रदर्शन किया।","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: बिना प्रशिक्षण का दुर्बल ध्यान तरीका, जीवित वीडियो की गति को 1.58 गुना तक बढ़ाता है - Aioga AI समाचार","description":"Apple ने तेल अवीव विश्वविद्यालय के साथ मिलकर CalibAtt प्रस्तुत किया, एक ऐसी प्रशिक्षण-रहित कैलिब्रेटेड स्पर्स एटेंशन विधि, जो ऑफ़लाइन यह पहचानती है कि किस टोकन के बीच के निम्न स्को...","url":"https://www.aioga.com/hi/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:05:57.682Z"},"it":{"title":"CalibAtt: metodo di attenzione sparsa senza addestramento, che aumenta la velocità dei video generati da testo di 1,58 volte","summary":"Apple e l'Università di Tel Aviv hanno proposto congiuntamente CalibAtt, un metodo di attenzione sparsa calibrata senza addestramento, che identifica offline le connessioni a bassa priorità tra i token che possono essere saltate e le compila in operazioni ottimizzate, saltando così i calcoli irrilevanti durante l'inferenza. Su Wan 2.1 14B, Mochi 1 e modelli di distillazione a pochi passi, CalibAtt raggiunge fino a 1,58 volte di accelerazione end-to-end, superando i metodi senza addestramento esistenti mentre mantiene la qualità video e l'allineamento testo-video.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: metodo di attenzione sparsa senza addestramento, che aumenta la velocità dei video generati da testo di 1,58 volte - Aioga Notizie IA","description":"Apple e l'Università di Tel Aviv hanno proposto congiuntamente CalibAtt, un metodo di attenzione sparsa calibrata senza addestramento, che identifica offline le connessioni a bassa...","url":"https://www.aioga.com/it/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:06:54.269Z"},"nl":{"title":"CalibAtt: een spars attention-methode zonder training, die de snelheid van tekst-naar-video verhoogt tot 1,58 keer","summary":"Apple en de Universiteit van Tel Aviv hebben samen CalibAtt geïntroduceerd, een trainingsvrije methode voor het kalibreren van spaarzame aandacht. Deze methode identificeert offline de zwakke verbindingen tussen tokens die kunnen worden overgeslagen en compileert ze tot geoptimaliseerde bewerkingen, waardoor tijdens de inferentie irrelevante berekeningen worden overgeslagen. Op Wan 2.1 14B, Mochi 1 en enkele stappen van gedestilleerde modellen bereikt CalibAtt een maximale end-to-end versnelling van 1,58 keer, waarbij het bestaande trainingsvrije methoden overtreft en tegelijkertijd videokwaliteit en tekst-video-uitlijning behoudt.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: een spars attention-methode zonder training, die de snelheid van tekst-naar-video verhoogt tot 1,58 keer - Aioga AI-nieuws","description":"Apple en de Universiteit van Tel Aviv hebben samen CalibAtt geïntroduceerd, een trainingsvrije methode voor het kalibreren van spaarzame aandacht. Deze methode identificeert offlin...","url":"https://www.aioga.com/nl/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:06:41.283Z"},"tr":{"title":"CalibAtt: Eğitim gerektirmeyen seyrek dikkat yöntemi, metinden videoya hızını 1,58 kat artırır","summary":"Apple, Tel Aviv Üniversitesi ile birlikte CalibAtt'ı sundu; bu, eğitim gerektirmeyen bir doğrulama seyreklik dikkat yöntemi olup, tokenlar arasındaki atlanabilecek düşük puanlı bağlantıları çevrimdışı olarak tespit eder ve optimize işlemler olarak derler, bu sayede çıkarım sırasında ilgisiz hesaplamalardan kaçınılır. Wan 2.1 14B, Mochi 1 ve az adımlı damıtma modellerinde, CalibAtt uçtan uca en fazla 1,58 kat hızlanma sağlarken video kalitesini ve metin-video uyumunu korur ve mevcut eğitim gerektirmeyen yöntemlerden daha iyidir.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Eğitim gerektirmeyen seyrek dikkat yöntemi, metinden videoya hızını 1,58 kat artırır - Aioga AI Haberleri","description":"Apple, Tel Aviv Üniversitesi ile birlikte CalibAtt'ı sundu; bu, eğitim gerektirmeyen bir doğrulama seyreklik dikkat yöntemi olup, tokenlar arasındaki atlanabilecek düşük puanlı bağ...","url":"https://www.aioga.com/tr/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:07:36.026Z"},"vi":{"title":"CalibAtt: Phương pháp chú ý thưa thớt không cần huấn luyện, tăng tốc độ video sinh bởi văn bản lên 1,58 lần","summary":"Apple và Đại học Tel Aviv cùng đề xuất CalibAtt, một phương pháp chú ý thưa không cần huấn luyện, thông qua việc nhận diện ngoại tuyến các kết nối có điểm thấp có thể bỏ qua giữa các token và biên dịch thành các thao tác tối ưu, bỏ qua các tính toán không liên quan khi suy luận. Trên Wan 2.1 14B, Mochi 1 và các mô hình chưng cất một vài bước, CalibAtt đạt tốc độ xử lý đầu-cuối nhanh hơn tối đa 1,58 lần, trong khi vẫn duy trì chất lượng video và sự tương thích văn bản-video, vượt trội so với các phương pháp không cần huấn luyện hiện có.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Phương pháp chú ý thưa thớt không cần huấn luyện, tăng tốc độ video sinh bởi văn bản lên 1,58 lần - Tin tức AI Aioga","description":"Apple và Đại học Tel Aviv cùng đề xuất CalibAtt, một phương pháp chú ý thưa không cần huấn luyện, thông qua việc nhận diện ngoại tuyến các kết nối có điểm thấp có thể bỏ qua giữa c...","url":"https://www.aioga.com/vi/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:07:39.385Z"},"id":{"title":"CalibAtt: Metode perhatian jarang tanpa pelatihan, meningkatkan kecepatan video dari teks ke video hingga 1,58 kali","summary":"Apple dan Universitas Tel Aviv bersama-sama mengusulkan CalibAtt, sebuah metode perhatian jarang yang terkalibrasi tanpa pelatihan, yang melalui identifikasi offline koneksi skor rendah yang dapat dilewati antar token dan dikompilasi menjadi operasi yang dioptimalkan, dapat melewati komputasi yang tidak relevan saat inferensi. Pada Wan 2.1 14B, Mochi 1, dan model distilasi langkah sedikit, CalibAtt mencapai percepatan ujung-ke-ujung hingga 1,58 kali, sambil mempertahankan kualitas video dan keselarasan teks-video, lebih unggul dibandingkan metode tanpa pelatihan yang ada.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Metode perhatian jarang tanpa pelatihan, meningkatkan kecepatan video dari teks ke video hingga 1,58 kali - Berita AI Aioga","description":"Apple dan Universitas Tel Aviv bersama-sama mengusulkan CalibAtt, sebuah metode perhatian jarang yang terkalibrasi tanpa pelatihan, yang melalui identifikasi offline koneksi skor r...","url":"https://www.aioga.com/id/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:08:19.559Z"},"th":{"title":"CalibAtt: วิธีการสังเกตความสนใจแบบเบาบางที่ไม่ต้องฝึก ทำให้ความเร็วของการสร้างวิดีโอจากข้อความเพิ่มขึ้นเป็น 1.58 เท่า","summary":"Apple และมหาวิทยาลัยเทลอาวีฟร่วมกันเสนอ CalibAtt วิธีการปรับความสนใจแบบบางที่ไม่ต้องการการฝึกอบรม โดยระบุการเชื่อมต่อที่มีคะแนนต่ำซึ่งสามารถข้ามได้ระหว่าง token แบบออฟไลน์และคอมไพล์เป็นการดำเนินการที่ปรับให้เหมาะสม เพื่อข้ามการคำนวณที่ไม่เกี่ยวข้องในระหว่างการสืบหา ในโมเดล Wan 2.1 14B, Mochi 1 และแบบจำลองการกลั่นด้วยก้าวน้อย CalibAtt ทำให้ได้ความเร่งปลายทางต่อปลายทางสูงสุด 1.58 เท่า ในขณะที่ยังคงคุณภาพวิดีโอและการปรับวิดีโอกับข้อความได้ดีกว่าวิธีที่ไม่ต้องฝึกอบรมที่มีอยู่","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: วิธีการสังเกตความสนใจแบบเบาบางที่ไม่ต้องฝึก ทำให้ความเร็วของการสร้างวิดีโอจากข้อความเพิ่มขึ้นเป็น 1.58 เท่า - ข่าว AI Aioga","description":"Apple และมหาวิทยาลัยเทลอาวีฟร่วมกันเสนอ CalibAtt วิธีการปรับความสนใจแบบบางที่ไม่ต้องการการฝึกอบรม โดยระบุการเชื่อมต่อที่มีคะแนนต่ำซึ่งสามารถข้ามได้ระหว่าง token แบบออฟไลน์และคอมไพล...","url":"https://www.aioga.com/th/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:08:26.980Z"},"pl":{"title":"CalibAtt: Metoda rzadkiej uwagi, która nie wymaga treningu, zwiększa prędkość generowania wideo przez tekst o 1,58 razy","summary":"Apple i Uniwersytet w Tel Awiwie wspólnie zaproponowali CalibAtt, metodę skalibrowanej rzadkiej uwagi, która nie wymaga treningu. Poprzez offline rozpoznawanie niskoocenionych połączeń między tokenami, które można pominąć, i kompilowanie ich jako zoptymalizowane operacje, podczas wnioskowania pomija się nieistotne obliczenia. Na modelach Wan 2.1 14B, Mochi 1 oraz modelach destylacji z kilkoma krokami, CalibAtt osiąga maksymalne 1,58-krotne przyspieszenie end-to-end, przewyższając istniejące metody nie wymagające treningu, przy zachowaniu jakości wideo i zgodności tekst-wideo.","category":"论文研究","source":"Apple Machine Learning Research（RSS）","aggregationSource":"Apple Machine Learning Research（RSS）","pageTitle":"CalibAtt: Metoda rzadkiej uwagi, która nie wymaga treningu, zwiększa prędkość generowania wideo przez tekst o 1,58 razy - Aioga Wiadomości AI","description":"Apple i Uniwersytet w Tel Awiwie wspólnie zaproponowali CalibAtt, metodę skalibrowanej rzadkiej uwagi, która nie wymaga treningu. Poprzez offline rozpoznawanie niskoocenionych połą...","url":"https://www.aioga.com/pl/news/cmruw1v74005vbiymxxxmhf22/","contentTranslated":true,"sourceHash":"d092842ce769d342","translatedAt":"2026-07-23T02:09:12.635Z"}}}}