{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-28T05:20:57.982Z","headline":"SANA-Video 2.0：混合线性注意力与注意力残差实现高效视频生成","description":"SANA-Video 2.0 是一个混合视频扩散 Transformer，提供 5B 和 14B 两种规模，可在单 GPU 上生成最高 720p 视频。其 Hybrid Linear-Softmax Attention 以 3：1 比例混合线性与 softmax 注意力，配合 Block Attention Residuals 将深层有效秩提升约 12%。","url":"https://www.aioga.com/news/cmryvpwy905kerolgy47s0xdn/","mainEntityOfPage":"https://www.aioga.com/news/cmryvpwy905kerolgy47s0xdn/","datePublished":"2026-07-23T00:00:00.000Z","dateModified":"2026-07-23T00:00:00.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://arxiv.org/abs/2607.21553","https://aihot.virxact.com/items/cmryvpwy905kerolgy47s0xdn"],"canonicalUrl":"https://www.aioga.com/news/cmryvpwy905kerolgy47s0xdn/","directAnswer":{"@type":"Answer","text":"SANA-Video 2.0 是混合视频扩散 Transformer，包含 5B 与 14B 两种规模，公开材料称其可在单 GPU 上生成最高 720p 视频。","url":"https://www.aioga.com/news/cmryvpwy905kerolgy47s0xdn/","dateCreated":"2026-07-23T00:00:00.000Z","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":"arXiv source article","url":"https://arxiv.org/abs/2607.21553","datePublished":"2026-07-23T00:00:00.000Z","provider":{"@type":"Organization","name":"arXiv","url":"https://arxiv.org/abs/2607.21553"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmryvpwy905kerolgy47s0xdn","datePublished":"2026-07-23T00:00:00.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmryvpwy905kerolgy47s0xdn"}}],"aggregationSource":"HuggingFace Daily Papers（社区热门论文）","originalPublisher":{"name":"arXiv","url":"https://arxiv.org/abs/2607.21553"},"article":{"id":"cmryvpwy905kerolgy47s0xdn","slug":"cmryvpwy905kerolgy47s0xdn","url":"https://www.aioga.com/news/cmryvpwy905kerolgy47s0xdn/","title":"SANA-Video 2.0：混合线性注意力与注意力残差实现高效视频生成","title_en":"SANA-Video 2.0： Hybrid Linear Attention with Attention Residuals for Efficient Video Generation","summary":"SANA-Video 2.0 是一个混合视频扩散 Transformer，提供 5B 和 14B 两种规模，可在单 GPU 上生成最高 720p 视频。其 Hybrid Linear-Softmax Attention 以 3：1 比例混合线性与 softmax 注意力，配合 Block Attention Residuals 将深层有效秩提升约 12%。","source":"HuggingFace Daily Papers（社区热门论文）","sourceUrl":"https://arxiv.org/abs/2607.21553","aiHotUrl":"https://aihot.virxact.com/items/cmryvpwy905kerolgy47s0xdn","publishedAt":"2026-07-23T00:00:00.000Z","category":"模型更新","score":77,"selected":true,"articleBody":["SANA-Video 2.0 是一个混合视频扩散 Transformer，提供 5B 和 14B 两种规模，可在单 GPU 上生成最高 720p 视频。","其 Hybrid Linear-Softmax Attention 以 3：1 比例混合线性与 softmax 注意力，配合 Block Attention Residuals 将深层有效秩提升约 12%。"],"articleImages":[],"mediaStatus":"none","articleBodyZh":[],"translationStatus":"","bodyOrigin":"summary-fallback","editorial":{"summary":"SANA-Video 2.0 是混合视频扩散 Transformer，包含 5B 与 14B 两种规模，公开材料称其可在单 GPU 上生成最高 720p 视频。","background":"该模型采用 Hybrid Linear-Softmax Attention，按 3：1 比例混合线性与 softmax 注意力，并引入 Block Attention Residuals。","viewpoint":"Aioga 判断，方案重点是在视频生成中组合不同注意力机制，并通过残差设计改善深层表示；其实际效率与生成质量仍需更多公开结果支撑。","implications":"材料称 Block Attention Residuals 可将深层有效秩提升约 12%。这可能为控制视频扩散 Transformer 的计算负担与深层表示能力提供研究参考。","nextStep":"值得关注后续是否披露 5B 与 14B 模型的生成速度、显存占用、质量评测和测试条件，以验证单 GPU 生成最高 720p 视频的适用范围。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-07-26T03:30:39.095Z","sourceHash":"f4bce65640b0c620","review":{"approved":true,"groundedness":95,"clarity":94,"duplicationRisk":16,"blockingIssues":[],"notes":["候选内容准确覆盖了模型规模、单 GPU 最高 720p、3：1 混合注意力和深层有效秩提升约 12% 等来源信息。","“实际效率与生成质量仍需更多公开结果支撑”明确表述为 Aioga 的判断，未冒充来源事实。","可选地说明“深层有效秩提升约 12%”的具体测量口径和对照基线未在所给材料中披露。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["模型更新","HuggingFace Daily Papers（社区热门论文）"],"translations":{"zh-CN":{"title":"SANA-Video 2.0：混合线性注意力与注意力残差实现高效视频生成","summary":"SANA-Video 2.0 是一个混合视频扩散 Transformer，提供 5B 和 14B 两种规模，可在单 GPU 上生成最高 720p 视频。其 Hybrid Linear-Softmax Attention 以 3：1 比例混合线性与 softmax 注意力，配合 Block Attention Residuals 将深层有效秩提升约 12%。","category":"模型更新","source":"arXiv","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0：混合线性注意力与注意力残差实现高效视频生成 - Aioga AI资讯","description":"SANA-Video 2.0 是一个混合视频扩散 Transformer，提供 5B 和 14B 两种规模，可在单 GPU 上生成最高 720p 视频。其 Hybrid Linear-Softmax Attention 以 3：1 比例混合线性与 softmax 注意力，配合 Block Attention Residuals 将深层有效秩提升约 12%。","url":"https://www.aioga.com/news/cmryvpwy905kerolgy47s0xdn/"},"en":{"title":"SANA-Video 2.0: Efficient Video Generation with Hybrid Linear Attention and Attention Residuals","summary":"SANA-Video 2.0 is a hybrid video diffusion Transformer, available in 5B and 14B sizes, capable of generating videos up to 720p on a single GPU. Its Hybrid Linear-Softmax Attention mixes linear and softmax attention at a 3:1 ratio, and combined with Block Attention Residuals, it increases the effective rank of deep layers by about 12%.","category":"Models","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Efficient Video Generation with Hybrid Linear Attention and Attention Residuals - Aioga AI News","description":"SANA-Video 2.0 is a hybrid video diffusion Transformer, available in 5B and 14B sizes, capable of generating videos up to 720p on a single GPU. Its Hybrid Linear-Softmax Attention...","url":"https://www.aioga.com/en/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:36:42.370Z"},"ja":{"title":"SANA-Video 2.0：混合線形注意と注意残差による効率的なビデオ生成の実現","summary":"SANA-Video 2.0 はハイブリッド動画拡散トランスフォーマーで、5B と 14B の二つの規模を提供し、単一 GPU 上で最大 720p の動画を生成できます。そのハイブリッド・リニア-ソフトマックス注意機構は、3:1 の比率で線形注意とソフトマックス注意を混合し、ブロック注意残差と組み合わせることで深層の有効ランクを約 12% 向上させます。","category":"モデル更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0：混合線形注意と注意残差による効率的なビデオ生成の実現 - Aioga AIニュース","description":"SANA-Video 2.0 はハイブリッド動画拡散トランスフォーマーで、5B と 14B の二つの規模を提供し、単一 GPU 上で最大 720p の動画を生成できます。そのハイブリッド・リニア-ソフトマックス注意機構は、3:1 の比率で線形注意とソフトマックス注意を混合し、ブロック注意残差と組み合わせることで深層の有効ランクを約 12% 向上させます。","url":"https://www.aioga.com/ja/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:36:47.241Z"},"ko":{"title":"SANA-Video 2.0：혼합 선형 어텐션과 어텐션 잔차를 통해 효율적인 비디오 생성 구현","summary":"SANA-Video 2.0은 혼합 비디오 확산 Transformer로, 5B 및 14B 두 가지 규모를 제공하며 단일 GPU에서 최대 720p 비디오를 생성할 수 있습니다. 그 Hybrid Linear-Softmax Attention은 선형과 softmax 주의를 3:1 비율로 혼합하며, Block Attention Residuals와 함께 깊은 층의 유효 랭크를 약 12% 향상시킵니다.","category":"모델 업데이트","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0：혼합 선형 어텐션과 어텐션 잔차를 통해 효율적인 비디오 생성 구현 - Aioga AI 뉴스","description":"SANA-Video 2.0은 혼합 비디오 확산 Transformer로, 5B 및 14B 두 가지 규모를 제공하며 단일 GPU에서 최대 720p 비디오를 생성할 수 있습니다. 그 Hybrid Linear-Softmax Attention은 선형과 softmax 주의를 3:1 비율로 혼합하며, Block Attention Re...","url":"https://www.aioga.com/ko/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:37:28.072Z"},"es":{"title":"SANA-Video 2.0: Atención lineal híbrida y residuo de atención para la generación eficiente de videos","summary":"SANA-Video 2.0 es un Transformer de difusión de video híbrido, disponible en dos tamaños, 5B y 14B, y puede generar videos de hasta 720p en una sola GPU. Su Atención Híbrida Lineal-Softmax mezcla la atención lineal y softmax en una proporción de 3:1, y junto con los Residuos de Atención por Bloques eleva el rango efectivo profundo en aproximadamente un 12%.","category":"Modelos","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Atención lineal híbrida y residuo de atención para la generación eficiente de videos - Aioga Noticias de IA","description":"SANA-Video 2.0 es un Transformer de difusión de video híbrido, disponible en dos tamaños, 5B y 14B, y puede generar videos de hasta 720p en una sola GPU. Su Atención Híbrida Lineal...","url":"https://www.aioga.com/es/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:37:26.061Z"},"fr":{"title":"SANA-Video 2.0 : Attention linéaire mixte et résidu d'attention pour une génération vidéo efficace","summary":"SANA-Video 2.0 est un transformateur de diffusion vidéo hybride, disponible en deux tailles de 5B et 14B, capable de générer des vidéos jusqu'à 720p sur un seul GPU. Son attention Hybrid Linear-Softmax mélange l'attention linéaire et softmax dans un rapport de 3:1, et avec les Block Attention Residuals, il augmente le rang effectif profond d'environ 12%.","category":"Modèles","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0 : Attention linéaire mixte et résidu d'attention pour une génération vidéo efficace - Aioga Actualités IA","description":"SANA-Video 2.0 est un transformateur de diffusion vidéo hybride, disponible en deux tailles de 5B et 14B, capable de générer des vidéos jusqu'à 720p sur un seul GPU. Son attention...","url":"https://www.aioga.com/fr/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:38:04.414Z"},"de":{"title":"SANA-Video 2.0: Effiziente Videoerzeugung durch gemischte lineare Aufmerksamkeit und Aufmerksamkeitsrest","summary":"SANA-Video 2.0 ist ein hybrider Videodiffusions-Transformer, der in den Größen 5B und 14B erhältlich ist und auf einer einzelnen GPU Videos mit bis zu 720p erzeugen kann. Seine Hybrid Linear-Softmax Attention mischt lineare und Softmax-Attention im Verhältnis 3:1 und steigert mit Block Attention Residuals die effektive Tiefe um etwa 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Effiziente Videoerzeugung durch gemischte lineare Aufmerksamkeit und Aufmerksamkeitsrest - Aioga KI-News","description":"SANA-Video 2.0 ist ein hybrider Videodiffusions-Transformer, der in den Größen 5B und 14B erhältlich ist und auf einer einzelnen GPU Videos mit bis zu 720p erzeugen kann. Seine Hyb...","url":"https://www.aioga.com/de/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:38:06.578Z"},"pt-BR":{"title":"SANA-Video 2.0: Atenção linear híbrida e resíduo de atenção para geração eficiente de vídeo","summary":"SANA-Video 2.0 é um Transformer de difusão de vídeo híbrido, disponível em duas escalas, 5B e 14B, capaz de gerar vídeos de até 720p em uma única GPU. Sua Atenção Linear-Softmax Híbrida mixa atenção linear e softmax na proporção de 3:1, e junto com os Resíduos de Atenção em Bloco aumenta a classificação efetiva profunda em cerca de 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Atenção linear híbrida e resíduo de atenção para geração eficiente de vídeo - Aioga Notícias de IA","description":"SANA-Video 2.0 é um Transformer de difusão de vídeo híbrido, disponível em duas escalas, 5B e 14B, capaz de gerar vídeos de até 720p em uma única GPU. Sua Atenção Linear-Softmax Hí...","url":"https://www.aioga.com/pt-BR/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:38:46.567Z"},"ru":{"title":"SANA-Video 2.0: гибридное линейное внимание и остаточное внимание для эффективной генерации видео","summary":"SANA-Video 2.0 — это гибридный видео-диффузионный трансформер, доступный в двух масштабах: 5B и 14B, способный генерировать видео с разрешением до 720p на одном GPU. Его гибридное линейно-softmax внимание смешивает линейное и softmax внимание в соотношении 3:1, а блоковые остатки внимания (Block Attention Residuals) повышают эффективный ранг глубоких слоев примерно на 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: гибридное линейное внимание и остаточное внимание для эффективной генерации видео - Aioga Новости ИИ","description":"SANA-Video 2.0 — это гибридный видео-диффузионный трансформер, доступный в двух масштабах: 5B и 14B, способный генерировать видео с разрешением до 720p на одном GPU. Его гибридное...","url":"https://www.aioga.com/ru/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:38:46.655Z"},"ar":{"title":"SANA-Video 2.0: الجمع بين الانتباه الخطي المختلط وبقايا الانتباه لتحقيق توليد الفيديو بكفاءة","summary":"SANA-Video 2.0 هو محول نشر فيديو مختلط، ويقدم مقاسين 5B و14B، ويمكنه توليد فيديوهات تصل إلى 720p على GPU واحد. يجمع الانتباه الهجين الخطي-سوفتماكس Linear-Softmax Attention بين الانتباه الخطي والانتباه سوفتماكس بنسبة 3:1، ومع تحويليات الانتباه الكتلية Block Attention Residuals يزيد الرتبة الفعالة العميقة بحوالي 12٪.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: الجمع بين الانتباه الخطي المختلط وبقايا الانتباه لتحقيق توليد الفيديو بكفاءة - Aioga أخبار الذكاء الاصطناعي","description":"SANA-Video 2.0 هو محول نشر فيديو مختلط، ويقدم مقاسين 5B و14B، ويمكنه توليد فيديوهات تصل إلى 720p على GPU واحد. يجمع الانتباه الهجين الخطي-سوفتماكس Linear-Softmax Attention بين الان...","url":"https://www.aioga.com/ar/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:39:28.551Z"},"hi":{"title":"SANA-वीडियो 2.0: मिश्रित रैखिक ध्यान और ध्यान अवशेष के साथ कुशल वीडियो उत्पादन को लागू करना","summary":"SANA-Video 2.0 एक हाइब्रिड वीडियो डायफ्यूज़न ट्रांसफ़ॉर्मर है, जो 5B और 14B दो आकार प्रदान करता है, और एकल GPU पर अधिकतम 720p वीडियो उत्पन्न कर सकता है। इसका हाइब्रिड लिनियर-सॉफ़्टमैक्स अटेंशन 3:1 अनुपात में लिनियर और सॉफ़्टमैक्स अटेंशन को मिश्रित करता है, और ब्लॉक अटेंशन रेसिडुअल के साथ मिलकर गहरी परतों की प्रभावी रैंक को लगभग 12% बढ़ाता है।","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-वीडियो 2.0: मिश्रित रैखिक ध्यान और ध्यान अवशेष के साथ कुशल वीडियो उत्पादन को लागू करना - Aioga AI समाचार","description":"SANA-Video 2.0 एक हाइब्रिड वीडियो डायफ्यूज़न ट्रांसफ़ॉर्मर है, जो 5B और 14B दो आकार प्रदान करता है, और एकल GPU पर अधिकतम 720p वीडियो उत्पन्न कर सकता है। इसका हाइब्रिड लिनियर-सॉफ़्ट...","url":"https://www.aioga.com/hi/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:39:33.937Z"},"it":{"title":"SANA-Video 2.0: attenzione lineare ibrida e residuo di attenzione per la generazione efficiente di video","summary":"SANA-Video 2.0 è un Transformer di diffusione video ibrido, disponibile nelle dimensioni 5B e 14B, in grado di generare video fino a 720p su una singola GPU. La sua Hybrid Linear-Softmax Attention combina attenzione lineare e softmax in un rapporto di 3:1, e insieme ai Block Attention Residuals aumenta il rango efficace degli strati profondi di circa il 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: attenzione lineare ibrida e residuo di attenzione per la generazione efficiente di video - Aioga Notizie IA","description":"SANA-Video 2.0 è un Transformer di diffusione video ibrido, disponibile nelle dimensioni 5B e 14B, in grado di generare video fino a 720p su una singola GPU. La sua Hybrid Linear-S...","url":"https://www.aioga.com/it/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:40:12.862Z"},"nl":{"title":"SANA-Video 2.0: Efficiënte videoproductie realiseren door gemengde lineaire aandacht en aandachtresiduen","summary":"SANA-Video 2.0 is een hybride video-diffusie Transformer, beschikbaar in twee schalen van 5B en 14B, die video’s tot 720p kan genereren op een enkele GPU. De Hybrid Linear-Softmax Attention mengt lineaire en softmax-attentie in een verhouding van 3:1, en samen met Block Attention Residuals verhoogt dit de effectieve diepe rang met ongeveer 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Efficiënte videoproductie realiseren door gemengde lineaire aandacht en aandachtresiduen - Aioga AI-nieuws","description":"SANA-Video 2.0 is een hybride video-diffusie Transformer, beschikbaar in twee schalen van 5B en 14B, die video’s tot 720p kan genereren op een enkele GPU. De Hybrid Linear-Softmax...","url":"https://www.aioga.com/nl/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:40:12.580Z"},"tr":{"title":"SANA-Video 2.0: Karma Doğrusal Dikkat ve Dikkat Rezidüsü ile Verimli Video Üretimi","summary":"SANA-Video 2.0, tek bir GPU üzerinde en fazla 720p video üretebilen, 5B ve 14B boyutlarında iki ölçek sunan bir hibrit video yayılım Transformer’ıdır. Hibrit Lineer-Softmax Dikkati, lineer ve softmax dikkatini 3:1 oranında karıştırır ve Blok Dikkat Rezidülleri ile derin verimli rütbe yaklaşık %12 artırılır.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Karma Doğrusal Dikkat ve Dikkat Rezidüsü ile Verimli Video Üretimi - Aioga AI Haberleri","description":"SANA-Video 2.0, tek bir GPU üzerinde en fazla 720p video üretebilen, 5B ve 14B boyutlarında iki ölçek sunan bir hibrit video yayılım Transformer’ıdır. Hibrit Lineer-Softmax Dikkati...","url":"https://www.aioga.com/tr/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:40:56.298Z"},"vi":{"title":"SANA-Video 2.0: Kết hợp chú ý tuyến tính và phần dư chú ý để thực hiện tạo video hiệu quả","summary":"SANA-Video 2.0 là một Transformer khuếch tán video hỗn hợp, cung cấp hai quy mô là 5B và 14B, có thể tạo video lên đến 720p trên một GPU duy nhất. Chú ý Hỗn hợp Tuyến tính-Softmax của nó trộn tuyến tính và chú ý softmax theo tỷ lệ 3:1, kết hợp với Khối Dư Chú ý để nâng cấp bậc hiệu quả sâu khoảng 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Kết hợp chú ý tuyến tính và phần dư chú ý để thực hiện tạo video hiệu quả - Tin tức AI Aioga","description":"SANA-Video 2.0 là một Transformer khuếch tán video hỗn hợp, cung cấp hai quy mô là 5B và 14B, có thể tạo video lên đến 720p trên một GPU duy nhất. Chú ý Hỗn hợp Tuyến tính-Softmax...","url":"https://www.aioga.com/vi/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:40:51.576Z"},"id":{"title":"SANA-Video 2.0: Menggabungkan Perhatian Linear dan Sisa Perhatian untuk Meningkatkan Efisiensi Generasi Video","summary":"SANA-Video 2.0 adalah sebuah Transformer difusi video hybrid, tersedia dalam dua skala yaitu 5B dan 14B, dapat menghasilkan video hingga 720p pada satu GPU. Hybrid Linear-Softmax Attention-nya mencampur perhatian linier dan softmax dalam rasio 3:1, dan dengan Block Attention Residuals meningkatkan peringkat efektif lapisan dalam sekitar 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Menggabungkan Perhatian Linear dan Sisa Perhatian untuk Meningkatkan Efisiensi Generasi Video - Berita AI Aioga","description":"SANA-Video 2.0 adalah sebuah Transformer difusi video hybrid, tersedia dalam dua skala yaitu 5B dan 14B, dapat menghasilkan video hingga 720p pada satu GPU. Hybrid Linear-Softmax A...","url":"https://www.aioga.com/id/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:41:33.572Z"},"th":{"title":"SANA-Video 2.0: การรวมความสนใจเชิงเส้นแบบผสมกับเศษเหลือความสนใจเพื่อให้เกิดการสร้างวิดีโอที่มีประสิทธิภาพ","summary":"SANA-Video 2.0 เป็น Transformer กระจายวิดีโอแบบผสมผสาน ซึ่งมีขนาด 5B และ 14B สามารถสร้างวิดีโอความละเอียดสูงสุด 720p บน GPU เดียว Attention แบบ Hybrid Linear-Softmax ของมันผสมระหว่าง linear และ softmax ในอัตราส่วน 3:1 ร่วมกับ Block Attention Residuals ทำให้ rank ที่มีประสิทธิภาพในชั้นลึกเพิ่มขึ้นประมาณ 12%","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: การรวมความสนใจเชิงเส้นแบบผสมกับเศษเหลือความสนใจเพื่อให้เกิดการสร้างวิดีโอที่มีประสิทธิภาพ - ข่าว AI Aioga","description":"SANA-Video 2.0 เป็น Transformer กระจายวิดีโอแบบผสมผสาน ซึ่งมีขนาด 5B และ 14B สามารถสร้างวิดีโอความละเอียดสูงสุด 720p บน GPU เดียว Attention แบบ Hybrid Linear-Softmax ของมันผสมระหว่...","url":"https://www.aioga.com/th/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:41:39.728Z"},"pl":{"title":"SANA-Video 2.0: Mieszane liniowe uwagi i reszty uwag umożliwiające wydajną generację wideo","summary":"SANA-Video 2.0 to mieszany transformator dyfuzji wideo, oferujący dwa rozmiary: 5B i 14B, mogący generować wideo w maksymalnej rozdzielczości 720p na pojedynczym GPU. Jego Hybrydowa Uwaga Liniowo-Softmaxowa miesza uwagę liniową i softmax w stosunku 3:1, a w połączeniu z Resztami Blokowej Uwagi zwiększa efektywny rząd głębokiej warstwy o około 12%.","category":"模型更新","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"SANA-Video 2.0: Mieszane liniowe uwagi i reszty uwag umożliwiające wydajną generację wideo - Aioga Wiadomości AI","description":"SANA-Video 2.0 to mieszany transformator dyfuzji wideo, oferujący dwa rozmiary: 5B i 14B, mogący generować wideo w maksymalnej rozdzielczości 720p na pojedynczym GPU. Jego Hybrydow...","url":"https://www.aioga.com/pl/news/cmryvpwy905kerolgy47s0xdn/","contentTranslated":true,"sourceHash":"8535efad877fd82e","translatedAt":"2026-07-26T01:42:24.727Z"}}}}