7月31日,字节跳动旗下豆包大模型正式发布新一代视频创作模型Seedance 2.5,支持单次生成30秒高质量视频片段,并支持多轮延长生成更长视频。模型正在陆续上线即梦AI、豆包专业版,API服务也将于近期上线火山方舟。
据官方介绍,Seedance 2.5延续了Seedance 2.0统一的多模态音视频联合生成架构,重点突破了长时一致性和多模态控制能力。
核心升级点
第一,30秒连贯+多轮延长。相比上一代,Seedance 2.5单次生成时长显著提升,且支持对话式延长——用户可以基于已生成的视频继续扩展,理论上没有上限。这让短视频创作从"生成片段"变成了"生成完整作品"。
第二,多模态输入能力增强。支持图片、视频、音频等多种输入形式的混合控制,创作者可以用参考图定风格、用参考视频定动作、用音频定节奏。
第三,抖音生态深度整合。生成的视频可以在抖音、剪映、即梦之间无缝流转,直接使用热门BGM库、特效和模板。
"视频生成的竞争不再是'谁能生成更长的视频',而是'谁能把生成能力嵌入用户的真实工作流'。Seedance选择从抖音生态出发向下扎根,Sora选择从技术上限出发向上突破——两条路线,两个终点。"—— Dawn Vision编辑部
AI视频的竞争格局
2026年的AI视频赛道,已经从"比谁能生成"变成了"比谁能用"。Sora依然保持技术标杆地位,但在产品落地和生态整合上,中国公司速度更快。
对独立AI视频创业公司来说,出路越来越清晰:做深垂直场景——电商商品视频、教育课件、广告制作,这些需要的不是最先进的模型,而是最贴合工作流的解决方案。
明天见。
Sources · 参考来源
声明:本文为 Dawn Vision 基于公开信息的二次创作与独立分析,仅供参考。
本文基于 Dawn Vision 认知引擎处理的 6 个源信号生成,经编辑部人工审核。素材来源:IT之家、网易新闻。
相关入库笔记:Seedance 2.5 · 字节跳动 · 豆包 · AI视频 · 30秒生成 · 多模态输入
On July 31, ByteDance's Doubao officially launched Seedance 2.5, a new video creation model supporting 30 seconds of high-quality video per generation with multi-round extension. Rolling out on Jimeng AI and Doubao Pro, with API access coming to Volcano Ark.
Seedance 2.5 retains the unified multimodal audio-video joint generation architecture of 2.0, with breakthroughs in long-sequence consistency and multimodal control.
Key Upgrades
First, 30-second coherence + multi-round extension. Significantly longer single-generation duration, plus conversational extension — users can build on generated videos with no theoretical upper limit.
Second, enhanced multimodal input. Mixed control via images, video, and audio — reference images for style, reference video for motion, audio for rhythm.
Third, deep Douyin ecosystem integration. Seamless flow across Douyin, Jianying, and Jimeng, with direct access to trending BGM libraries, effects, and templates.
"Video generation competition is no longer about 'who makes longer videos' — it's about 'who integrates into real user workflows.' Seedance roots downward through the Douyin ecosystem; Sora pushes upward from the technical frontier. Two routes, two destinations."—— The Dawn Vision Editorial Desk
AI Video Competitive Landscape
In 2026, AI video competition has shifted from "who can generate" to "who delivers utility." Sora maintains its technical benchmark, but Chinese companies move faster on product integration and ecosystem.
For independent AI video startups, the path is clear: go deep in vertical scenarios — e-commerce, education, ad production. These need workflow solutions, not just the latest model.
See you tomorrow.
Sources · 参考来源
声明:本文为 Dawn Vision 基于公开信息的二次创作与独立分析,仅供参考。
Processing 6 source signals with editorial review. Sources: ITHome, NetEase News.
Notes: Seedance 2.5 · ByteDance · Doubao · AI video · 30s generation · multimodal input
视频生成的竞争不再是'谁能生成更长的视频',而是'谁能把生成能力嵌入用户的真实工作流'。Seedance选择从抖音生态出发向下扎根,Sora选择从技术上限出发向上突破——两条路线,两个终点。
—— Dawn Vision编辑部
Video generation competition is no longer about 'who makes longer videos' — it's about 'who integrates into real user workflows.'
—— The Dawn Vision Editorial Desk
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 6 个源信号生成,经编辑部人工审核。素材来源:IT之家、网易新闻。
Processing 6 source signals with editorial review. Sources: ITHome, NetEase News.