[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-seedance-2-5-enterprise-api-b2b":3,"news-related-6ed14a36-a62a-43e8-949a-cf9df4405d98":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"6ed14a36-a62a-43e8-949a-cf9df4405d98","Seedance 2.5 把视频生成送进 B 端:30 张参考图、API 上火山方舟、徐工小鹏首批接入","字节跳动 Seed 团队 7 月 31 日发布 Seedance 2.5,把单次生成时长拉到 30 秒、多模态参考升级到 30 张图\u002F10 段视频\u002F10 段音频,并支持时间戳精准编辑。API 服务即将登陆火山方舟,徐工集团、小鹏汽车、灵初智能、微分智飞、穹彻智能等成为首批接入企业,落地工业制造、具身智能、自动驾驶仿真等 B 端场景。","## 视频生成从「生成片段」走向「完成创作」\n\n7 月 31 日,字节跳动 Seed 团队正式发布新一代视频创作模型 **Seedance 2.5**。相比 2.0 时代把焦点放在「生成一个 15 秒片段」,2.5 把重心放到「完成一段创作」——单次生成时长翻倍到 30 秒、多模态参考上限拉高一个数量级、并且把编辑能力做到了时间戳级别。\n\n更关键的是,**Seedance 2.5 发布当日即接入火山方舟 API,并官宣首批 5 家企业客户**:徐工集团、小鹏汽车、灵初智能、微分智飞、穹彻智能。模型能力被推到了「工业可用」门槛,视频生成模型正式从「演示」走向「生产」。\n\n## 三个核心升级\n\n**(1) 长叙事能力。** 单次 30 秒、多轮延长。模型能在一个片段里组织「铺垫—推进—转折—收尾」的镜头叙事,而不是把同一画面拉长。官方示例里一段 30 秒演唱会短片,从化妆间后台、穿过走廊、与伴舞互动、登上舞台到最终体育馆全景,完整跑完一套电影化的一镜到底。\n\n**(2) 多模态参考能力。** 单次输入上限提升到 **30 张图片 + 10 段视频 + 10 段音频**,用于还原画面构图、场景、风格、人物、道具。多人同框、群像叙事下也能同时稳定多个人物的形象与声音。新增的「白模参考」允许用无纹理 3D 模型先搭出空间结构、主体姿态和镜头机位,模型再按这套结构渲染真实材质——这对工业、汽车、电影预可视化场景是实打实的工作流增益。\n\n**(3) 时间戳级编辑能力。** 在生成阶段,用户能用 prompt 控制某个时间段发生的事件、视角、运镜和节奏;在生成后,可以针对指定片段的角色、动作、声音或剧情做定向修改,同时保持修改前后的连贯性。配合绿幕编辑、视角\u002F运镜编辑、参考编辑,Seedance 2.5 把「成片后期」这一工序压进了生成流程。\n\n## 产业落地的两条主线\n\n字节跳动 Seed 在官方文章里点出了 **Seedance 2.5 在 B 端的两个应用方向**:\n\n- **机器人与具身智能**:模型生成高质量合成视频数据,用于训练机器人的感知和操作能力。灵初智能、穹彻智能、微分智飞正是国内具身智能赛道的主要玩家。\n- **自动驾驶仿真**:模拟极端天气、复杂路况等长尾场景,为系统测试和训练提供更多样本。\n\n加上 **徐工集团** 这类传统工业制造龙头把 Seedance 用于工业仿真、流程培训和设备演示,以及 **小鹏汽车** 把模型接入到内容生产链路(车载营销、宣传物料、车展演示等场景),首批 5 家客户的组合勾勒出了一个清晰的信号:**视频生成模型正在从「消费者玩具」转向「企业生产资料」**。\n\n## API 商业化与版图补全\n\nAPI 服务上线火山方舟之后,Seedance 2.5 与字节既有的即梦 AI、豆包专业版形成「C 端入口 + 工具侧入口 + B 端 API」的三层分发。对标国内另一极的可灵(Kling),Seedance 在 2.5 这一代重点押在了 **「长叙事 + 可控编辑 + B 端产业落地」** 这条差异化路径上,而不是单纯卷时长或分辨率。\n\n火山引擎同步官宣了徐工、小鹏、灵初智能、微分智飞、穹彻智能「发布即接入」的合作节奏。这意味着 **Seedance 2.5 不是「先发模型、等客户」**,而是「客户在等着上线」。\n\n## 行业影响:视频生成进入「工业化」分水岭\n\n视频生成模型在 2025-2026 年经历了三轮迭代:第一轮拼「能不能生成」(Sora 初登场),第二轮拼「够不够长、够不够清」(1080p \u002F 4K \u002F 60 秒),第三轮——也就是 Seedance 2.5 这一代——拼的是 **「能不能直接进生产线」**。\n\n分水岭的标志有三:\n1. **可编辑性**:能精确控制片段级元素,而不是一次性抽奖式生成。\n2. **参考上限**:多模态参考数量足够复杂,能支撑真实创作的工作量。\n3. **B 端落地的具体场景**:不再是「品牌方做了一条 30 秒广告」的零星案例,而是进入机器人训练、自动驾驶仿真这种每天产生大量合成数据需求的生产流程。\n\nSeedance 2.5 在这三点上同时迈过门槛,加上 API 立刻可用、首批 5 家客户覆盖工业\u002F汽车\u002F具身智能三个高价值方向,**对国内视频生成赛道的「工业化时间表」是一次实质推进**。\n\n所以呢:视频生成模型的下半场不是「谁更长、谁更真」,而是「谁能最早嵌入真实生产链路」。Seedance 2.5 在这一步上跑得最快,但市场不会给它太久——可灵、Vidu、智谱、阿里通义万相都会在 B 端 API 和产业落地上加速跟上。","https:\u002F\u002Fmp.weixin.qq.com\u002Fs\u002F_LtI8-8PkyQckW9CPQVw5A","d4a24db7-b6c2-410c-ba99-c16625c61305",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"64c1622c-6473-4ad6-9cb6-01e4802cc279","en","Seedance 2.5 goes enterprise: 30 reference images, Volcano API","ByteDance Seed released Seedance 2.5 on July 31, extending single-shot generation to 30 seconds, expanding multimodal references to 30 images \u002F 10 videos \u002F 10 audio clips, and adding timestamp-precise editing. The API is about to launch on Volcano Ark, and XCMG, XPeng, Lingchu Intelligent, Weifen Zhifei, and Qiongche Intelligent have been named as the first batch of enterprise customers spanning industrial manufacturing, embodied AI, and autonomous driving simulation.","## Video Generation Moves From \"Generating a Clip\" to \"Finishing a Creation\"\n\nOn July 31, ByteDance's Seed team officially released the new-generation video creation model **Seedance 2.5**. Compared with the 2.0 era's focus on \"generating a 15-second clip,\" version 2.5 centers on \"finishing a creation\" — doubling single-shot generation length to 30 seconds, lifting multimodal reference limits by an order of magnitude, and pushing editing precision down to the timestamp level.\n\nMore importantly, **Seedance 2.5 launched on Volcano Ark's API on day one** and announced its first five enterprise customers: XCMG Group, XPeng Motors, Lingchu Intelligent, Weifen Zhifei, and Qiongche Intelligent. Model capability has crossed the \"industrial-usable\" threshold, and video generation models have officially moved from \"demo\" to \"production.\"\n\n## Three Core Upgrades\n\n**(1) Long-form narrative capability.** Single-shot 30 seconds, multi-round extension. The model can organize a \"setup → progression → twist → resolution\" shot narrative within a single clip, rather than stretching one frame. The official demo — a 30-second concert short — runs a full cinematic one-shot from backstage makeup, through the corridor, a meet-cute with backup dancers, the stage entrance, and a final wide of the stadium.\n\n**(2) Multimodal reference capability.** The single-input ceiling is now **30 images + 10 videos + 10 audio clips** to recover composition, scene, style, characters, and props. Even in multi-character group scenes, the model simultaneously stabilizes multiple subjects' appearance and voice. A new \"white-model reference\" feature lets users lay out spatial structure, subject posture, and camera position with untextured 3D models first; the model then renders realistic material on top — a real workflow gain for industrial, automotive, and pre-vis scenarios.\n\n**(3) Timestamp-precise editing capability.** During generation, users can use prompts to control what happens, from which perspective, with what camera motion, in which time segment. After generation, users can perform targeted edits on the role, action, sound, or plot of a specified segment while maintaining coherence before and after the edit. Combined with green-screen editing, viewpoint\u002Fcamera-motion editing, and reference editing, Seedance 2.5 compresses the \"post-production\" stage into the generation pipeline.\n\n## Two Main Lines of Industrial Deployment\n\nByteDance Seed's official post points out **two B-side application directions for Seedance 2.5**:\n\n- **Robotics and embodied AI**: the model generates high-quality synthetic video data for training robot perception and manipulation. Lingchu Intelligent, Qiongche Intelligent, and Weifen Zhifei are key players in China's embodied-AI track.\n- **Autonomous driving simulation**: simulating extreme weather, complex road conditions, and other long-tail scenarios to provide more samples for system testing and training.\n\nAdd **XCMG Group** — a traditional industrial manufacturing leader using Seedance for industrial simulation, process training, and equipment demos — plus **XPeng Motors** integrating the model into its content-production chain (in-car marketing, promotional assets, auto-show demos, etc.). The combination of the first five customers sketches a clear signal: **video generation models are shifting from \"consumer toy\" to \"enterprise production material.\"**\n\n## API Commercialization and Map Completion\n\nWith the API on Volcano Ark, Seedance 2.5 forms a three-tier distribution with ByteDance's existing Jimeng AI and Doubao Pro: **C-end entry point + tool-side entry point + B-end API**. Aiming at domestic rival Kling, Seedance 2.5 doubles down on the differentiation path of **\"long narrative + controllable editing + B-side industrial deployment\"** rather than competing purely on length or resolution.\n\nVolcano Engine simultaneously announced the \"launch-day integration\" pace with XCMG, XPeng, Lingchu, Weifen Zhifei, and Qiongche. This means **Seedance 2.5 isn't \"ship the model, then wait for customers\"** — it is \"customers waiting for go-live.\"\n\n## Industry Impact: Video Generation Crosses an \"Industrialization\" Watershed\n\nVideo generation models went through three rounds of iteration in 2025-2026: round one was \"can it generate at all\" (Sora's debut), round two was \"is it long enough, sharp enough\" (1080p \u002F 4K \u002F 60 seconds), and round three — the Seedance 2.5 generation — is **\"can it go directly onto the production line.\"**\n\nThree signs mark the watershed:\n1. **Editability**: precise control over segment-level elements instead of lottery-style generation.\n2. **Reference ceiling**: enough multimodal references to support real creative workloads.\n3. **Concrete B-side scenarios**: not isolated cases like \"a brand made a 30-second ad\" but entry into daily high-volume synthetic-data workflows such as robot training and autonomous-driving simulation.\n\nSeedance 2.5 crosses all three thresholds at once. Combined with the immediately-available API and first-customer coverage of three high-value directions — industrial, automotive, and embodied AI — **this is a substantive push on the \"industrialization timeline\" for the domestic video-generation track.**\n\nSo what: the second half of video generation isn't \"who's longer, who's more realistic\" — it's \"who embeds in real production loops first.\" Seedance 2.5 is the fastest on this step, but the market won't give it much runway — Kling, Vidu, Zhipu, and Alibaba Tongyi Wanxiang will all accelerate on B-side APIs and industrial deployment.","seedance-2-5-enterprise-api-b2b","2026-08-01T04:30:00Z","2026-08-01T04:03:46.499557Z","2026-08-01T04:03:46.499566Z",true,"agent",207,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"2fbfa6c5-3bf5-4279-a353-6324396b2d36","字节 Seedance 2.5 把单段视频拉到 30 秒：视频生成终于\"能用\"了？","bytedance-seedance-2-5-30s-video-model","2026-07-31T06:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"f0ea091f-d030-4a32-827a-ae21c1b61c8f","昆仑万维 WAIC 大会将一次性放出四款全模态模型：Matrix-3.5 把\"状态-动作\"塞进一套参数","kunlun-waic-matrix-3-5","2026-07-17T12:08:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"b4214f43-353e-42e3-b48e-92dd4fc64290","京东开源 EchoWM 全模态世界模型:720p 音画同步,能跟着你走","jd-echowm-omnimodal-world-model","2026-08-25T23:10:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"6f9e9f94-9dcc-4c6c-b254-6c5d0fe8ed37","京东开源 JoyAI-Video-Edit:16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-realtime-diffusion","2026-08-10T00:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"bcdc10bc-2f08-4c39-8ffa-e7e34041c112","京东开源 JoyAI-Video-Edit:用 16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-real-time-streaming","2026-08-05T03:00:00+00:00"]