文 | 伯虎财经，作者 | 秦三钱
10月26日，即梦AI上线短剧工作室，新增了一项颇具传统影视制作色彩的功能：3D导演台。
创作者可以调整机位、设计运镜，对分镜进行预演。与此同时，单镜头定点重拍、声音重构、角色与场景资产管理等功能也被放进同一个工作台。
在传统影视工业中，这些环节由导演、摄影、剪辑等专业岗位协同完成。AI视频工具曾试图通过一句提示词，将这些复杂的岗位职能压缩为一次性的自动化任务，如今却又将核心的控制权重新交还给创作者。
就在该功能发布的两天前，字节跳动内部刚刚完成了一轮AI应用资源的集中收拢。
据36氪报道，剪映旗下的AI短剧创作平台“小云雀”正式并入即梦。尽管保留了独立产品入口且团队保持稳定，但运营资源已全面向即梦倾斜。
今年3月，小云雀推出短剧Agent时，仍主打一键生成成片。半年多后，即梦却开始强调分镜、修改与协作。
从尽一切可能摒弃人工介入，到重新强调分镜、机位与局部修正，字节产品形态的演变，构成了AI视频赛道一个耐人寻味的反差。
这并非AI视频走回头路。对于偶尔尝鲜的用户，一句提示词生成一段短片已经足够；但对于必须按要求修改、按周期交付的制作团队，生成只是漫长流程中的一个环节。
从小云雀到即梦，字节调整的不只是产品架构，更折射出整个AI视频赛道的新命题：当“做出画面”变得门槛极低，谁能将零散的生成能力组织成一套稳定可交付的生产系统？
01 一键成片的边界
在AI视频爆发的早期，“一键成片”几乎是所有产品追求的终极范式。
在这套逻辑里，剧本、角色、分镜和视频都可以由模型自动完成，创作者只要给出需求，小云雀是其中的典型。
小云雀能够一键解析最多10万字剧本，依托Seedance等模型能力，自动化完成角色设定、分镜切分与画面生成。相比传统多模型调用叠加人工拼贴的模式，这套方案极大地缩短了初稿产出时间。在营销短视频、知识解说等高度标准化、对叙事连续性要求较低的场景中，“自动生成”确实展现出了惊人的降本增效能力。
然而，当应用场景从单点短视频延伸至具备连续叙事的剧集时，一键成片的局限性迅速暴露。其核心瓶颈不在于单个镜头的清晰度，而在于跨镜头组合时的逻辑与视觉连续性。
例如，一段15秒的独立画面中，人物动作与光影可能完美无暇；但若置于一部数十集的短剧里，创作者必须确保角色在不同场景中的五官特征、服装道具、空间位置保持高度一致，且镜头衔接符合视听语言。
今年4月，中央广播电视总台中国之声曾走访陕西白泽影视的AI短剧制作团队。当时，这支20人的团队正在制作一部60集、总时长超100分钟的半科幻短剧。制作负责人郑杰艺透露，即使借助AI模型，不同制作人员分工跟进时，仍频繁出现“前一集人物站在场景左边，后一集却无故跳跃至右边”的低级错误。
为解决这些“幻觉”与失控问题，团队不得不花大量精力提前统一步长站位，并在后期进行繁重的补拍与修正。
这些不可控带来的隐性成本，并未随着算力单价的下降而自动消失。传统影视的成本消耗在演员、场地与拍摄周期上；引入AI后，前端拍摄成本被大幅压缩，但反复抽卡却构成了全新的生产成本。
即梦此次增加的定点重拍和3D导演台，正是在试图降低这种不确定性。前者让创作者修改局部镜头，后者将机位规划前置，减少先生成、再发现镜头不符合要求的浪费。
随着市面上第三方AI视频Agent大量涌现，单纯依赖“一键生成”所建立的差异化壁垒已极易被复制。当生成能力成为越来越容易获得的服务，仅靠一次点击完成视频，已经不足以解释一款工具的长期价值。
02 谁能掌握AI视频的生产流程？
当底层的视频生成能力逐步变成行业的基础设施，AI视频平台的竞争焦点悄然发生了转移。
过去，用户挑选工具的核心指标是画质、时长与价格。这种单点能力的竞争导致用户粘性极低，字节系的产品也难免遭遇早先发布的Minimax M3与可灵的冲击。
但在专业影视生产流程中，创作者更多需要的是剧本拆解、角色资源、场景搭建、道具分库、多轨音视频切片等。
这些环节无法依靠一次独立的API调用解决。谁能将这些零碎的节点串联并固化下来，就能成为创作者无法离开的“生产力工作台”。
字节跳动正在顺应这一逻辑加速布局。9月，剪映推出剪映Hub，将无限画布与多轨道编辑器结合。创作者既能在画布里组织剧本、生成素材，也能把结果送入熟悉的剪辑流程。即梦和小云雀提供的生成能力，成为这个工作台中的一环。
以视频剪辑切入的剪映向上游剧本与素材生成延伸，以生成见长即梦则向下游剪辑与资产管理渗透，两者在生产流程的中游汇合。
这种产品演进，也改变了平台之间的竞争尺度。
快手旗下可灵选择从模型能力入手。可灵3.0强化人物一致性、多镜头叙事与精细控制，希望在生成环节就减少角色变化、镜头失控造成的后期修补。当模型更理解导演意图，一部分过去依赖外部工作流处理的问题，可能直接在模型层解决。
LibTV等第三方聚合平台， 侧重于通过多模型聚合、无限画布与Skill（技能）体系打造差异化。通过在不同环节调度最适合的模型（例如用A模型做人设、B模型做动作、C模型做配音），并将成熟的制作经验沉淀为可复用的技能，依然能够创造跨模型的附加价值。
当原生大模型的控制力越来越强，随着大厂产品开始补齐工作流，单纯的“模型聚合+画布界面”是否足以建立足够深的防御护城河？
真正的终极战场，大概率将发生在项目数字资产与制作经验的产品化上。
对于一家长期产出AI短剧的团队而言，专属的角色3D资产、特定的美术风格模型、经过验证的分镜预设与工作流规范，是可以在后续项目中高频复用的资产。一旦这些资产沉淀在某个平台上，团队的迁移成本将变得极其高昂。
与此同时，传统影视导演的拆镜经验、动作控制技巧，正在被转化为可被Agent调用的标准化模板与技能。工具平台的商业模式，也逐渐从卖Token“增值税”，转变为按团队订阅、资产托管与协同溢价的SaaS模式。
值得注意的是，这场控制权争夺战中，内容制作公司并未完全将生产命脉让渡给平台。例如芒果TV已推出自研的 “芒果灵创” 平台，将模型调度、角色资产库与协同制作深度结合。对于头部内容生产方而言，将自身核心的影视制作 Know-how 转化为自主可控的生产系统，远比单纯依赖外部工具更具安全感。
不同定位的AI视频厂商，不可避免地踏入同一条河流，介入作品从立项、制作到最终交付的全生命周期。
03 专业化之后，低质供给会出清吗？
生产力的大跃进还在持续，生产工具也得以在更长体量、更高规格的影视作品中接受真实市场的检验。
今年8月底，AIGC长剧《后西游记》登录芒果TV与湖南卫视。该剧第一季共30集，单集时长约40分钟，全剧画面与人物演绎均由AI辅助完成。面对庞大的工业化体量，芒果灵创为此生成了109个深度绑定的人物资产与143个场景资产，极大地压缩了制作周期与制片成本。
如果说长剧集验证了AI在流媒体端承接长周期生产的能力，那么院线电影则直接将AI推向了视听消费金字塔尖的考场。
10月23日，国内首部获公映许可证的AI全流程院线电影《三星堆：未来往事》正式上映。这部100分钟的电影历时两年多制作，近百人团队参照传统电影工业设置了17个专业职能岗位，迭代了超120万张底图。在千万级的制作投入下，AI展现出了在视听大片制作上极其惊人的成本优势。
然而，技术层面的破局，并没有无缝转化为票房上的商业奇迹。
猫眼专业版数据显示，截至10月28日晚，该片上映6天累计票房仅约为137万元（不含点映）。这一成绩与前期庞大的研发与制作投入之间，形成了极为显眼的落差。
这一现象为整个AI影视热潮敲响了警钟。技术可以帮创作者将想象力带入预算许可的范围，可以免去庞大的实景搭建与光影调度，但它无法替创作者撰写出打动人心的故事，更无法自动完成与观众的情感共鸣。
从古早的希腊戏剧到如今的AI成片，不论任何时候，内容生产的核心都是对作品的判断。
这种矛盾在供给过剩的短剧市场尤为尖锐。DataEye数据显示，仅2026年9月单月，抖音端原生AI剧与漫剧的新增数量就达到了5.18万部，创下历史新高。但排名前1.46% 的头部新剧，拿走了全网 64.6% 的播放增量。
极低门槛带来的供给爆炸，并没有带来观众注意力的等比例增加。相反，内容市场陷入了更剧烈的马太效应。
专业工具反而可能同时推动两种趋势。一方面，低成本、标准化的短剧继续依靠批量制作和题材试错争夺流量；另一方面，依靠AI降低特效与工业门槛，将省下的成本赋能给编剧、剧本叙事与美学表达，尝试去冲击电影、长剧等高溢价市场。
这两种路线并不互斥，但对工具提出了截然不同的要求。自动化要的是极致效率，而品质交付则需要专业化工作流的支撑。
在这场分化中，工具平台也面临一道选择题，是服务更大规模的低成本制作，还是服务项目数量较少、但管理要求更复杂的专业团队。
回看字节将小云雀整合进即梦的举措，其本质并非否定“一键成片”的价值，而是试图将原本孤立的“自动生成能力”，塞进更为庞大、严密的影视工业体系之中。
生成能力推倒了创作的围墙，但作品真正的价值，依然要在围墙之外的市场上重新被证明。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Qin's money.
On 26 October, Jimeng AI opened the short theatre studio and added a more traditional video production feature: 3D director.
The creators can reset the plane, design the mirrors and pre-run storyboard. At the same time, functions such as single shot spot reshooting, voice re-engineering, role and scene asset management were placed in the same workstation.
In the traditional video industry, these are carried out in collaboration with professional positions such as directors, photography and editing. The AI video tool, which tried to compress these complex job functions into a one-time automated task by means of a prompt, now gives control of the core back to the creator.
Just two days before publish, the byte beat had just completed a round of centralization of AI application resources.
According to Krypton, CapCut's platform for the creation of the Ai short play, the Little Skylark, was officially incorporated into Jimeng. Despite the maintenance of an independent product portal and the stability of the team, the operational resources have been fully tilted towards Jimeng.
In March this year, when Little Skylark launched Agent's short play, it was still the main key to produce a film. More than six months later, Jimeng began to emphasize storyboard, modify and collaborate.
From doing everything possible to abandon manual intervention to re-emphasizing storyboard, the location and local modifications, the evolution of the byte product pattern constitutes an interesting contrast in the AI video track.
This isn't AI's turn back. For occasional users, it is sufficient that a short piece of prompt is generated; however, for production teams that have to be modified as required and delivered on a cycle basis, the generation is only a link to long pipeline.
From little sparrows to Jimeng, Z113Z adjusts not only to product architecture, but also to reflect the new proposition of the entire AI video track: When the threshold of “making images” becomes extremely low, who can organize fragmented generating capacity into a stable, deliverable production system?
One-clicked boundary.
In the early days of the AI video outbreak, "one-clicked film" was the ultimate paradigm for almost all products.
In this logic, scripts, characters, storyboards and videos can be done automatically by models, and the creators are typical of them if they give a demand.
Small sparrows are able to interpret up to 100,000 scripts in a single key, based on model capabilities such as Seedance, automating role setting, storyboard cut and image generation. The first draft output is significantly shorter than the traditional multi-model model, which calls for supersing and hand-coding. In highly standardized scenarios, such as marketing short video, intellectual discourse and lower requirements for narrative continuity, “auto-generated” does show amazing efficiency gains.
However, when the application scene extends from a single point short video to a series of series with continuous narratives, the limitations of a single-click piece are quickly exposed. The core bottleneck lies not in the clarity of a single shot, but in the logic and visual continuity of a combination of shot.
For example, in a 15-second stand-alone picture, the movement of the person and the light may be perfect; but if placed in a series of dozens of short plays, the creator must ensure that the roles are highly consistent with the five official features in the different scenes, clothing props, space locations and that shot links to audio-visual language.
In April this year, the Central Radio and Television station, the Voice of China, visited the Ai short play production team in Shaan Baizawa. At that time, the team of 20 people was producing a semi-sci-fi drama of 60 episodes lasting over 100 minutes. The chief producer, Zheng Jiaxuan, revealed that even when following up with the AI model, different producers were following up on the division of labour, there were frequent low-level errors in “the former set of people standing on the left side of the scene, while the latter series jumped to the right without cause”.
In order to address these “fantasy” and uncontrolled issues, the team had to devote considerable effort to harmonizing long steps ahead of time and to carrying out heavy catch-up and correction at post-production.
The hidden costs of these uncontrollable costs do not automatically disappear as the unit price of the calculator falls. The cost of traditional video is consumed by actors, the site and the shoot cycle; after the introduction of AI, the cost of front-end filming has been significantly reduced, but repeated cards constitute entirely new production costs.
Jimeng is trying to reduce this uncertainty precisely by reshooting this additional target and 3D director. The former allowed the creators to modify the local shot, while the latter pre-empted the scheduling of the slots to reduce the waste of Mr. S.S. and the discovery of the non-conformity of shot.
With the emergence of third-party AI video Agent on the market, the differential barriers created by reliance solely on “one-key generation” are easily replicated. When generating capacity becomes an increasingly accessible service, a single click to complete the video is no longer sufficient to explain the long-term value of a tool.
02 Who has access to the production process of AI video?
As the bottom-up video generation capacity is gradually becoming the infrastructure of the industry, the competition focus of the AI video platform is shifting.
In the past, the core indicators of the user selection tool were graphic quality, length and price. This competition for single-point capacity has led to extremely low user viscosity, and products from the byte system have inevitably been hit by earlier Minimax M3 of publish and Kling.
However, in the professional video production process, creators are more likely to need script dismantling, role resources, set-up of scenes, prop library, multitrack video slices, etc.
These links cannot be solved by an independent API call. Those who are able to tie and entrench these fractional nodes can become “productivity workshops” from which creators are unable to leave.
byte beats are following this logic of acceleration. In September, CapCut launched CapCutHub, combining an unlimited canvas with a multi-orbit editor. The creator is able to draw a screenplay, generate asset, and send the results into the familiar editing process. The generation capacity provided by Jimeng and the Lark became part of this platform.
The CapCut, which is cut in video clips, is extended to the upstream scripts and asset to produce the long Jimeng, which penetrates downstream clips and asset management, and which converge in the middle of the production process.
This evolution of products has also changed the scale of competition between platforms.
Kling under the Fast Hand flag chooses to start with model capabilities. Kling3.0 strengthens the consistency of character, multi-shot narratives and precision control, hoping to reduce role change at the production chain and to repair post-production as a result of shot ' s loss of control. When the model better understands the director ' s intentions, some of the issues that were previously dependent on external workflow may be addressed directly at the model level.
A third-party aggregation platform such as LibTV focuses on creating differences through multi-model aggregation, unlimited canvass and Skill systems. The value added of cross-models can still be created through the most appropriate models at different stages (e.g., man-made with model A, action with model B, voiceover with model C) and by settling mature production experience into reusable skills.
As the original LLM becomes more and more controlled, and as large plant products begin to complete workflow, is a simple “model + canvas interface” sufficient to create a sufficiently deep defensive moat?
The real ultimate battleground would probably be the product of project digital assets and production experience.
For a team with a long-term AI short play, exclusive role 3D assets, specific art style models, certified storyboard presets and workflow specifications are assets that can be reused by high frequency in follow-up projects. Once these assets are deposited on a platform, the costs of team relocation will become extremely high.
At the same time, traditional video directors ' mirror-breaking experience and action control techniques are being converted into standardized templates and skills that can be used by Agent. The business model of the tool platform has also evolved from selling Token's “value added tax” to the SaaS model based on group subscriptions, asset hosting and synergistic premiums.
It is noteworthy that during this control battle, the content-producing companies did not fully cede the lifeline to the platform. Mango TV, for example, has launched a self-researched Mango Spring platform that combines model movements, role asset banks and collaborative production depths. For head content producers, transforming their own core video production, Know-how, into an autonomous, controlled production system is much safer than relying solely on external tools.
Different AI video manufacturers inevitably enter the same river and intervene in the entire life cycle of their work, from project creation to production to final delivery.
After professionalization, will low-quality supplies come out?
The productivity leap is continuing, and production tools are being tested for real markets in longer-sized, higher-specified visual works.
At the end of August this year, AIGC's chronicle of the West was posted on Mango TV and Hunan Watch. The first season of the show consisted of 30 episodes, each lasting about 40 minutes, and the full picture and character were completed by AI. In the face of enormous industrialization, Mango Spirit generated 109 deep-seated person assets and 143 scene assets, significantly reducing production cycles and production costs.
If the long series validates the ability of AI to deliver a long cycle of production at the end of streaming media, the cinema films directly push AI to the top of the audio-visual consumption pyramid.
On 23 October, the country ' s first public screening of the full-flow-basket film, The Three Stars: The Past, was officially launched. The 100-minute film has been produced over two years, and nearly 100 people have set up 17 professional jobs in line with the traditional film industry, with over 1.2 million background maps. With tens of millions of levels of production input, AI has demonstrated an amazing cost advantage in large audio-visual productions.
However, the breakdown at the technical level has not been seamlessly translated into a commercial miracle on the ballot box.
Data from the cateye professional edition show that, as of the night of 28 October, the 6-day cumulative ticket room was only about $1.37 million (excluding the spot). This achievement contrasts significantly with the large research and development and production inputs of the preceding period.
This phenomenon is a wake-up call for the whole Ai video boom. Technology can help creators to bring imagination into the range of budget permissions, and it can avoid massive landscape-building and light movement, but it cannot write inspiring stories for creators, much less automatically completing emotional resonance with audiences.
From the ancient Greek drama to today's AI film, the core of content production is the judgement of the work at all times.
This contradiction is particularly acute in the short-time market for excess supplies. Data from DataEye show that in September 2026 alone, the number of new Douyin progens and comics reached 5.18 million, a record high. But the top 1.46 percent of the top one, took off the Internet 64.6 percent of the increase.
The supply explosion caused by the very low threshold has not been matched by an increase in audience attention. On the contrary, the content market has been caught up in a much more intense horse-to-mart effect.
Rather, professional tools may drive both trends simultaneously. On the one hand, low-cost, standardized short plays continue to rely on bulk production and subject-matter trials to compete for traffic; on the other hand, AI has been used to lower special effects and industrial thresholds, to give saved costs to writers, script narratives and aesthetic expressions, and to try to shock high-priced markets such as movies, long plays, etc.
The two routes are not mutually exclusive, but they place very different demands on tools. Automation requires extreme efficiency, while quality delivery requires support from specialized workflow.
In this fragmentation, the tool platform also faces the question of whether to serve a larger, low-cost production or a more complex team of professionals with fewer service lines but more complex management requirements.
Looking back at byte's initiative to integrate small sparrows into Jimeng, the essence of the initiative is not to deny the value of “one-click-in-one” but to attempt to insert the previously isolated “automated generation capacity” into a larger and more sophisticated video industry system.
The capacity to generate has pushed down the wall of creation, but the true value of the work remains to be re-proven in the market outside the wall.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.