Google于10月27日发布新一代人工智能世界模型 Genie 4，进一步提升了AI生成互动环境的真实感，并加入实时修改场景和自动生成音频等能力。与传统的图像生成模型主要输出静态图片或视频不同，Genie 4能够创建用户可以参与其中、进行交互的虚拟世界，相关技术有望应用于游戏开发、影视制作以及机器人训练等领域。
Genie 4此次主要带来四项升级：生成更加逼真的虚拟环境、允许用户通过自然语言提示实时改变场景、为生成的世界自动添加音频，以及采用GoogleGemini 3.8 Flash模型辅助模拟过程。此外，新模型能够生成最长约两分钟的互动体验，是上一代Genie 3所支持时长的一倍。
世界模型是近年来人工智能研究的重要方向之一。传统生成式AI通常根据文本提示生成图像、视频或其他内容，而世界模型更进一步，尝试学习环境中的空间关系、物体运动规律以及动作与结果之间的联系。模型生成的内容不再只是供用户观看，而是能够随着用户的操作持续变化。
例如，在一个由AI生成的虚拟环境中，用户可以控制角色移动、改变观察方向，或者与场景中的物体进行互动。模型需要根据这些操作持续生成后续画面，并尽可能保持环境的连贯性。这类能力对于游戏、机器人模拟和复杂环境训练具有潜在价值。
Genie 4的一项重要改进，是用户能够在体验过程中通过自然语言指令修改虚拟世界。用户不必退出当前场景、重新编写提示词并从头生成，而是可以直接要求模型改变环境中的某些元素，从而快速调整场景设计。
这一功能可能对游戏开发者和影视制作团队尤其有用。在正式投入大量资源制作之前，创作者可以利用AI生成初步的场景原型，尝试不同的环境布局、视觉风格和互动方式，再决定哪些方案值得进一步开发。
对于电影制作而言，这类技术还可能用于前期视觉预演。制作团队可以快速构建虚拟场景，探索镜头运动、人物行动路线和环境效果，而不必立即搭建真实布景或制作复杂的三维资产。
Genie 4还加入了音频生成能力。此前的世界模型主要集中于视觉环境和交互过程，而新模型能够为生成的场景配上声音，使虚拟体验更加完整。音频可以帮助模拟环境氛围，也可能为游戏原型提供初步的声音效果。
不过，自动生成音频并不意味着模型已经能够完整取代专业音效制作。对于需要精确控制节奏、空间定位、人物对白以及声音与动作同步关系的项目，仍然需要进一步的人工调整和专业工具支持。
Google还表示，Genie 4能够用于创建机器人训练场景。机器人在真实世界中学习和测试动作，往往涉及设备成本、场地限制以及潜在的安全风险。如果能够通过世界模型生成可交互的模拟环境，研究人员就可以在虚拟场景中测试机器人如何移动、抓取物体以及应对不同环境条件。
这类模拟能力的关键，不只是让虚拟场景看起来逼真，还要尽可能准确地呈现物体的空间关系、运动规律和动作后果。只有当模拟环境能够较可靠地反映现实世界中的物理约束时，机器人在虚拟环境中学到的策略才更有可能迁移到真实设备上。
在技术架构方面，Genie 4采用Gemini 3.8 Flash协助模拟过程。Google希望借助这一模型的能力，改善生成环境中的逻辑连贯性和互动表现。不过，公司并未在此次报道中披露足够详细的技术指标，因此目前还无法准确衡量其相较上一代模型的整体性能提升幅度。
更长的生成时长也是此次更新的重要变化。Genie 3生成的互动体验最长约一分钟，而Genie 4将这一时间延长至约两分钟。虽然两分钟对于完整游戏或复杂模拟来说仍然相当有限，但对于快速原型设计和短时互动体验而言，已经能够提供更多探索空间。
与此同时，Google也承认，随着模拟持续时间延长，虚拟环境仍可能出现不一致的问题。例如，场景中的细节可能逐渐发生变化，某些物体也未必能够始终保持原有的位置和状态。这涉及世界模型的一项核心难题：模型不仅需要生成看起来合理的下一帧画面，还必须在较长时间内维持一个稳定、连续的环境。
如果一个物体在用户离开视野后消失，或者重新进入场景时突然改变位置，即使画面本身足够逼真，整个体验仍会显得不可靠。因此，物体持久性、空间记忆和长期状态一致性，仍是世界模型需要继续改进的关键能力。
在发布安排方面，Genie 4目前首先以研究预览版形式向经过筛选的测试者开放。Google表示，未来将向Google AI Ultra订阅用户推出，并已经开放申请等候名单，允许用户申请通过Project Genie体验新模型。
目前无法独立判断新模型与上一代Genie 3相比究竟进步多大。现阶段公开的演示主要来自Google官方，实际用户体验、长时间交互表现以及模型在复杂场景中的稳定性，仍有待后续测试验证。
从行业发展方向来看，Genie 4代表着生成式人工智能正在从内容创作向环境模拟和交互式体验扩展。未来的AI系统不仅可能根据一句提示生成一张图片或一段视频，还可能直接构建一个能够探索、修改和交互的虚拟空间。
不过，要让这种技术真正成为游戏开发、机器人训练或影视制作中的实用工具，仍需要解决长期一致性、物理模拟精度、交互响应速度以及生成成本等问题。Genie 4目前更像是这一发展方向上的重要实验性进展，而非已经能够取代传统游戏引擎或专业三维制作工具的成熟产品。
在 Project Genie 中申请 Genie 4 访问权限：
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Google, on October 27, publish, a new generation of artificially intelligent world models, Genie 4, further enhanced the authenticity of AI ' s interactive environment and added real-time modification of scenes and automatic audio generation. Unlike traditional image-generation models, where the main output of static pictures or videos is static, Genie 4 is able to create a virtual world where users can participate and interact with each other, technologies are expected to be applied in areas such as game development, video production and robotic training.
Geneie 4 brought four major upgrades: to create a more robust virtual environment, to allow users to change the scenes in real time through natural language tips, to add automatic audio to the world that was generated, and to use GoogleGemini 3.8 Flash to aid the simulation process. In addition, the new model can generate up to a maximum of two minutes of interactive experience, double the length of the previous generation of Geneie 3.
World models are one of the important directions of artificial intelligence research in recent years. Traditionally generated AI usually generates images, videos or other content based on text tips, while the world model goes further, trying to learn about space relations in the environment, the behaviour of objects and the connection between actions and results. Model-generated content is no longer simply for viewing by the user, but can change as the user operates.
For example, in an AI-generated virtual environment, users can control the movement of roles, change the direction of observations or interact with objects in the scene. The model needs to generate follow-up images on a continuous basis and, to the extent possible, to maintain environmental consistency. Such capabilities are potentially valuable for games, robotic simulation and complex environmental training.
An important improvement in Genee4 is the ability of users to modify the virtual world through natural language instructions in the course of experience. Instead of exiting the current scene, rewriting prompt and generating it from scratch, the user can directly require the model to change certain elements in the environment, thereby quickly adjusting the scene design.
This function may be particularly useful for game developers and video production teams. Before officially investing in large-scale production, creators can use AI to generate a prototype of the initial scene, try different environmental layouts, visual styles and ways of interaction, and then decide which options merit further development.
For film production, such technologies may also be used in front of visual rehearsals. Production teams can quickly build virtual scenes and explore the shot movement, the man's course of action and environmental effects without immediately setting up real scenes or producing complex three-dimensional assets.
Genie 4 also added audio generation. Previously, the world model had focused mainly on the visual environment and interactive processes, while the new model had been able to match the resulting scenes with sound, making virtual experience more complete. Audio can help to simulate the environment and may also provide the initial acoustic effect for the prototype.
However, the automatic generation of audio does not mean that the model has been able to replace professional sound production in its entirety. Further manual adjustments and professional tool support are still required for projects that require precise control of the rhythm, spatial positioning, character-to-person whiteness and synchronization of sound and action.
Google also indicated that Genie 4 could be used to create robotic training scenarios. Robots learn and test their movements in the real world, often involving equipment costs, site limitations and potential safety risks. If interactive simulation environments can be created through world models, researchers can test robots in virtual scenarios how to move, capture objects and respond to different environmental conditions.
The key to such simulation capabilities is not only to make the virtual scene look real, but also to present as accurately as possible the space relationship, the pattern of motion and the consequences of movement of the object. Only when the simulated environment reflects more reliably the physical constraints in the real world can robotics learn more likely to migrate to real equipment in a virtual environment.
In terms of technical architecture, Genie 4 uses Gemini 3.8 Flash to assist in the simulation process. Google hopes to build on the capabilities of the model to improve logical coherence and interaction in the generation environment. However, the company did not disclose sufficiently detailed technical indicators in the present report to be able to measure accurately its overall performance improvement compared to the previous generation of models.
The longer generation is also an important change in this update. Geneie 3 generates an interactive experience of up to one minute, while Geneie 4 extends this period to about two minutes. While two minutes are still quite limited for complete games or complex simulations, more exploration space is available for fast prototype design and short-term interactive experience.
At the same time, Google acknowledges that there may still be inconsistencies in the virtual environment as the simulation lasts longer. For example, the particulars in the scene may gradually change, and certain objects may not always remain in their original position and state. This relates to the core challenge of the world model: not only does the model need to generate a seemingly reasonable picture but it must also maintain a stable and continuous environment over a longer period of time.
If an object disappears after the user has left the view or suddenly changes its location when re-entering the scene, the whole experience will appear unreliable even if the picture itself is sufficiently authentic. Therefore, the persistence of objects, spatial memory and long-term state consistency remain key capabilities for continued improvement in world models.
In terms of publish arrangements, Genie 4 is currently open to selected testers first in the form of a research preview. Google indicated that in the future it would be available to Google AI Ultra subscribers and had opened the waiting list to allow users to apply for new models through Project Genie.
It is not possible to judge independently how far the new model has progressed compared to the previous Genie 3. The public presentation at this stage, mainly from Google official, has practical user experience, long-term interaction and the stability of the model in complex scenarios, which is still to be tested for subsequent testing.
In terms of the direction of industry, Genie 4 represents an expansion of the generation of artificial intelligence from content creation to environmental simulation and interactive experience. The future AI system may not only generate a picture or video based on a hint, but may also directly construct a virtual space that can be explored, modified and interactive.
However, to make such technology a practical tool in game development, robotic training or video production, long-term consistency, physical simulation precision, interactive response speed and generation costs still need to be addressed. Genie 4 is now more like an important experimental development in this direction than a mature product that can replace a traditional game engine or a professional three-dimensional production tool.
Application in Project Genee 4 for access to permission:
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.