内容来源：https://aibusiness.com/generative-ai/runway-releases-gen-4-5-video-model

内容总结：

AI视频生成技术再迎突破，Runway发布Gen-4.5模型

近日，AI初创企业Runway发布了新一代视频生成模型Gen-4.5，该模型在视觉准确性、创意控制及物理细节表现上均有显著提升。用户仅需输入文字描述，即可生成高清视频片段，尤其擅长处理复杂构图、物理运动及角色表情。

尽管Runway在短格式视频生成领域表现突出，主要面向社交媒体等短时内容创作，但其仍面临来自OpenAI的Sora及谷歌Veo等竞品的激烈竞争。行业分析师指出，不同模型已呈现差异化定位：Runway专注于秒级短视频，而谷歌Veo则瞄准分钟级的长视频市场。

随着生成式AI视频技术日趋逼真，其带来的“真实性困境”也引发关注。业内专家指出，AI生成内容与实拍影像的界限日益模糊，可能引发虚假信息传播风险。目前，行业对是否标注AI生成内容尚未形成统一标准，部分游戏公司已就此采取相反立场。

此外，现有AI视频模型仍存在因果逻辑、物体持续性等方面的技术局限，例如“门先开、后压把手”等逻辑错误。分析师认为，尽管技术持续进步，但要在生成长镜头时保持场景一致性，仍需进一步突破。

在生成式AI应用落地的初期，专家建议企业优先关注能提升信息交互体验的场景，并考虑对AI生成内容添加标识，以应对技术带来的伦理与实用挑战。

中文翻译：

由谷歌云赞助
选择你的首个生成式AI应用场景
要开始使用生成式AI，首先应关注能够优化人类信息交互体验的领域。
这种新模型及其他同类技术所能达到的逼真程度，给企业和其他用户带来了一些挑战。
本周一，Runway推出了一款新的视频生成模型，与大型生成式AI厂商的视频及图像生成模型展开竞争。
这家成立于2018年的初创公司表示，Runway Gen-4.5在视觉精度和创意控制方面表现出色。该公司专注于面向视频、媒体和艺术的生成式AI系统，并已取得一定成功。
用户只需通过文字描述所需的具体动作和行为，即可利用该模型生成高清视频。据Runway介绍，该模型在细节构图、物理准确性以及角色表现力方面表现卓越，同时能实现多样化的风格控制和视觉连贯性。
Runway基于英伟达GPU构建该模型，并利用这些芯片进行预训练、后训练及推理。
高德纳分析师阿伦·钱德拉塞卡兰指出，Runway Gen-4.5再次证明了AI模型正在持续进步。然而，这款新模型正面临来自OpenAI视频模型Sora和谷歌Veo 3.1的激烈竞争。
钱德拉塞卡兰表示，尽管Gen-4.5和Veo 3.1都是视频模型，但它们面向不同的受众和应用场景。Runway生成的视频主要应用于社交媒体信息流。
"Runway的应用场景始终是短视频。"他补充道。
而谷歌Veo则瞄准更长篇幅的视频，例如持续数分钟的产品营销视频，而非仅数秒的短片。相较之下，Gen-4.5更适合制作适用于Instagram等平台的秒级社交媒体短视频。
不过，通过新模型，Runway提升了生成物体和角色的能力，在质量与清晰度方面实现了更好的连贯性。
钱德拉塞卡兰认为："他们似乎更专注于如何重现更复杂的视频场景。"
他还提到，该模型生成的现实世界影像有时令人难辨真伪。但这并非Runway独有的现象，许多视频生成模型已发展到难分虚实的地步。
弗雷斯特分析师威廉·麦基翁-怀特指出，真实与虚拟的界限模糊引发了两种不同观点。
"我建议添加免责声明，通常在短片末尾标注'本视频借助AI技术生成'。"他表示。
他注意到游戏公司对AI生成视频持不同立场：例如Epic Games公开表示不主张标注AI生成内容，而另一家游戏公司Valve则支持明确标注。
麦基翁-怀特说："关于企业该如何抉择的争论正在持续。"虽然真实性问题带来了道德困境，但Runway的技术本身也凸显了AI视频生成的一些局限。
例如，该模型在因果推理方面存在缺陷，特别是当结果先于原因出现时——比如门在把手被按下之前就自行开启。另一挑战是物体恒存性问题，即物体可能意外消失或出现。
麦基翁-怀特指出："虽然模型对物体记忆和交互的处理在不断改进，但要实现更持久、更连贯的镜头呈现，现有特定模型仍需完善。"
您可能还喜欢

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The level of realism that this new model and others can achieve poses some challenges for enterprises and other users.
Runway on Monday introduced a new video generation model that competes with video and image-generating models from bigger generative AI vendors.
Runway Gen-4.5 offers a high level of visual accuracy and creative control, according to the 2018 startup, which has found success with generative AI systems aimed at video, media and art.
Users can use the model to generate high-definition videos simply by writing prompts that detail the desired motion and action. The model excels in detailed compositions, physical accuracy, and expressive characters, according to Runway. It can also handle different stylistic control and visual consistency.
Runway built the model on Nvidia GPUs and uses the chips for pre-training, post-training and inference.
Runway Gen-4.5 is another example of how AI models are continually improving, said Arun Chandrasekaran, an analyst at Gartner. However, the new model faces stiff competition from OpenAI's video model, Sora, and Google's Veo 3.1.
Even though both Gen-4.5 and Veo 3.1 are video models, they target different audiences and applications, Chandrasekaran said. Runway videos are mostly used in social media feeds.
"The use case for Runway has always been short videos," Chandrasekaran said.
With Google Veo, meanwhile, the tech giant is targeting longer-form videos, such as product marketing videos that run in minutes, as opposed to those that run for seconds. Gen-4.5 is more suited for Instagram and social media reels that run for seconds rather than minutes.
However, with its new model, Runway has improved its ability to produce objects and characters with greater consistency in quality and clarity.
"It also looks like they are focusing a lot more on how to recreate more complex video scenes," Chandrasekaran said.
He added that the model is also generating footage of the physical world, sometimes making it difficult to distinguish what is real from what is not. This isn’t particular to Runway, though, as many video-generating models have become so advanced that it’s hard to distinguish what is fake or not.
The inability to distinguish between what is real and what is not has led to two different points of view, said William McKeon-White, an analyst at Forrester.
"What I would recommend is a disclaimer, usually at the end of a short to say they're done with the assistance of AI," he said.
He noted that gaming companies using AI-generated videos have recently taken opposing stances. For example, Epic Games has stated that it is in favor of not labeling AI-generated materials. On the other hand, Valve, another video game company, supports labeling AI-generated content.
"There's actively a debate about where organizations fall on this," McKeon-White said. While the realism issue presents a somewhat moral dilemma, Runway's approach itself highlights some limitations of AI video generation technology.
For example, the model has exhibited some problems with causal reasoning, specifically when the effect precedes the cause. One example is a door opening before the handle is pressed. Another challenge is object permanence, when objects disappear or appear unexpectedly.
"While memory and the interactions of objects are still improving, there remains a bit of work when it comes to the specific models being used here to create more persistent or more consistent shots over time," McKeon-White said.
You May Also Like