谷歌最新视频生成模型Veo-2于12月16日发布，展示了令人印象深刻的视频生成能力。尽管在复杂运动和一致性方面仍存在局限性，但其生成的逼真广告令人惊叹，大部分人难以辨别其为AI生成。该广告由一人在三周内制作完成，虽然经过人工编辑和配音，但所有镜头均由AI生成，除了幕后花絮中的人物采访。制作过程中，角色一致性是最大挑战，需要通过快速剪辑和避免复杂场景来弥补。未来，通过技术进步，有望解决角色一致性和复杂场景的难题。

🎬Veo-2模型展示了强大的视频生成能力，尽管在复杂运动和一致性上仍有不足，但已能生成高度逼真的广告。

🎭 角色一致性是制作中的最大挑战，目前只能通过文字描述来控制角色，需要快速剪辑和避免复杂场景来掩盖不一致性。

🗣️ 广告制作中，除了幕后采访部分，所有镜头均由AI生成，配音和导演由人工完成，没有使用AI生成的语音。

💡 解决角色一致性的一个潜在方案是从现有视频角色开始，并在此基础上进行延续生成，但可能存在侵犯知识产权的风险。

⚙️ 未来，AI唇形同步技术有望解决口型不匹配的问题，而复杂场景的生成可能需要借助AI辅助的物理引擎。

Published on January 21, 2025 7:13 PM GMT

Veo-2 is google's latest video-generation model. Released Dec 16th, it's quite impressive! Of course, there are still limitations (available in that previous link), especially w/ more complex movements (e.g. skateboarder & ballerina) and consistency. 

Then we have this, very realistic ad created by one person in ~3 weeks. 

Most people would not be able to tell this is AI-generated (maybe 1/100k people could tell unprompted?). It is still human-edited and uses human voices. Some quick facts from the author: 

How long did it take?: yup, not sure the hours. What I can backtrack is I generated 12 days for the ad, and 4 for the BTS. add editing, sound design, etc. Maybe 3 weeks?All AI?: Every shot is AI except the talking head in the BTS part.Hardest Part: Character consistency. You can only use text now, so you have to make sure the characters stay the same as much to trick the viewer it’s the same person!

Some tricks they employed:

Very quick shots/edits. Notice the transitions are quick, which likely helps with loss of consistency across shots.No AI-generated videos w/ voice (the narrator was from Fiverr, and the "director" is the creator)No complex scenes (e.g. the skateboarder)

For the limitations:

Consistency across scenes & characters: currently the interface only allows production from text. For character consistency, starting from an existing character in video, and continuing from there is one solution.
Google might not allow this though due to easily violating others IP and other concerns.
AI lip sync:  I've seen several impressive demos from research papers these past two years, so I'd guess it's already mostly solved. It'd just need to be integrated in the interface
Looking at this recent Kling AI video though, the results aren't great.
Complex Scenes: More scale and data usually does the trick. Possibly could use AI-assisted physics engines to help render these (or produce mass amounts of data), but that's definitely not my expertise, just speculation.

Discuss