Google DeepMind takes a step forward in AI video creation with the launch of I see 2 to openly compete with Sora. Accessible via the experimental tool VideoFXthe tool is already showing promising results despite temporary constraints. Thanks to advanced techniques and artistic collaborations, DeepMind seems to want to position itself as a leader in this rapidly expanding field!
A new advance in video generation
The DeepMind AI research laboratory has just reached an important milestone with I see 2a revolutionary model of video generation. This new version is capable of producing clips of more than 120 seconds in resolution 4k (4096 x 2160 pixels). An exploit that far exceeds the capabilities of OpenAI’s Sora, limited to 20 second videos and a maximum resolution of 1080p.
For now, However, these impressive capabilities only represent theoretical potential.. Indeed, Videos created with Veo 2 on VideoFX are currently limited to 8 seconds at 720p resolution.
Despite these constraints, Google plans to open access to this tool to more users in the coming weeks. In addition, the mountain view firm will integrate Veo 2 into its development platform Vertex AI as soon as it is ready for large-scale use!
Improved capabilities and promising results
It must be said that Veo 2 stands out for its better understanding of physicscamera controls and textures. In tests, this generator produced videos with sharper images, even in complex or very animated scenes. It allows you to generate videos fromtext prompts ortext-image associations.
Furthermore, it offers extensive control over perspectives, camera movements and cinematographic effects. According to Eli Collins, VP of Product at DeepMind, the model excels in modeling realistic phenomena such as fluid dynamics. He admits, however, that some they persist in defianceparticularly in terms of character coherence and details in fast or complex scenes.
To refine its models, DeepMind collaborates with renowned artists and creators, including Donald Glover et The Weeknd. These partnerships aim to integrate user feedback from the earliest phases of development.
ALSO READ: Meta asks California to block OpenAI’s transition to for-profit status
Ethical and technical issues
Like many generative models, Veo 2 is based on intensive training from videos available online. DeepMind remains vague on the sources used, but several experts believe that it is based more on the works available on YouTube. This approach therefore raises questions about respect for copyright… For good reason, creators do not have the possibility of remove their works existing training sets.
DeepMind nevertheless defends the legitimacy of these practices by invoking theusage equitable. That said, this position is not unanimous… Even more so with growing concerns about the impact of AI on creative jobs. To limit abuse, SynthID technology integrates invisible digital anchors into generated videos even if this solution is not completely foolproof. At the same time, filtering measures are applied to avoid sensitive content, such as violent or explicit scenes.
Additionally, Google’s compensation policy for copyright disputes will only apply to Veo 2 once the model is widely available. Aside from Veo 2, DeepMind also announced improvements to the image model Image 3 which is now integrated into the ImageFX tool. These updates promise more detailed renderings with better loyalty to text invitations and an optimized interface for users.
Source : TechCrunch