Google DeepMind has lifted the veil on Veo 2, its latest artificial intelligence model dedicated to video generation. Presented as a greatly improved version of its predecessor (Veo, therefore), this new tool is capable of creating videos of more than two minutes in 4K resolution. For comparison, its competitor Sora from OpenAI is limited to 20-second clips at 1080p.
Always further in realism
But these performances remain theoretical for the moment. The VideoFX tool, where Veo 2 is currently in testing, limits creation to 8-second videos at 720p. Google assures that access to this tool will be expanded in the coming days, and that Veo 2 will also be integrated into other products, such as its Vertex AI platform, as the technology is ready for large-scale use.
Eli Collins, Vice President of Product at DeepMind, explained: “ We will continue to improve Veo 2 based on user feedback and plan to share more new features next year.»
Beyond the impressive numbers, Veo 2 stands out for its technical capabilities. DeepMind claims that the model better understands the principles of physics, camera movements, and properties of light. The result: sharper textures, more realistic handling of shadows and reflections, and better modeling of fluids, such as coffee poured into a cup or maple syrup dripping on pancakes.
In terms of style, Veo 2 can produce videos as varied as a Pixar-style animated clip or a scene shot with specific cinematic effects. For example, it is possible to request a “close-up” or “18mm lens” and the model adapts its creations accordingly.
However, not everything is perfect. Despite progress, Veo 2 does not escape certain common flaws in AI models: inconsistencies in complex details, characters that struggle to remain identical over time, or even difficulties with rapid movements.
Eli Collins recognizes these limitations: “ There are still challenges to overcome in terms of consistency and realism. We work with creators to refine our tools. » Among the artists collaborating with DeepMind, we find prestigious names like Donald Glover or The Weeknd.
Like many artificial intelligence-based tools, Veo 2 raises ethical questions. DeepMind remains unclear on the video sources used to train its model, but Google-owned YouTube is a likely candidate. This use of public content without explicit consent from the creators raises eyebrows, particularly in artistic circles.
To limit abuse, Veo 2 integrates watermarking technology, called SynthID, which discreetly marks videos to indicate that they have been generated by AI. However, these protections are not foolproof.
At the same time, DeepMind announced an update to its image model, Imagen 3, which now generates more detailed and stylized images, ranging from realism to Japanese animation. These tools, while certainly promising, show that the future of digital creation remains closely linked to the advances — and limits — of artificial intelligence.
🟣 To not miss any news on the WorldOfSoftware, , .
