Google’s Impressive Second Iteration of Video Models
This is Veo 2, Google’s second iteration, and somehow it manages to surpass every single video model currently available. This includes the recently released Sora 2 from OpenAI. It’s quite surprising considering Google hasn’t had the best track record, but it seems that this December, Google has exceeded everyone’s expectations regarding AI development. We’ve seen update after update showing us that Google is now clearly the industry leader when it comes to AI development.
This marks a historic moment as AI is more competitive than ever. Being able to top the leaderboards in terms of software that is not only the best but the best by a decent margin means that Google is back in the AI game and setting new standards for other industry leaders.
In terms of actual benchmarks, Google’s model outshines others. For example, Meta’s MovieGen, a 1080p video generator, is preferred only around 30% of the time, whereas over 50% of the time, Google’s model takes the lead. Models like Cing 1.5, highly regarded in the creative industry, also fall short. Furthermore, Minx, one of the top models, ranks just 30% for user preference.
Unmatched Physical Simulations
One of the standout aspects of Google’s Veo 2 model is its incredible physics capabilities. Many video generation models struggle with understanding and simulating the physical world at a granular level. Google, however, has cooked up something exceptional. For instance, cutting a tomato with all its subtle movements and changes in texture is rendered beautifully.
Liquids, being highly unpredictable, are another challenging aspect of generative AI systems. Yet, Google’s Veo 2 handles fluid simulations remarkably well. Examples include the pouring of coffee with subtle nuances and the flow of syrup with precise detail. These demonstrate the model’s coherence.
Creative and Unique Use-Cases
Beyond accurate physics, Veo 2can generate creative and unique characters. Imagine a sitcom TV show with potatoes—this showcases the model’s character consistency. Another example involves a car driving at top speed through a road until it reaches a waterfall. The physics in these scenarios are executed with high precision.
Google’s Frontier Image Model: IM3
Google didn’t stop with video models. They’ve also launched their Frontier Model, a text-image model called IM 3. This model outperforms every other model on the leaderboard with an outstanding ELO rating. Google’s mastery in prompt adherence and UI interface is evident, allowing users to alter and control image components easily.
A noteworthy example is a close-up of a man’s eye reflecting garlic bread—an intricate prompt that IM 3 handles with ease. Another creative output includes a photorealistic image of a potato fighting a vampire on the moon, highlighting the model’s flexibility and capability.
Overall, with the release of Veo 2 and IM3, Google has reclaimed its position as a leader in text-to-video and text-to-image technology. Are you more bullish on Google now? I’m certainly looking forward to their future innovations, especially with potential releases in January.
Source: https://www.youtube.com/watch?v=gSypQljcZgM
Comments