I see 2 and Figure 3 are proof that generative AI has no intention of stopping, and December of this year was loaded with evidence in this sense. Now with these two models, Google DeepMind is trying to revolutionize the creation of visual content, improving the generation of videos and images with such high quality and accuracy that no one has seen before. And from what they’ve shown so far, it looks like they’re on the right track.

On the one hand, Veo 2 takes video generation to a new standardwith longer clips, higher resolution and stunning realism in movement and detail. Image 3, on the other hand, refines image creation, offering impressive fidelity to user guidelines and the versatility to adapt to any artistic style.

Not only these technologies Google DeepMind is consolidated as one of the leaders in this sectorbut they also put the Veo 2 and Imagen 3 in a prime position against rivals like OpenAI’s Sora, which was recently released for most of the world (except for the European Union, you know, commonplace for a while now). ) or tools like MidJourney and DALL-E. what’s the news We will tell you about it below.

I see 2: Video generation reaches 4K and more than 2 minutes

Veo 2 is the next generation of Google DeepMind’s artificial intelligence model designed specifically for video generation. This version represents a significant leap compared to its predecessor and features that make it one of the most advanced models today. Its most notable improvements include higher resolution, longer clip duration, and surprising realism in motion, textures, and detail.

As confirmed by Google, Veo 2 can generate videos longer than 2 minutes with resolutions up to 4K (4096 x 2160 pixels). This represents an important advance over current models such as OpenAI’s Sora, which is currently limited to 20-second clips and 1080p resolution. In its experimental phase through VideoFX, a Google tool where it’s only available on a limited basis, the Veo 2 generates clips of up to 8 seconds in 720p resolution.

The Veo 2’s technical improvements go beyond resolution and durability. Google DeepMind achieved model s a much more accurate understanding of physicswhich allows realistic representation of complex scenes such as fluid movement, falling objects or interactions between elements. In addition, the model has improved the control of the virtual camera, which translates into smoother movements and the ability to capture objects and people from different angles, imitating the cinematographic effects that we see in large productions.

As if that wasn’t enough, Google reports Veo 2 is also capable of generating videos with a wide variety of stylesfrom animations in the purest Pixar style to sequences that seek a hyper-realistic surface. Details such as the behavior of viscous liquids – for example syrup or coffee poured into a cup – and the processing of lights, shadows and reflections with a precision that we have not seen before in models of this brand stand out in the demos. type.

However, the Veo 2 still faces challenges. Despite the advances, the model still struggles with element consistency in longer scenes or complex instructions. The famous “uncanny valley” still present in details such as human expressions, unrealistic eyes or in fast moving sceneswhere visual artifacts or inconsistencies may still occur.

At the moment, Veo 2 is in an experimental phase within VideoFXtool available only to users selected through a waiting list. Google has announced that it plans to offer it at scale in the future through its Vertex AI platform, which will allow developers and industries to use the technology to create innovative visual content.

Figure 3: Sharper details and truer to challenges

Figure 3 is the latest version of the Google DeepMind image generation modeldesigned to create visual compositions with an unprecedented level of detail, precision and versatility. With this update, the company tries to consolidate its position in an increasingly competitive market, where tools like MidJourney, DALL-E 3 or Stable Diffusion have shown impressive progress in recent months. And if he has had trouble in this regard until now, now he seems to be able to achieve it.

One of the most important improvements in Image 3 is its ability to follow instructions provided by the user much more faithfully and accuratelyespecially if the challenges are complex or detailed. This solves one of the most common problems with previous generations of models: the tendency to ignore or misinterpret specific parts of requirements. Now Image 3 is able to capture more complex concepts, respecting the main elements as well as the small details that enrich the scene.

In addition to its accuracy, Figure 3 stands out for the quality of the results in terms of textures, lighting and composition. The generated images show a significant improvement in the handling of highlights and shadows and achieve a more realistic visual balance. Details such as depth of field, edge sharpness and use of color have also been fine-tuned, allowing for more vivid and balanced images even in complex art styles such as photorealism, impressionism or anime.

Another new feature of this update is the ImageFX tool, the Google platform where Image 3 is available. The interface has been improved with new features such as “chiplets”: automatic suggestions of expressions related to the challenge which facilitate iteration and refinement of images. For example, when you type “hummingbird next to a strawberry,” ImageFX offers options like “realistic detail,” “natural lighting,” or “background blur,” allowing the user to fine-tune the generation.

Of course, despite these advances Figure 3 still faces some of the most common limitations in this type of service. As with other similar models, there are still difficulties in generating certain elements, such as human hands in complex positions or interactions between objects. However, the improvements are evident and rank Imagen 3 as an increasingly advanced and versatile tool, capable of competing with the best models on the market.

Risks and Ethical Considerations: The Generative Artificial Intelligence Debate

The development of technologies such as I See 2 and Image 3 brings with it an increasingly topical debate: how can we use these tools without posing a risk to society? One of the most important challenges is disinformation. Veo 2’s ability to generate high definition videos with increasing realism could be used to create deepfakes and maliciously manipulated content. Aware of this risk, DeepMind has implemented its SynthID technology, an invisible watermarking system that allows you to identify whether a video has been generated by artificial intelligence. However, as with any solution of this type, SynthID is not invulnerable and it remains to be seen whether this will be enough to prevent abuse.

On the other hand, tools like Image 3, capable of generating images in seconds with an impressive level of detail, They invite illustrators, photographers and other creative professionals.. The automation of tasks that previously required hours of human labor raises complex questions about the future of the arts sector. Is it possible to protect creators in a world where artificial intelligence can replicate – or even surpass – their work in some cases? Although Google maintains its intention to work with the creative community, the balance between technological innovation and job preservation remains an open question.

Beyond these risks, it’s clear that generative AI is advancing at breakneck speed. Tools like Veo 2 and Imagen 3 open the door to new forms of visual expression and content creation, but also they require a deep discussion of the ethical limits and measures necessary to minimize their negative impact.

And that is also clear Models like Veo 2 and Imagen 3 are not only the present, but also the future of visual creation. Competition with other companies such as OpenAI ensures that the pace of development will continue to break even, driving new forms of expression and production. The question now is not whether these technologies will change the way we create content, but how they will and how far we will be able to take them.

More information: 1/2