Google unveils Veo 2 and expands access to Imagen 3

Google unveiled Veo 2, its new video-generation model, on Monday, aiming to reduce physical errors and give people creating clips with artificial intelligence more creative control. The company has also expanded the availability of Imagen 3 in ImageFX and launched Whisk, an experiment for creating images from other images.

The announcement intensifies competition with OpenAI and its Sora model, which became available to paying subscribers last week. AI-generated video has become one of the industry’s most hotly contested fronts: the technology can already produce striking sequences, but it still struggles with complex movement, continuity between shots and believable physical actions.

Veo 2 aims to better understand how the world moves

Google DeepMind says Veo 2 improves its understanding of physics, human movement and cinematography. In practical terms, users can describe not just a scene—for example, someone running in the rain—but also the lens, camera angle, lighting or a specific movement, such as a tracking shot or a low-angle shot.

The model can generate video at resolutions of up to 4K and clips lasting more than two minutes, according to Google. That is a significant capability because many of the generators available so far have focused on very short sequences. Still, the maximum length a model can produce does not necessarily translate into a video ready for publication: keeping characters, objects and spaces consistent throughout a long scene remains one of the technology’s central challenges.

Google says Veo 2 delivers better results than other models in comparative human evaluations of realism and adherence to instructions. Because the study was commissioned by the company itself, it serves as an indication of its progress but does not replace independent testing or answer a crucial question: how will the tool perform on creative assignments that differ substantially from those used in the evaluation?

Veo 2 will initially be available to a limited group of VideoFX users, Google Labs’ experimental video-generation platform. The company plans to bring it to Vertex AI, its enterprise platform, later, opening the door to uses in advertising, training, audiovisual prototyping and social media content.

Imagen 3 expands across ImageFX

Alongside its new video model, Google has expanded Imagen 3 to ImageFX users in more than 100 countries. Imagen 3 is not an entirely new model: Google introduced it in August. Monday’s announcement is about its broader rollout within the free, experimental tool.

The generator is particularly good at rendering people, handling lighting and following detailed instructions, according to the company. Google also highlights the quality of text inside images—an area in which visual models have made progress but still frequently produce errors when they need to reproduce words, signs or labels accurately.

All images generated with Imagen 3 and videos created with Veo 2 include SynthID, Google DeepMind’s digital watermark. The signal is designed to be imperceptible and withstand common alterations such as cropping or compression. It helps identify AI-generated content, although it does not prevent an image or video from being shared outside Google’s ecosystem, nor does it by itself solve the problems of attribution and misinformation.

Whisk swaps text prompts for visual references

Google used the announcement to launch Whisk, another experiment available through Google Labs in the United States. Instead of writing a long description, users can provide reference images for the subject, scene and style. Whisk uses Imagen 3 to create a new composition based on those visual prompts.

The tool is designed for quickly exploring ideas, not for copying an image pixel by pixel. Google says it captures broad traits from the references and interprets them in a new creation. That approach is useful for designing concepts, but it also illustrates a debate that will continue to grow: how much control creators want over the use of their work as references by generative systems.

With Veo 2, Google has a more ambitious answer to Sora and other video models such as Runway and Kling. The first real test will be its availability: the industry needs to determine whether the advertised level of control holds up outside demonstrations and whether creators can integrate these clips into their workflows without spending more time fixing flaws than producing content.