Google drops Veo 2 with 4K cinematic video and upgrades Imagen 3

For the past year, generative video has mostly been stuck in a funhouse mirror: gorgeous five-second clips that melt into nightmare fuel the moment a camera tries to track a moving subject.

Google just dropped its next big counterpunch. On December 16, the company unveiled Veo 2, alongside an updated version of its Imagen 3 image generator and a clever new remix experiment called Whisk.

According to Google, both models scored state-of-the-art results in head-to-head blind tests against rival leading models judged by human raters.

Talking to AI like a real director

The biggest leap in Veo 2 is raw fidelity combined with camera control. The model can output clips up to 4K resolution, and Google says it can extend generations to run for several minutes in length.

Under the hood, Veo 2 demonstrates a substantially better grasp of real-world physics, fluid human movement, and subtle facial expressions. The classic video hallucination problem—random extra fingers and drifting background objects—happens far less frequently here.

Even better, it understands cinematic shorthand. If you type "18mm lens," Veo 2 renders the wide field of view that focal length is known for. If you ask for a "shallow depth of field," it actually isolates the subject and blurs the background.

From gliding low-angle tracking shots to tight macro close-ups, the camera behavior finally follows cinematic logic. All generated outputs will embed Google's invisible SynthID watermark to trace AI provenance.

Veo 2 is currently rolling out via a waitlist on Google Labs' VideoFX tool, with plans to bring it to YouTube Shorts and Vertex AI next year.

Crisper Imagen 3 and an image-remixing toy

Alongside the video news, Google refreshed Imagen 3. The updated model produces brighter lighting, tighter composition, and better prompt fidelity across styles from photorealism to impressionism and anime. It is now live in ImageFX across more than 100 countries.

Google also launched Whisk, an experimental tool currently available in the U.S. that ditches text prompting entirely.

Instead of typing descriptions, users upload reference images for subject, scene, and style. Gemini reads those images and automatically writes detailed captions, which it then feeds into Imagen 3 to remix them into novel digital assets like enamel pins, stickers, or plushies.

Why it matters

For filmmakers, designers, and creators, Veo 2 represents a shift from lottery-ticket prompting to deliberate production. Being able to call out specific lenses, focal lengths, and camera trajectories turns generative video into an actual camera rig rather than a random B-roll machine.

For the AI landscape, Google is putting intense pressure on OpenAI's Sora, Runway, and Kling. Pushing outputs to 4K and multi-minute timelines raises the technical ceiling, especially with Google sitting on YouTube, the world's largest video distribution network.

We think pairing Gemini's visual analysis with Imagen 3 in Whisk is a brilliant, lightweight demonstration of how foundational models should cooperate. However, delivering 4K video spanning several minutes carries an eye-watering compute bill. The fact that Veo 2 remains locked behind a VideoFX waitlist shows that while Google has cracked the visual quality, making it fast and cheap enough for everyday creators is still the real bottleneck.