one year on
Google unveils Veo 2 video model with 4K output and improved physics
The second-generation video generation model achieves state-of-the-art results against leading models with cinematic controls and fewer hallucinations, as Google also upgrades Imagen 3 and launches a new image-remixing tool called Whisk.
Google today announced Veo 2, a major upgrade to its video generation model that can produce clips at resolutions up to 4K and extend to minutes in length. The model, available via the VideoFX waitlist, demonstrates a much stronger grasp of real-world physics and human movement compared to its predecessor, and includes explicit support for cinematic language – users can specify genre, lens type, and effects like shallow depth of field or low-angle tracking shots.
The company also released an updated Imagen 3 model, now rolling out globally in ImageFX, with brighter images, more faithful prompt following, and a wider range of art styles. The new Whisk experiment, launching today in the US, lets users combine subject, scene, and style images – powered by Gemini captioning and Imagen 3 remixing.
In head-to-head comparisons judged by human raters, Veo 2 achieved state-of-the-art results against leading models. Google says Veo 2 hallucinates unwanted details less frequently, and all outputs carry invisible SynthID watermarks. The company emphasizes a measured rollout, with expansion to YouTube Shorts and other products planned for next year.
The record
Report notes Veo 2 reduces hallucinations and includes SynthID watermarking, with measured rollout to VideoFX waitlist.
view the original post →One year later — open only if you can handle spoilers
Veo 2 quickly became the go-to model for professional video generation tasks, while Sora struggled with consistency and availability. Google expanded Veo 2 to YouTube Shorts in early 2025, and the model's API on Vertex AI saw rapid enterprise adoption.