What Veo 3.1 is good at — and what it's not

Veo 3.1 is the model you reach for when the final cut has to actually look like a finished film — when "AI video" with the usual telltale lighting bugs, melting hands, and texture noise won't pass. Google DeepMind built the Veo line with a heavy lean on cinematic vocabulary in the prompt parser. Tell Veo 3.1 "dolly in slowly, anamorphic lens flare from camera-right, golden-hour low key with the subject's face in shadow," and it will land all four of those instructions correctly — most other video models will execute two of the four and improvise the rest.

The headline feature of the 3.1 update is Ingredients to Video. Upload up to three reference images of a character, product, or object, and Veo holds them consistent across scenes, camera angles, and even lighting changes. This solves the single hardest problem in AI video: face drift. In every prior generation of AI video, the protagonist's face would subtly morph between shots — different cheekbones, different eye color, even when the prompt explicitly tagged them. Ingredients to Video locks the reference; the rendered character is the same person in every cut.

The second flagship-tier feature is output up to 4K with real detail. Veo 3.1 renders cleanly at high resolution without the noise artifacts and blurry stretch you get from upscaling a low-res source. For brand work, OOH placement, or any context where the final delivery is 4K, Veo finishes the path most other AI video models can't.

The third pillar is native synchronized audio — dialogue, ambient, and sound effects produced in the same pass as the picture, lip-synced and timed without a separate Foley step. Combined with cinematic prompt control and 4K output, this is the closest current model to producing a finished short in one generation.

Where it's weaker: for rapid prompt iteration, a lighter "Fast"-tier model is the better tool — use one to find the right composition, then graduate to Veo for the finish. Multi-modal reference inputs (audio reference, motion-reference video) are stronger on Seedance 2.0. And on raw text-to-video Elo, Seedance 2.0 currently sits slightly ahead.

A reasonable mental model: Veo 3.1 is the default for cinematic finish quality and resolution. For reference-heavy shots, Seedance 2.0. For multi-shot storyboards, Kling V3.