A video model paired synchronised sound with the picture

The process wall

Google DeepMind's model card for Veo 3 states it was published on 23 May 2025. Before this model, a video generator's soundtrack was a separate job: a maker rendered silent motion, then scored, foleyed or dubbed it afterward in its own pass with its own tools. The model card describes Veo 3 as producing 'high quality, high-resolution video with audio' from a single text prompt or image, folding that second pass into the same generation step.

What the documents show

The accompanying technical report names the mechanism: latent diffusion applied jointly to 'temporal audio latents' and 'spatio-temporal video latents,' with video and audio each compressed by their own autoencoder before a shared transformer denoises both together. That is a different claim from adding a separately generated audio track: the two signals come from one diffusion process conditioned on each other, which is what the model card means by native audio. The model card adds a second kind of evidence: Veo 3 was scored against Meta's MovieGenBench prompts and against named competitors, including Kling 2.0, OpenAI Sora and Runway Gen-3 and Gen-4, with results reported as human-rater preference rather than an automated score.

Authorship and limitation

The model card states its own limit plainly: 'maintaining complete consistency throughout complex scenes or those with complex motion, remains a challenge.' Its bias testing found the model 'appears to skew towards lighter skin tones when race is not specified in the prompt,' a finding attributed to the card's own red-teaming rather than an outside audit. On provenance, the card records that Veo 3 can produce deepfakes of 'worse quality than those made by dedicated deepfake tools,' a risk it says is mitigated 'in part' through SynthID watermarking, an embedded rather than visible credential, and explicitly a partial mitigation rather than a guarantee of detectability.

What survives

What a maker keeps is the writing: the prompt still authors the scene, the dialogue and the sound cues, while the joint diffusion step is what removes the separate scoring or foley pass, not the creative decision behind it. Editorially, anyone citing Veo 3's current capabilities should check the live product page rather than this record, since Google's own page, retrieved 16 September 2026, already frames Veo 3 as one step behind its successor, Veo 3.1.

• Ask whether a generated clip's audio was produced jointly with the video or added afterward, since only the former is what the technical report calls native audio.
• Check a generated video for the lighter-skin-tone skew the model card itself reports before treating demographic representation as neutral.
• Look for a SynthID or equivalent embedded credential before relying on a watermark-free appearance as evidence of a clip's origin.

Read together, the two documents describe a specific technical change, joint audio-video diffusion, and name its own bounds: scene consistency, demographic skew and partial watermarking, rather than a general claim of finished, trustworthy output.