Veo 3 Tech Report

TL;DR

Google DeepMind’s tech report for Veo 3, their text-to-video+audio generation system. The architectural content is extremely thin: Veo 3 is a single latent-diffusion model that runs the diffusion process jointly over temporal audio latents and spatiotemporal video latents, with modality-specific autoencoders and a transformer-based denoising network. The bulk of the document is the safety / responsibility stack — data filtering, multimodal-classifier content safety, fairness evaluation across 140 professions, CBRNE red-teaming, SynthID watermarking, and the assurance-evaluation governance process. No parameter counts, no benchmark numbers, no architecture diagrams beyond the high-level claim of joint A+V latent diffusion.

Key claims

Veo 3 uses latent diffusion applied jointly to temporal audio latents and spatiotemporal video latents, encoded by separate audio and video autoencoders [Model & Data, “Latent diffusion model”].
The denoising network is a transformer trained to remove noise from noisy latent vectors, then iteratively applied to Gaussian noise at sampling time [Model & Data, “Latent diffusion model”].
Training data is annotated with text captions at multiple levels of detail using multiple Gemini models, with safety filtering for unsafe captions and PII removal [Data].
Training videos are filtered for compliance, safety, and quality, and the corpus is semantically deduplicated across sources to reduce memorization risk [Data].
Synthetic captions are generated to broaden the concept-association distribution in training data [Mitigations].
Veo 3 is “still poor at generating text” — the report cites this as a key reason it remains unsuitable for misuse scenarios that require legible in-video text [Dangerous Capabilities].
Veo 3 has a strong bias toward “cinematic footage” with frequent camera cuts and dramatic angles, which the report frames as a side effect that makes realistic coercive-video generation difficult [Dangerous Capabilities].
Image-to-video can produce “quite good deep-fakes” but with worse fidelity and less speech-controllability than dedicated deepfake tools [Dangerous Capabilities].
Veo 3 demonstrates limited capability across CBRNE domains (nuclear physics, explosives physics, missile flight dynamics) — outputs consistently fail to accurately render radioisotopes, shielding geometry, critical-mass configurations, blast geometry, or missile dynamics [CBRNE].
The fairness evaluation generates 16 videos for each of 140 professions and categorizes outputs by perceived skin tone (Monk scale), age, and gender; the model “skews toward lighter skin tones when race is not specified in the prompt” [Unfair bias].
SynthID watermarking is applied as a post-training mitigation alongside production-time content filtering [Mitigations].
Four evaluation categories at the model level: development evaluations, assurance evaluations (run by a team outside model development, with held-out datasets), red teaming, and product-deployment review by Google DeepMind’s Responsibility and Safety Council [Responsibility & Safety Evaluations].

Method

The report describes Veo 3 only at a block-diagram level. Audio and video are encoded by separate autoencoders into compressed latent representations: video into spatiotemporal latents, audio into temporal latents. A transformer-based denoising network is trained with a standard diffusion objective on the joint latent representation (the report explicitly says “the diffusion process is applied jointly to the temporal audio latents, and the spatio-temporal video latents”). Sampling iteratively denoises an input Gaussian noise tensor through the same network.

No details are provided on: the dual-stream vs single-stream choice, cross-modal coupling mechanism, text-conditioning architecture, parameter counts per branch, training compute, latent shapes, RoPE choices, or sampling/CFG configuration. Compared to the open-weights joint-T2AV literature (LTX-2: Efficient Joint Audio-Visual Foundation Model, MOVA: Towards Scalable and Synchronized Video-Audio Generation, SkyReels-V4: Multi-modal Video-Audio Generation, Inpainting and Editing model, The Design Space of Tri-Modal Masked Diffusion Models), the report is silent on every design axis those papers explicitly characterize.

The safety pipeline is described in more detail: pre-training data is filtered for compliance and safety risk areas with semantic deduplication; multiple Gemini models generate captions at varying detail levels; synthetic captions diversify concept coverage; assurance datasets are held out from model development; SynthID watermarking and production filtering act at output time.

Results

The report does not include quantitative results. No FVD, VBench, Artificial Analysis ranking, human-preference Elo, or audio-quality metric is reported. The only numerical detail is the fairness-evaluation design (140 professions × 16 videos each). All “results” are qualitative safety claims:

Safety mitigations “effectively reduced content safety violations for the final Veo model” [Development Evaluations].
The model has “limited domain-specific capability” across CBRNE domains [CBRNE].
Veo 3 is “still poor at generating text” and “is generally prone to small hallucinations that mark videos as clearly fake” [Dangerous Capabilities].
The model “appears to skew towards lighter skin tones when race is not specified in the prompt” [Unfair bias].

Why it’s interesting

This is the closed-source counterpart documentation for the joint-T2AV recipe the wiki has been tracking on the open side — but the report deliberately discloses almost nothing about architecture, scale, or evaluation. The one substantive technical commitment is that Veo 3 is a single latent-diffusion model with joint audio + video denoising (not a cascaded T2V→V2A pipeline), which aligns with the open-source consensus argued in LTX-2: Efficient Joint Audio-Visual Foundation Model §1 that decoupled pipelines cannot model the joint distribution. Whether Veo 3 uses a dual-stream DiT with bidirectional cross-attention (as LTX-2, MOVA, SkyReels-V4 all converged on) or a single-trunk discrete-MDM recipe (à la The Design Space of Tri-Modal Masked Diffusion Models) or something else is not disclosed.

For the World Foundation Models cluster this is a useful data point on how closed flagships document themselves: the disclosure surface is policy + evaluation + governance, not architecture or benchmarks — mirroring the pattern in Project Genie: Experimenting with infinite, interactive worlds and The Waymo World Model: A New Frontier For Autonomous Driving Simulation. Veo 3 is referenced as a closed comparison point in multiple open-source filings (LTX-2: Efficient Joint Audio-Visual Foundation Model §6.3 cites Veo 3’s 12 s synchronized A+V duration; OmniForcing: Unleashing Real-time Joint Audio-Visual Generation uses Veo 3 as a closed baseline; RISE-Video: Can Video Generators Decode Implicit World Rules? places Veo 3.1 in the top cluster on reasoning-style TI2V evals) — this report is what those papers are actually comparing against, and it is striking how little is there to compare against directly.

The CBRNE / dangerous-capabilities section is also notable as a template: it makes explicit that text-rendering weakness and cinematic bias are currently load-bearing for misuse resistance, which is a fragile mitigation as those weaknesses get fixed in successor models.