Google Veo 3 Video Generation

Google Veo 3 is the one closed video model the open ecosystem genuinely hasn't caught up to, and the reason is one checkbox on this node: generate_audio. Veo 3 produces video and synchronized audio - dialogue, sound effects, ambience - natively, in a single pass. As of mid-2026 there's still no open-weight model that does that. LTX got close, but the KB's standing verdict is that "Veo 3's native audio has no open-source equivalent." So when you reach for this node, you're not just buying quality; you're buying a capability that doesn't exist locally.

It's an API node like the rest of the built-in partner/video family: ships with ComfyUI core, calls Google's Veo API through Comfy Org's proxy, bills your Comfy account in credits. Nothing runs on your GPU. Your prompt goes in, a finished clip comes back as VIDEO, and in between there's a queue and a poll.

The inputs that matter

• prompt - the video description. This is a frontier text-to-video model; it rewards proper sentences about subject, action, camera, and mood.
• model - the biggest decision in the node. Five options: veo-3.1-generate (flagship), veo-3.1-fast-generate, veo-3.1-lite (cheapest), and the two veo-3.0 variants. The 3.1 line adds resolution control up to 4K; the lite and 3.0 models cap at 1080p and reject 4K outright.
• generate_audio - the headline feature, and it's off by default. Audio costs more per second, but this is the only node in the whole built-in video category where flipping it on gets you real synchronized sound instead of a silent clip.
• aspect_ratio - 16:9 or 9:16. duration_seconds - 4, 6, or 8.
• negative_prompt - genuinely supported (many closed APIs ignore it), so use it.
• person_generation - ALLOW or BLOCK, Google's gate for people in the output. Default ALLOW.
• image - an optional reference image for image-to-video.
• enhance_prompt - present, deprecated, and ignored. It's a dead widget; don't toggle it expecting anything.

Where people get burned

• Google's content filters run server-side and hard. The node surfaces "blocked by Google's Responsible AI filters" as an actual error when the moderation flags a frame. If you're used to unrestricted local generation, adjust expectations before you start.
• 4K is a trap. It's only available on the veo-3.1-generate family, and the price badge makes clear it's the premium tier. Lighter models silently reject it at validation.
• It's real money per second. The badge scales with model, resolution, and audio. A full-quality 8-second clip with audio is the most expensive single generation in this whole category.
• Non-deterministic, as always here. The seed is a re-run toggle, not a reproducibility knob.

The community has been measuring the open-versus-closed video gap against Veo for over a year, and it's still the benchmark. If you need audio-synced footage and a budget, this is the built-in to use. If you need cheap iteration, look elsewhere.