Genie 4 vs CARI4D: One Generates a World, the Other Solves a Video

The only thing Genie 4 and CARI4D reliably share is a label. Genie 4, announced by Google DeepMind on October 27, 2026, is a generative world model: you describe a scene, steer it in natural language, and it produces video with synchronized audio for a couple of minutes. CARI4D, NVIDIA Research's CVPR 2026 reconstruction method, runs the other direction. It takes a monocular video of a person handling an object and returns that person's pose, the object's six-degree-of-freedom trajectory and two hand-contact scores for every frame. One invents a plausible world. The other measures a real one. Put them on the same page and the useful question stops being "which is better" and becomes "which direction is my problem facing."

The confusion is understandable, because both products get filed under "4D" and both produce something that moves through time. But a generator and a reconstruction method are not competing implementations of the same task; they fail in opposite ways and they are bought by different people. Genie 4 can invent a convincing street that never existed, and cannot tell you where anything in it actually is. CARI4D can tell you, to the frame, how a hand closed around a kettle, and cannot invent a kettle it has not seen. Below, the two models are separated by what they take in, what they hand back, and what each one structurally cannot do.

What Genie 4 returns: a session, not a scene file

The first thing to understand about Genie 4 is that its output is consumed in real time and does not survive the session as geometry. Google's own description is a world you set in motion and steer: move the camera, change the weather, add a character, and a Gemini model adapts the instruction to what is already on screen. The frames are generated autoregressively, which is why coherence is measured in seconds and minutes rather than in scene complexity.

The additions over Genie 3 are the ones every headline picked up. Audio is generated natively — ambient sound, music and interactive dialogue — rather than layered on afterward. Steering is live rather than a prompt that restarts the world. And the coherence window is reported at roughly two minutes, double its predecessor, although Google's blog says only "short sessions" and the two-minute figure comes from launch briefings.

What is absent matters just as much. There is no mesh export, no scene graph, no camera track, no depth buffer and no way to take the generated world into Blender, Unity or Unreal. A studio that wants to block a shot can watch Genie 4 do it and cannot hand the result to a pipeline. There is also no published resolution, frame rate, model size, benchmark or price, and access is a limited trusted-tester pool with a public waitlist — AI Ultra subscribers through Project Genie and Gemini API early-access partners are promised "soon," with no date.

Google does name the failure modes, which is more than most launch posts do: detail and object permanence drift across longer sessions and after repeated edits, spatial accuracy is imperfect, in-world text is frequently illegible, and audio accuracy is unsolved. All output carries a SynthID watermark. For a pre-visualization tool those caveats are survivable. For anything that needs a measurement, they are disqualifying.

What CARI4D returns: poses, trajectories and contact

CARI4D is a pipeline, not a single network, and its output is deliberately narrow. Given a monocular RGB video, it reconstructs the human with a parametric body model, reconstructs the manipulated object's mesh from the first frame, places both at metric scale, and tracks them through the sequence. The learned component that does the correcting is CoCoNet, a render-and-compare transformer that takes the current estimate, renders it into RGB, depth and masks, and regresses per-frame pose corrections plus one contact logit per hand.

The dependencies are explicit in the repository: UniDepth for metric depth, NLF for human pose, Hunyuan3D for the object mesh, FoundationPose for pose search and tracking, and CoCoNet for the correction pass. That matters when you budget the deployment, because each of those is a separate checkpoint with its own licence and its own hardware profile.

CARI4D ships in two forms. The research checkpoint, nvidia/CARI4D, is marked research and development only and links to arXiv 2512.11988, published December 12, 2025. A commercial variant, cari4d_commercial, covers NVIDIA's Momentum Human Rig and is governed by the NVIDIA Open Model Agreement. Neither is deployed by any inference provider on Hugging Face, so "available" means weights plus a Docker setup plus an NVIDIA GPU, not an endpoint. There is no public leaderboard row and no independent reproduction; the numeric evidence is the CVPR paper's own ablations on BEHAVE, which is vendor-produced research.

The comparison, dimension by dimension

• Input — Genie 4 takes a text description plus live steering instructions; CARI4D takes a monocular RGB video that already exists.

• Output — Genie 4 returns generated video and audio, streamed in a session; CARI4D returns per-frame human pose and shape, object rotation and translation, and hand-contact logits.

• Grounding — Genie 4 is plausible but ungrounded; nothing in the output corresponds to a measured coordinate. CARI4D is metric, recovering scale through depth estimation and pose search against the observed frames.

• Persistence — Genie 4's world ends when the session ends and cannot be exported. CARI4D's output is a data structure you can store, re-render and compare.

• Audio — Genie 4 generates it, including dialogue; CARI4D has none and does not model sound.

• Access — Genie 4 is invite-only with no price; CARI4D is downloadable weights with a research licence and a commercial variant under NVIDIA's agreement.

• Evidence — Genie 4 has demos and vendor claims, no published benchmark; CARI4D has a peer-reviewed paper, vendor-run ablations and no third-party reproduction.

The overlap that exists, and the one that does not

The tempting composition is to use CARI4D to reconstruct a real interaction, then use Genie 4 to imagine variations of it. On paper the handoff looks natural: CARI4D produces a metrically grounded human-object scene, and Genie 4 produces alternate outcomes. In practice there is no interface between them. Genie 4 accepts text and steering input, not a reconstructed scene, a mesh or a pose track. Nothing published by either team describes feeding one into the other, and the fact that both outputs look three-dimensional in a screenshot does not make them compatible formats.

The overlap that does exist is narrower and more useful: both are tools for reasoning about what happens next in a physical scene. CARI4D gives you the observed dynamics — contact, grip, trajectory — which is exactly the kind of signal a robot-learning pipeline needs to label demonstrations. Genie 4 gives you counterfactuals, which is what a policy needs in order to rehearse situations it has not recorded. A lab building embodied data could plausibly use both, on different sides of the same dataset, without either model ever talking to the other.

The failure modes are also mirror images. Genie 4 can render a wrong object beautifully and drift away from it over two minutes. CARI4D can track a badly shaped object accurately through a whole video and still hand you a poor mesh, because its metrics are about pose and contact rather than surface fidelity. Neither model fixes the other's weakness.

What you can actually run this week

CARI4D is the one with a path to execution. The weights are public, the Docker image exists, and an A100-class GPU is the documented target. It is still a research pipeline with a paper to read and several external checkpoints to assemble, but a team with the hardware can be running it this week. Genie 4 is the one with a waitlist.

It is worth being direct about our own catalogue: OrcaRouter hosts neither model. CARI4D is a research reconstruction pipeline with no hosted endpoint anywhere, and Genie 4 has no public API to route. What we do route is part of the reasoning layer Google built around Genie 4 — Gemini 3.8 Flash, the model that watches generated frames and re-plans, and Gemini Robotics ER 1.6 preview — on one API across 200+ models with provider list prices passed through at 0% markup and automatic failover across providers. If your actual job is to ask questions about a video rather than reconstruct it or generate one, a hosted multimodal model is the cheaper answer, and it does not require an invitation.

For everyone else the decision reduces to a sentence each. If the video already exists and you need to know what the person and the object did, CARI4D is the only one of the two that answers the question. If the world does not exist yet and you need to see it move, Genie 4 is the only one that will show you — once someone lets you in.