Genie 3 and SHARP both turn an image into a scene you can walk through. One generates every frame as you move; the other builds geometry you can keep. Here is what actually separates them — and the company whose products already sit on both sides of the line.
Give Genie 3 an image and you can walk through it; give a feed-forward reconstruction model a handful of photographs and you can move through the scene they captured. The headline is the same, which is why the two get confused constantly — but one is generative AI producing new frames as you move, and the other is reconstruction, estimating the geometry that a real photograph shows. This article explains what actually separates them, who ships which, and the one company whose products already sit on both sides of the line.
A Genie-class world model is a generative video model with a control input. At every step, it takes the frames generated so far plus your action — a key press, a camera movement, a text instruction — and generates the next frame. There is no scene being built anywhere: the world exists only as a stream of pixels, each frame conditioned on the trajectory that preceded it. DeepMind's Genie 3, announced in August 2025, generates worlds this way at 720p and 24 frames per second, keeps them consistent "for a few minutes," and holds a visual memory that reaches "as far back as one minute ago" — those phrases are quoted from DeepMind's blog because a blog post is all that exists; no paper or technical report has been published.
The consumer release is instructive about where the limits sit. Project Genie, launched in January 2026 for Google AI Ultra subscribers in the US, caps each world at 60 seconds, after which you keep a video of your exploration and nothing else. That gap — "a few minutes" of consistency in the announcement, one minute per session in the product — is the honest measure of how hard persistence is.
Persistence, not image quality, is the unsolved problem of this family. Decart's Oasis 3, released in June 2026, can simulate hours of photorealistic driving at about two cents per second — but revisiting a street generates a different street, and cars can pass through each other. The model has no geometry to be consistent with; every frame is a fresh act of generation, disciplined only by what fits in its context.
A feed-forward reconstruction model like SHARP produces the opposite artifact: a million explicit Gaussians with positions, sizes and colours. The scene is a file. Rendering it is rasterization — hundreds of frames per second on ordinary hardware, with no model in the loop. Turn around and everything is exactly where it was, not because a network remembered it, but because nothing exists to forget: consistency is a property of the representation, not an achievement of the model. You can export the file, edit it, measure it, load it into an engine, and render it forever at no further cost.
What reconstruction cannot do is imagine. It reports what the input photographs support, and where they support nothing, the honest output is nothing — the blur and holes that every reconstruction method shows in unobserved regions. A world model has no such restraint: it will invent the street around the corner, weather that changes, a door that opens when you reach for it. Genie 3 accepts "promptable world events" — text commands that change the world while you are inside it. No reconstruction can do that, because a reconstruction is answerable to evidence.
A reconstruction is a document: fixed, exportable, consistent by construction, and limited to what was observed. A generated world is a performance: unlimited in content, interactive, and gone when the model stops — with consistency only as long as the model's memory. Which one you need depends entirely on whether your application needs the world to be true or needs it to be endless.
As of mid-2026 the frame-generation side is crowded. Alongside DeepMind's Genie line: Runway's GWM-1 (December 2025) generates about two minutes of action-conditioned 720p in real time, with variants aimed at robotics and avatars; Odyssey-2 (October 2025) generates a frame every 50 milliseconds with prompt-level interactivity; NVIDIA's Cosmos 3 (June 2026) targets robotics and driving, where the customer is a training pipeline rather than a player; Tesla trains its driving stack partly in a synthetic world simulator of its own. Meta's V-JEPA 2 is a world model in a different sense worth keeping distinct: it predicts embeddings for robotic planning, and generates no pixels at all. OpenAI, notably, ships video generation (Sora 2) but no interactive world model.
The clean dichotomy above has one prominent exception, and it is the most interesting data point in the field. World Labs — Fei-Fei Li's startup — ships both paradigms as products. Marble, generally available since November 2025, is a generative world model whose output is explicit 3D: exportable Gaussian splats and triangle meshes, built from text, images or video, from twenty dollars a month. Generation on the way in; a persistent, editable, engine-loadable scene on the way out. RTFM, the same company's research model, is the purest version of the opposite bet — an autoregressive "learned renderer" with no 3D representation at all, holding the world together with posed frames as spatial memory, sized to run on a single H100.
Marble outputting Gaussian splats does not fold it into the reconstruction family — but the boundary is subtler than "generation versus reconstruction." Give Marble a photograph and it stays faithful to what the photograph shows, exactly as a reconstruction would; past the photograph's edge, everything is generated, because what was never observed can only be invented. Seen this way, every system in this article reconstructs the evidence it is given and generates the rest — what separates them is the ratio, and what they do with the unobserved. SHARP generates the least. It is built for small viewpoint changes — Apple's target is the head movement of an AR/VR user, not walking through the photograph — so the only unseen content it ever has to produce is the thin regions that appear at object edges when the camera shifts, and it fills those deterministically from learned priors rather than by sampling a generative model. Marble inverts the ratio: the photograph is an anchor, and whole rooms of plausible invention are built around it. Genie drops the anchor entirely once a session starts, generating whatever your movement calls for.
Two questions, then, that are usually merged and should not be. One is form: frames generated on the fly, or an explicit scene you keep. The other is evidence: how much of the scene is answerable to input photographs, and how much is invention. Genie generates frames; SHARP estimates geometry; Marble generates geometry around whatever evidence you hand it. So "world models have no exportable geometry" — true of Genie, Oasis and GWM-1 — fails as a definition of the category. What survives without exception is the trade: frame-generation models buy unlimited content at the price of persistence, and explicit-3D systems buy persistence at the price of committing — to what the photographs show, or to one frozen version of what the model imagined.
If your application must revisit, measure, edit or ship the scene — robotics simulation you can rerun, digital twins, film assets, anything loaded into Unreal or Unity — you need explicit 3D, whether reconstructed from photographs (SHARP, and everything in the timeline) or generated (Marble). If your application is the experience itself — an endless game, an interactive concept, training data whose variety matters more than its repeatability — frame generation is the point, and its impermanence is a feature. And if you are betting on where this converges: the reconstruction stack and the generation stack now meet at the same output format, 3D Gaussians, which is not a coincidence. It is the format cheap enough to render and explicit enough to keep.
No on both counts. Genie 3 holds no 3D representation — the world exists only as generated frames — so there is no scene file to export. In the public Project Genie release, a session lasts 60 seconds and what you keep afterwards is a video of your exploration, nothing more.
It is a generative world model whose output is explicit 3D: exportable Gaussian splats and meshes. Given a photograph, it stays faithful to what the photo shows and generates everything beyond it — whole plausible rooms no camera observed. A reconstruction method's output, by contrast, is limited to what its input images support. Marble sits between the categories: reconstruction-faithful where it has evidence, generative everywhere else.
No. SHARP is feed-forward reconstruction: a regression model that estimates metric-scale Gaussians for what a photograph actually shows, designed for nearby views — small, natural head movements rather than walking through the photo. It does not respond to actions, does not invent places beyond the photograph's evidence, and its output is a fixed scene — the opposite of a performance. I cover SHARP in full in a dedicated article.
Both roads lead to the same primitive. My 3D Gaussian Splatting Course builds the full 3DGS pipeline from first principles in PyTorch only — no CUDA, no C++ — covering initialization, rasterization, backward passes and training loops.
Explore the Course →I share practical posts on 3DGS, NeRF, PyTorch implementations, research breakdowns, and the engineering details that usually get skipped.
Subscribe →At Qubit Analytics, we help companies put 3D Gaussian Splatting to work — classic per-scene 3DGS, feed-forward 3DGS, and the 3D foundation models covered on this page. We integrate these techniques into products, build custom reconstruction pipelines, and turn recent research into working prototypes.
For consulting inquiries:
contact@qubitanalytics.be