Comparison
H3 Max Camera Control vs World Models
Camera control on a video model looks a lot like a world model in a short demo. The difference shows up when you ask what still exists after the model stops running.
Use this workspace to turn a loose idea into a focused video brief with a subject, motion direction, reference style, and next scene cue.
fal's H3 Max Multi Angle turns one image into a video with keyframe control of camera orbit, elevation, and distance.
H3-World is a community model on Hugging Face built on MiniMax-H3 that steers video with keyboard input.
World Labs' Marble exports Gaussian splats and meshes that open in ordinary 3D tools.
The one question that separates them
Turn the model off. Is the world still there? For H3 Max Multi Angle and H3-World, the answer is no: what you have is a video that was rendered along a path. For World Labs' spatial models, the answer is yes: the scene is stored as geometry you can download, reopen, and render again without the model.
Layer one: camera-controlled video
H3 Max Multi Angle takes a single image and generates a clip along a camera path you define with keyframes. fal's own description covers orbit, elevation, and distance in 3D space. It does not describe a persistent scene, stored geometry, or an exportable 3D asset. It is a video generator with a steering wheel, and a very good one.
Layer two: keyboard-steered video
H3-World, published on Hugging Face by an independent researcher, adds live keyboard input on top of MiniMax-H3. You press WASD to move the character and IJKL to move the camera, and the model keeps generating frames that respond. This is the same family as DeepMind's Genie and Skywork's Matrix-Game: consistency is something the model remembers from frame to frame, not something stored in a scene file. It feels like walking through a place. There is no place.
Layer three: generated worlds you can keep
World Labs approaches the problem from the other side. Marble generates an explicit 3D scene that can be exported as Gaussian splats or meshes and imported into a game engine, a simulator, or a video pipeline. Consistency is not a memory trick; the geometry simply exists. The tradeoff is that a generated world is mostly a static place, while video models can show weather, crowds, deformation, and story beats that geometry cannot easily hold.
Why the difference changes what you can build
A camera-controlled video costs compute for every second someone watches, forever. A generated world costs compute once, then renders on ordinary hardware. If you are building a live channel, a product demo, or a story stream, layer one and two are the tools. If you are building a game level, a robotics test scene, or an asset a client will reuse, layer three is the tool. The mistake is treating them as competing versions of the same thing.
FAQ
Did H3 Max just do what World Labs does?
No. H3 Max Multi Angle gives you camera control over a generated video. World Labs gives you a stored 3D scene. Both let you move a camera, which is why short demos look similar, but only one produces an asset that survives closing the tool.
Is H3-World an official fal or MiniMax release?
H3-World is a community project published on Hugging Face, built on the open-weights MiniMax-H3 model. It is not the same product as fal's H3 Max Multi Angle endpoint.
Which one should a creator try first?
If you want a cinematic clip from a still image, start with H3 Max Multi Angle. If you want to explore a scene and export it, look at World Labs. If you want to experiment with keyboard-driven video and can run models yourself, H3-World is the open option.