AlayaWorld: Open Source, Long Horizon and Playable Video World Generation

AlayaLab released AlayaWorld as an open source model under an Apache 2.0 license, accompanied by arXiv paper 2607.06291. The model generates consistent video sequences from a scene while responding to camera instructions, text prompts, and player interactions, maintaining coherence across durations that typically cause other video generation models to drift or reset.

Four capabilities define the release: camera controllable generation, a game world visual aesthetic mode, prompt driven scene interactions, and long horizon video output that holds environmental consistency across extended sequences. Each is demonstrated in the videos below.

Camera Controllable Video Generation

AlayaWorld accepts explicit camera instructions as part of the generation process, producing video that follows specified trajectories through a 3D scene. The model maintains scene consistency across cuts, so a pan or dolly move does not produce the visual flickering or content drift that affects most short form video generation models.

That stability across camera movement separates AlayaWorld from models that generate a fixed clip per prompt. The camera becomes a parameter of the generation itself, which means a user can define a path through the generated world the same way a cinematographer would call a shot.

Game Style Video Generation

AlayaWorld includes a generation mode that renders scenes in a stylized game world aesthetic. The visual language draws on the conventions of real time 3D environments. Clear depth separation, stylized lighting, and environment design that reads as navigable space rather than a rendered still are the defining properties of the mode.

This mode is useful for filmmakers working on projects where a game engine aesthetic is part of the design, or for rapid previsualization of environments where a world building visual register is needed before committing to full photorealistic production assets.

Prompt Driven Interaction

AlayaWorld accepts text prompts describing actions or events to occur within an existing scene, then generates video of those events unfolding in the established environment. The scene context persists across the prompt, so the generated action reads as belonging to the world already visible in the video.

The two examples below show distinct generated environments, each receiving a text prompt after the base scene was established. The model generated the resulting event without reconstructing the scene.

Long Horizon Video Generation

Most video generation models degrade after a few seconds. Scene coherence breaks down, objects shift position, and the visual logic of the environment stops holding across frames. AlayaWorld addresses this with a generation architecture designed to maintain scene consistency across durations significantly longer than standard generation windows. Temporal consistency across extended sequences is the specific problem the long horizon design targets.

The videos below are played at 2× speed to show the full extent of each generated sequence. Green Valley and Ancient Hall are each a single continuous generation with stable environment geometry, consistent lighting, and no visible reset or transition between segments.

Try Your Own Magic

The AlayaWorld project page includes an interactive demo built around the pyramid scene from the paper's evaluation set. The walk video below shows a character approaching the pyramid. Selecting one of the ten spells triggers the model's prompt driven generation for that action within the scene. Videos load from the AlayaWorld project server and may take a moment to start on first request.

Try Your Own Magic

Walk up to the pyramid. Cast a spell and watch it unfold.

Fire

Ice

Water

Earth

Light

Void

Poison

Rain

Fireworks

Summon

What Filmmakers Can Do With It

AlayaWorld's camera control capability is immediately applicable to previsualization. A filmmaker can generate a scene and then specify a camera path through it to see how a dolly, crane, or orbit move would read before committing to a physical shoot. The model maintains scene geometry across the move, which makes it more useful for this purpose than models that regenerate the scene from scratch for each clip.

The game world generation mode has direct applications for projects combining live action with digital environments, or for concept development in virtual production and gaming. A production designer can generate candidate environment designs quickly, iterating on visual tone without building assets. This aesthetic register is distinct from Nvidia Cosmos 3, which targets autonomous vehicle and robotics simulation; AlayaWorld's stylized output is more suited to creative world building and pre production concept work.

The long horizon generation capability addresses a persistent problem in AI video production. Temporal consistency across extended scenes is what the architecture specifically targets. DreamX World and Lingbot World have approached similar challenges through different architectural designs. AlayaWorld adds the prompt driven interaction layer on top of long horizon stability, meaning a filmmaker can generate a world, navigate through it, and trigger events within it without losing scene coherence. Its Apache 2.0 license makes it a strong candidate for local deployment and integration into video workflows in the AI FILMS Studio video workspace.

Try AI FILMS Studio

Generate text-to-video and image-to-video with the latest AI models in the video workspace.

Nodes Graph Editor

Build custom AI workflows by connecting models visually in the Nodes Graph Editor.

Sources

arXiv: AlayaWorld: Long-Horizon and Playable Video World Generation GitHub: AlayaLab/AlayaWorld Project Page: alaya-lab.github.io/AlayaWorld