HY-World 1.5 Made Generated Video Playable—Then Tencent Added Persistent 3D

HY-World 1.5, also known as WorldPlay, remains publicly available as an open-source model that generates video in response to navigation commands. It did not become a frictionless holodeck running entirely inside an ordinary browser, however, and Tencent’s subsequent HY-World work now produces persistent 3D assets alongside this earlier approach.

The original advance still matters: WorldPlay turns video generation into a responsive sequence rather than delivering one finished clip. The newer system also clarifies the boundary that beginners need to understand—convincing movement through generated imagery is not the same as possessing a stable, editable game world underneath.

WorldPlay generates the view, not a finished level

HY-World 1.5 is best understood as an action-conditioned streaming video model. It begins with an image or text description and predicts successive groups of frames while incorporating keyboard, mouse or camera-control inputs. The scene therefore appears explorable because movement affects what the model generates next.

A conventional game engine works differently. It maintains explicit geometry, objects, collision boundaries, rules and state, then renders an image from that stored world. WorldPlay primarily generates the visual observation itself, so paths, walls and landmarks remain credible through frame prediction and memory rather than through a complete scene graph prepared before exploration begins.

The authors’ revised WorldPlay paper reports long-horizon 720p generation at 24 frames per second and identifies three central mechanisms: keyboard-and-mouse action representations, a memory system that retrieves useful earlier views, and memory-aware distillation for faster generation. The paper was first submitted on December 16, 2025, and its current version was posted on June 9, 2026; these are research results from the authors’ configuration, not a guaranteed speed for every installation.

Memory makes the illusion navigable

A short generated video can show an attractive corridor without preserving it as a place. After the viewer turns around or takes a long detour, an ordinary autoregressive generator may replace the entrance, alter the room or lose objects that have fallen outside its recent context.

WorldPlay’s reconstituted context memory is designed to reduce that drift. It rebuilds the working context from earlier groups of frames and retains views considered geometrically important, allowing an old location to influence the image when the user returns. This tackles a problem that becomes much more visible in interactive generation than in a clip watched only once from beginning to end.

Long-term geometric consistency is still a model objective, not literal permanence. Remembering the appearance and relative position of a doorway does not create a stored mesh, collision map or editable door object. The system may present a coherent visual route while lacking many of the structures that game engines and simulation software use to enforce repeatable behavior.

Open source does not remove the hardware barrier

The public release is a technical local workflow, although a browser can act as the front end for an online or remotely hosted demonstration. The maintained HY-WorldPlay repository dates the open-source release to December 17, 2025, requires an NVIDIA CUDA GPU and lists 28 GB, 34 GB or 72 GB of GPU memory for its HunyuanVideo-based distilled inference setup, depending on sequence parallelism. Its installation also requires model downloads, Python dependencies and configuration outside the browser.

The same codebase uses a 480p image-to-video base model for its HunyuanVideo pipeline and publishes several 480p checkpoints. That does not negate the paper’s 720p result: the figures describe different published configurations rather than one universal output specification. Resolution, speed and memory demand depend on the selected checkpoint, pipeline, parallelism and optimizations.

The repository also includes a lighter five-billion-parameter WAN-based route, but labels its action control and long-term memory as compromised relative to the recommended eight-billion-parameter HunyuanVideo pipeline. Quantization can reduce computational pressure, yet the documented requirements do not support treating WorldPlay as a normal client-side web application that runs on any laptop.

Tencent’s next system preserves the generated world

The major update after WorldPlay is a change in output, not merely a higher-resolution stream. The official HY-World 2.0 repository records an initial technical release on April 16, 2026, additional world-generation components in May and a HY World 2.1 product update in July; the system accepts text, images or video and produces meshes or 3D Gaussian-splatting representations.

Those representations persist after generation and can be explored, edited or imported into established content and simulation tools. HY-World 2.0 divides the process into panorama generation, trajectory planning, world expansion, composition and 3D reconstruction. Instead of calculating every future view solely as another stretch of video, it creates an asset that another program can inspect and render repeatedly.

The two approaches therefore solve different parts of the generative-world problem. Streaming video can respond before an elaborate 3D asset has been completed, while persistent 3D is more suitable when a creator needs repeatable camera positions, editable geometry, collision handling or integration with an existing engine.

What “playable” means in HY-World 1.5

For WorldPlay, playable means that user input changes subsequent generated views quickly enough to support interactive navigation. It does not by itself establish deterministic physics, reliable object interaction, durable game state or a world that can be saved and edited as conventional 3D geometry.

Four distinctions help define the capability without either dismissing or overstating it:

• Action responsiveness asks whether the generated view follows the requested movement.
• Visual consistency asks whether a location remains recognizable after a turn or long detour.
• State persistence asks whether changes to the environment survive beyond the immediately generated frames.
• Asset persistence asks whether another application can inspect and edit the resulting world after generation ends.

HY-World 1.5 made the first two properties substantially more concrete by combining action-controlled video with an explicit long-range memory design. Tencent’s move toward meshes and Gaussian splattings addresses the latter two from a different direction. The result is less a single march toward a browser holodeck than a split engineering problem: one system must generate responsive experience, while another must preserve structure, state and reusable space.

Also read:

• Guy Raises $55K+ in One Day on Real-Concrete Tiny Construction Kit — Then Gets Roasted for AI Video
• Navigating the World of Video Conferencing:Best Practices for Engaging and Effective Virtual Meetings
• NVIDIA Lyra 2.0 Solves Spatial Forgetting and Temporal Drift in Generative Video