Overworld launches Waypoint-1, a real-time world-generating AI

Overworld has introduced Waypoint-1, an AI model that generates interactive video worlds in real time from text, images, and keyboard and mouse controls. Instead of simply creating a video for you to watch, it tries to build an environment you can enter and move through.

You can give it a few starting images, write an instruction such as "a game where you herd goats in a valley," and control the camera with your mouse or press keys. The system generates new frames while taking each action into account.

A world that responds to your movements

Unlike many video models, Waypoint-1 was trained from the start for interactive experiences. Overworld used 10,000 hours of video game footage, combined with controls and text descriptions.

In other systems, controls are usually applied every few frames. That creates lag and limits the feeling of being inside the environment. Waypoint-1 is designed to accept continuous camera movement and any keyboard key, with a response the company describes as zero latency.

The result is not a video game built with preprogrammed environments, characters, and rules. It is a sequence of images generated by the model as you play. Each new frame is produced using the previous frames and the actions you have just taken as context.

How it generates video on the fly

The model uses a diffusion technique, the same type of technology used by many image and video generators. Put simply, it learns to turn noisy information into coherent frames.

Its architecture processes frames in order: each one can use information from its own moment and from earlier moments, but it cannot look at frames that do not exist yet. This lets it create a continuous stream of new images instead of generating the entire video before showing it to you.

Overworld also applied a later training process called self-forcing. The idea is to prepare the model for the same conditions it will encounter during operation: generate a frame, use it as context, and repeat the process. This is intended to reduce error accumulation and image degradation during long sessions.

Performance on a consumer computer

The available model is Waypoint-1-Small, with 2.3 billion parameters. Overworld says that, when run on an RTX 5090 graphics card through its WorldEngine library, it reaches:

• 30 frames per second using four generation steps.
• 60 frames per second using two steps.
• Approximately 30,000 token passes per second in a single denoising pass.

These figures reflect the company’s test conditions and an RTX 5090, not just any computer. Even so, they point to an important shift: virtual world models are beginning to approach immediate interaction instead of being limited to prerecorded video demonstrations.

What you can do with Waypoint-1

WorldEngine provides a Python library for connecting the model to your own applications. The system receives context images, text instructions, and keyboard or mouse input, then returns new frames for real-time streaming.

For now, Overworld has announced Waypoint-1-Small. The Waypoint-1-Medium version will arrive later. It has also opened Overworld Stream for testing the experience and organized a hackathon to develop applications with WorldEngine.

For you, the important point is that AI is no longer just a tool for generating self-contained scenes. If these models improve their stability and maintain their performance, they could be used to create game prototypes, simulators, educational environments, or virtual spaces that respond to your actions without requiring every setting to be built manually.

The next challenge will not only be generating more frames per second. It will be maintaining consistent rules, remembering what happens in the world, and preventing the environment from deforming after several minutes of interaction. That will show whether Waypoint-1 is a quick demonstration or the beginning of a new way to create digital worlds.