Overworld open-sources real-time playable world model Waypoint-1

Waypoint-1 is an open-source real-time interactive video diffusion model from Overworld that lets you explore AI-generated environments using standard keyboard and mouse inputs, just like playing a game.

Most existing world models simply fine-tune generic pre-trained video generators with rigid steering commands. You usually get stuck rotating the camera every few frames alongside agonizing input latency. Waypoint-1 flips that formula by training from the ground up for true real-time agency, delivering instant responsiveness without artificial control constraints.

What it does and how it works

You feed Waypoint-1 a text prompt—like a cyberpunk city or herding goats in a valley—along with initial context frames, and then jump straight in with your mouse and keyboard. The model dynamically generates new frames on the fly, conditioning every single step on your immediate player inputs.

Under the hood sits a frame-causal rectified flow transformer operating on compressed latent representations. Overworld trained it on 10,000 hours of diverse video game footage paired directly with logged controller inputs and captions.

To make procedural generation work seamlessly, the team pre-trained the model using diffusion forcing so it learns to denoise individual frames autoregressively. They followed that up with DMD-based self-forcing post-training, which aligns training with actual runtime rollouts, slashes error accumulation, and enables one-pass classifier-free guidance with very few denoising steps.

Hitting 60 FPS on an RTX 5090

Alongside the model, Overworld released WorldEngine, a high-performance Python inference library built specifically for streaming world models. Running the 2.3B parameter Waypoint-1-Small on an RTX 5090, WorldEngine delivers around 30,000 token-passes per second. That translates to 30 FPS at 4 denoising steps, or a blistering 60 FPS at 2 steps.

That throughput comes from four focused engineering tweaks: AdaLN feature caching, a static rolling KV cache paired with Flex Attention, fused QKV matmuls, and full-graph torch.compile optimization.

The Waypoint-1-Small weights are live on Hugging Face at roughly 25.2 GB, with a larger Medium variant coming soon. You can hook it into your own code via the WorldEngine library or test it right in your browser at overworld.stream.

Keep expectations in check regarding fidelity. Current renders are relatively low-res, long-term spatial consistency degrades over time, and visual hallucinations crop up frequently during rapid movement.

Why it matters

We think Waypoint-1 marks an important pivot for world models, moving past static video generation and directly solving the crippling latency bottlenecks that kept interactive AI worlds impractical. Running a 60 FPS interactive visual simulation on a single consumer GPU proves that playable neural rendering is within reach of indie developers right now.

Game developers and creative technologists should definitely tinker with it, if only to see how fluid controller-conditioned diffusion actually feels under the fingers. While consistency and hallucination problems mean it cannot replace standard game engines anytime soon, baking input causality directly into pre-training lays down a very clean blueprint for the future of interactive synthetic media.