SCOPE turns Wan2.2 into a playable FPS world model

Generating a cinematic clip with an AI video model is trivial nowadays. But until now, you were just a spectator—the model decided where the camera walked and when the gun fired.

Researchers just changed the rules by open-sourcing SCOPE, an interactive world model explicitly built for playable first-person shooter environments. Plug in a standard controller, and you can actually steer the protagonist, aim, and shoot inside a video generated on the fly.

The project is built on top of the open-source Wan2.2 5B video model (Wan2.2-TI2V-5B). The research team has released the complete package: model weights across three DiT shards, inference scripts, and their custom CrossFPS dataset on GitHub and HuggingFace.

Decoupling the crosshairs from the scenery

Making video models respond to real-time gaming inputs has always hit a wall: real FPS gameplay involves chaotic, overlapping actions. Players move, steer the camera, jump, and reload all at once.

Previous game world models injected control signals globally across the entire frame. The result? The moment you fired a shot, the entire skybox wobbled and the background melted. SCOPE solves this by inserting specialized ActionModules into each of Wan2.2's 30 DiT transformer blocks.

The system splits the 10 degrees of freedom into two distinct pipelines. Continuous inputs, like the 4D dual-joystick movements, use temporal self-attention to slide the camera and surroundings naturally. Meanwhile, discrete buttons—such as firing, scoping, or reloading—are channeled through cross-attention that confines effects strictly around the weapon, without needing manual segmentation masks.

At present, the pipeline outputs 480x832 video at 20 FPS, spanning 81 frames (roughly 4 seconds). Users with hefty GPU memory can easily bump that up to 161 frames.

Seven games of training, zero-shot across new worlds

To give the model real shooter intuition, the authors built CrossFPS: 69,000 video clips captured across seven distinct FPS titles, frame-matched with 10-DoF gamepad telemetry.

In head-to-head testing against prior models like HY-World 1.5 and Matrix-Game 3.0, SCOPE wiped the floor with multi-action composition. While older baselines dropped to single-digit success rates when asked to move and shoot simultaneously, SCOPE hit 76%. Simultaneous scope-and-fire hit 82%, with an overall zero-shot task completion rate of 71.5%.

Even wilder is how it generalizes. Feed it a single static screenshot from completely unrelated titles like Black Myth: Wukong, Genshin Impact, or It Takes Two, supply an action sequence, and it generates a coherent FPS run through that universe without any fine-tuning.

Why it matters

While the rest of the AI industry is busy arguing over which video generator renders prettier 1080p b-roll, SCOPE takes world models in a far more practical direction: raw interactivity.

For game designers and indie developers, this is an enticing glimpse into next-generation prototyping. Instead of blocking out 3D greybox meshes, rigging weapon viewmodels, and coding controller state machines just to see how a level feels, a developer could sketch an environment concept, hook up a gamepad, and instantly test the vibe in first person.

Architecturally, the approach is remarkably smart. Rather than training a massive foundational model from scratch at immense cost, the authors frozen Wan2.2 and trained lightweight action adapters with zero-initialized residuals. It is a playbook any developer can borrow.

That said, nobody should expect it to replace Unreal Engine tomorrow. Generating four-second chunks at 480p on high-end silicon means latency and scale remain massive hurdles. But we think SCOPE provides one of the clearest blueprints yet for how neural video engines will eventually evolve into playable games.