z2tong / SCOPE

What is SCOPE?

SCOPE is an interactive world model for first-person shooter games that generates playable video footage conditioned on player actions. Given a starting image and a sequence of controller inputs (joystick movement, firing, jumping), it produces short video clips showing what that gameplay would look like. The system handles 10-dimensional action inputs combining continuous joystick movement with discrete button presses, and can generate approximately 4-second clips at 480x832 resolution. It is built in Python on top of the Wan2.2 diffusion architecture and runs inference through a straightforward command-line interface.

Why is it gaining traction?

The standout capability is cross-game generalization. Trained on footage from 7 different FPS titles, the model can generate believable gameplay for entirely unseen games without any fine-tuning. This is unusual in a space where models typically overfit to their training domains. The hybrid action space also matters for realism - it simultaneously processes aiming, movement, and weapon actions rather than treating them in isolation. Developers interested in game AI research, procedural content generation, or simulating player behavior have a clear reason to experiment here.

Who should use this?

Game AI researchers working on world models or player simulation will find this most relevant. Researchers studying cross-domain generalization in video generation have a concrete benchmark. Simulation and synthetic data pipelines that need action-conditioned game footage could integrate this. For production game development, the hardware requirements (24GB VRAM minimum, 80GB recommended) and generation speed make this impractical today. Indie developers hoping to generate actual game content will need significantly more maturity.

Verdict

This is promising academic research with real novelty in cross-game action conditioning, but it scores only 1.0% on credibility metrics. With 47 stars and a recent arXiv submission, the project is in early experimental stages. The model weights are available and the inference code is functional, so serious researchers can reproduce results. Production use cases should wait for community validation, better documentation, and hardware optimization. If you are exploring game world models or action-conditioned video generation for research, this is worth evaluating against alternatives.