World models
Wizard: a real-time game world model with agentic control
Researchers from Tencent, NUS and NTU have published Wizard, a world model that lets an agent track game rules while a video model draws the scene at 24 frames a second. It handles new characters and scenes without retraining.
On 29 October a team from Tencent, the National University of Singapore and Nanyang Technological University posted Wizard, a paper on a “game world model” that keeps the rules of a game outside the video model. An agent holds the state; the video model only draws it.
The aim is to fix two weaknesses of pure video world models: they struggle to keep game state consistent, and they usually only handle the characters and motions they were trained on. Wizard is built to take a new scene, a new character and a new move, and still run in real time.
How it works
The system splits the job in two. An agentic layer tracks the world and the rules. A video generator turns that state into pictures.
From a single scene image, Wizard estimates depth to build a rough, walkable 2.5D version of the space. Character reference images are converted into 3D meshes and rigged with skeletons. A vision-language model then picks matching animations from a motion library and retargets them to each character’s body. Setting all this up takes under five minutes before a session starts.
During play, the agent updates the skeletons, the scene geometry and combat state such as health, distance and facing. Those become depth and skeleton maps that condition the video model, which is fine-tuned from the open Wan2.2-TI2V-5B. A masking step limits each character’s reference image to the area where that character appears, which helps keep faces and outfits stable while they move.
Because the rules live in the agent, they hold. In the paper’s boss-fight demo, when either fighter’s health reaches zero the controller ends the fight and shows the result.
The numbers
The team built a training set covering 750 scenes and 178 characters, with 44,075 clips from 50.3 hours of paired footage. After distillation, the model produces video at 24 frames a second using two denoising steps and a rolling cache of 20 recent frames, so memory use stays bounded however long you play.
Limits
The evidence is qualitative. The paper shows example rollouts and does not report benchmark scores against other models. We found no hardware details for the 24 fps figure in the text, so treat it as a research result rather than a product spec.
The authors name two limits themselves. Characters with unusual bodies are hard to match with the motion library, and the generated video can drift from the scene geometry over longer rollouts. The demonstrations are also mostly combat.
Our take
Wizard is a clear example of a pattern we expect to spread: keep the facts of a world in explicit state, and let the generative model handle appearance. For anyone building interactive worlds, that is how you get rules that hold and a world that can still be reshaped by the player. Our explainer on how AI world models work covers the wider design space.