GameDirector Keeps Boss Fights Honest by Directing the Rules Outside the Video Model

GameDirector is out of paper-and-demo mode: the agentic control stack, the checkpoints and a large training corpus are public. The Hugging Face model repo onepiece1999/GameDirector was created on October 11, 2026 at 07:00 UTC (12:30 PM IST), with Apache-2.0 weights for AttackNet, SituationNet and a causal autoregressive world-model DiT. Matching code landed the same day at Jimntu/GameDirector_code, and the NRFTW plus Hollow Knight training splits went up under onepiece1999/GameDirector-Training-Data. The research team spans Tencent, Nanyang Technological University, the National University of Singapore and A*STAR.

The project page frames a blunt problem. End-to-end game world models can look convincing while ignoring health bars, cooldowns and termination. Traditional engines enforce those mechanics but leave little room for player-driven configuration. GameDirector tries to sit between those extremes.

Configure, direct, then paint

The loop has three named stages. The player configures the scene, characters and skills, including hit points, attack strength and skill meters. The director reads the latest frames: AttackNet detects hits, SituationNet estimates distance and facing, explicit combat state is updated, a local roughly 2-billion-parameter Gemma planner chooses the boss skill, and rule enforcement validates the command. Validated decisions become structured text prompts. A causal autoregressive world model built on Wan2.2-TI2V-5B renders the next quarter-second cell, and those frames feed back into the director.

That separation is the point. Gameplay logic lives in code that can refuse an illegal skill or apply damage exactly once. The video model is asked to render what the director already decided, not to invent the rules from pixels.

Three games and the numbers the team reports

The public materials cover three titles—No Rest for the Wicked, Vampire and Hollow Knight—with 13 characters (five players and eight bosses). An automated gameplay agent collected synchronized video, controls, boss actions and engine-internal states. The project page cites 51,758 clips, with 300 held out for evaluation (100 per game). The released training dataset card lists 23,966 NRFTW clips and 10,324 Hollow Knight clips as five-second, 80-frame samples; the Vampire training split used for the three-game checkpoints is not included in that upload.

On the comparison table published with the project, GameDirector reports 100% state alignment, mechanics fidelity of 98.5% under a GPT-5.5 judge and 99.6% under a Gemini judge, and decision quality of 88.4% and 93.0% under those same judges. End-to-end baselines in the same table sit far lower on mechanics fidelity, often in the low twenties to mid-thirties. AttackNet and SituationNet average above 90% accuracy on hit detection and spatial classification across the three games. Those figures are the authors’ reported results from their evaluation protocol.

The accompanying paper, arXiv 2609.25652, was posted September 22, 2026. The October 11 release is the first public code-and-weights drop for the full stack.

What you have to download to run it

The Hugging Face checkpoint pack is not self-contained. The world-model file is the trained DiT only (about 9.4 GB in the card’s layout). Operators still need Wan2.2 VAE and UMT5 encoder weights from Wan-AI’s repos, plus google/gemma-4-E2B-it for the planner after accepting that model’s license on the Hub. The code repository ships example/run_rollout.sh and a validate_bundle.py helper that checks the installed layout. SHA256 sums are included with the GameDirector weights.

License detail matters. The checkpoint card is Apache-2.0. The training dataset card is marked license: other. The GitHub code tree did not ship a top-level license file at release. Anyone building on the stack should read all three before commercial use, and should treat the footage-derived game data accordingly.

Why agent people should care

Most recent world-model demos ask one network to remember state, invent tactics and paint frames. GameDirector’s bet is that agent scaffolding—perception heads, an explicit state machine, a small planner and hard rules—can carry the parts generative video still bungles. The tradeoff is operational weight: multiple checkpoints, external foundation models and a closed loop that only knows the three games it was trained on. As a research template for rule-true interactive generation, though, the October 11 drop finally lets other labs reproduce the claim instead of watching the teaser.