GameDirector checkpoints

This repository contains the three trained checkpoint families used by GameDirector: AttackNet, SituationNet, and the causal autoregressive world model. The world-model checkpoint contains the trained DiT only; its Wan VAE, UMT5 encoder, and tokenizer must be downloaded separately.

Checkpoint manifest

Path	Contents
attacknet/attacknet.pt	three-game hit detector
situationnet/situationnet.pt	three-game spatial-state estimator
world_model/ar_world_model.safetensors	causal AR world-model DiT used by the agent (9.4 GB)

SHA256SUMS lists the expected hashes; check them with sha256sum -c SHA256SUMS.

Training data

The public NRFTW and Hollow Knight training splits are available at onepiece1999/GameDirector-Training-Data. The V Rising training split used by these three-game checkpoints is not included in that dataset.

Download

hf download onepiece1999/GameDirector --local-dir ./GameDirector-models

License

The checkpoints are released under the Apache License 2.0.