ReactiveGWM-Models: Reactive Game World Models for Street Fighter

📚 Models-Introduction

ReactiveGWM-Models is the model release accompanying ReactiveGWM, a game world model that decouples player control from NPC autonomy. Player actions are injected as a lightweight additive bias into the diffusion backbone, while NPC behaviour is conditioned through dedicated cross-attention modules over a structured Active_Behavior / Passive_Behavior / Strategy prompt.

The weights here are trained on ReactiveGWM-Datasets, which provides 5-second strategy-aligned rollouts on two classic Capcom fighting games:

Subset	Game (platform)
SF2	Street Fighter II: Champion Edition (Sega Genesis)
SF3	Street Fighter III: New Generation / 3rd Strike (CPS3)

For each game we release three checkpoints, all stored as .safetensors:

File	Role
ReactiveGWM_base.safetensors	Full ReactiveGWM trained from scratch on the structured strategy prompt + per-frame actions.
ReactiveGWM_transfer.safetensors	ReactiveGWM initialised from a vanilla checkpoint of Game 1 and transfer-trained on the strategy prompt.
Vanilla.safetensors	Vanilla world model baseline trained on the same clips with the free-form narration prompt (no NPC strategy decoupling).

🗂️ Directory layout

SF2/
├── ReactiveGWM_base.safetensors
├── ReactiveGWM_transfer.safetensors
└── Vanilla.safetensors
SF3/   (same layout)

🎮 Conditioning interface

ReactiveGWM_base and ReactiveGWM_transfer expect two parallel inputs:

1. Structured NPC prompt, with the same shape used in the dataset's metadata.csv:

   NPC: Active_Behavior(<move>: <desc>; ...),
        Passive_Behavior(<move>: <desc>; ...),
        Strategy(Offense | Control | Defense : <desc>)
   

2. Per-frame controller state for both players (buttons + d-pad), aligned 1-to-1 with the output frames — the same schema as actions.parquet in the dataset. This stream is injected as a lightweight additive bias into the diffusion backbone.

Vanilla.safetensors instead consumes a free-form English narration in the style of metadata_vanilla.csv, and serves as the text-to-video baseline for ablations.

Inference / training code, configs, and the data-generation pipeline live in a separate code release and are not distributed in this repository.

🤓 Acknowledgments

Training data was recorded with the stable-retro framework (SF3 via a CPS-3 custom integration on top of FBNeo). NPC strategy annotations were produced by Gemini. The Street Fighter games themselves are © Capcom Co., Ltd.; see the IP notice below. We thank the open-source community for their valuable contributions.

🔗 Citation

If you use these models, please cite:

---

⚖️ Copyright, IP, and licensing

Read this before using the models.

These models were trained on derivative material from two commercial video games whose copyright the model authors do not own:

• Street Fighter II: Champion Edition — © Capcom Co., Ltd.
• Street Fighter III: New Generation / 3rd Strike — © Capcom Co., Ltd.

All in-game visuals, character designs, sprites, backgrounds, audio, and game mechanics that the models can reproduce in their generated frames remain the exclusive property of Capcom Co., Ltd. No ownership over the underlying games is claimed or transferred by this release.

What is and isn't licensed

The CC BY-NC 4.0 license declared in the YAML header applies only to the contributions that are original to this release:

• the trained model weights in SF2/*.safetensors and SF3/*.safetensors,
• the model architecture, training recipe choices, and accompanying documentation.

The visual content these models can synthesise — sprites, character designs, stages, UI, audio, and any other recognisable Capcom IP — remains © Capcom Co., Ltd. and is not relicensed by this release. The weights are made available solely under fair-use / fair-dealing for non-commercial academic research on game AI, world models, and video understanding. CC BY-NC 4.0 does not grant you any right to the underlying Capcom IP. See the LICENSE and NOTICE files shipped alongside this README for the full statement.

What you may NOT do

• Use the weights, generated outputs, or any model derived from them in a commercial product, including a game, a streaming service, or a paid API.
• Use the models, or any further fine-tune of them, to reproduce, recreate, clone, or compete with Capcom's Street Fighter games or any other Capcom title.
• Redistribute the weights bundled with game ROMs, BIOS files, or emulator binaries.
• Claim ownership of, or relicense, any Capcom IP that the models reproduce in their outputs.

What you should do

• Treat this as an academic, non-commercial research artifact only.
• Cite both this model release and acknowledge Capcom as the copyright holder of the source games in any publication or release that uses it.
• If you publish a downstream model or a generated-content release built on these weights, carry forward the same non-commercial restriction and copyright notice.

ROMs

No game ROMs are included in this repository, and none will be provided. Reproducing the training data from the upstream generation code requires you to obtain legal copies of the original Street Fighter II (Genesis) and Street Fighter III (CPS3) ROMs yourself, in accordance with the laws of your jurisdiction.

Takedown / DMCA

If you are a rights holder (in particular Capcom Co., Ltd. or an authorised representative) and believe these model weights or any content they reproduce infringes your rights, please contact zeqing.wang@u.nus.edu with a description of the material in question. The maintainer will respond and, where appropriate, remove the contested content promptly.