SCOPE is an interactive world model for FPS games with 10-DoF action control, trained on 69K clips across 7 games.
[](https://z2tong.github.io/SCOPE/)
[](https://github.com/z2tong/SCOPE)
[](https://arxiv.org/abs/2605.23345)
[](LICENSE)
## Highlights
- **Hybrid Action Space** — Jointly processes continuous (4D dual-joystick) and discrete (6 binary buttons) control signals within a unified framework — the first FPS world model to do so.
- **Dense Per-Frame Conditioning** — Resolves overlapping actions at every single frame, enabling simultaneous multi-action composition (e.g., moving + aiming + firing) that reflects real gameplay complexity.
- **Cross-Game Generalization** — Trained on 7 diverse FPS titles, a single model generalizes zero-shot to entirely unseen game environments without fine-tuning.
- **In-Scope / Out-of-Scope Decoupling** — Spatially selective conditioning that separates localized in-scope effects (weapon recoil, HUD) from stable out-of-scope world generation — without any segmentation labels.
## Model Overview
SCOPE is an interactive world model for first-person shooter (FPS) games. Built on [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B), SCOPE inserts a conditioning module into each transformer block that reshapes features into per-pixel temporal sequences. Each spatial position computes its action response from local visual content, naturally separating in-scope effects (e.g., weapon firing, reloading) from out-of-scope world generation (e.g., stable surroundings) — without any segmentation labels.
Trained on [**CrossFPS**](https://huggingface.co/collections/zizhaotong/crossfps) (69K clips, 7 games, 10-DoF), SCOPE learns general visual-to-action mappings rather than game-specific patterns, enabling zero-shot transfer to unseen scenes.
## Architecture
| Component | Details |
|-----------|---------|
| **Base Model** | Wan2.2-TI2V-5B (DiT, 30 transformer layers) |
| **Action Module** | Per-block conditioning with per-pixel temporal sequences |
| **Text Encoder** | UMT5-XXL (4096-dim hidden) |
| **VAE** | Wan2.2 Video VAE (4× temporal compression, 8× spatial compression) |
| **Total Parameters** | ~5B (1575 tensors, of which 750 are action-related) |
| **Precision** | BFloat16 |
### ActionModule Design
Each of the 30 DiT blocks contains an `ActionModule` with two conditioning paths:
- **Mouse/Joystick Path**: Sliding-window temporal features → MLP fusion → pixel-wise temporal self-attention with RoPE
- **Keyboard/Button Path**: Button embedding → temporal windowing → cross-attention (video queries, keyboard keys/values)
Both output projections are zero-initialized for stable residual training on top of frozen pretrained weights.
### Generation Specs
| Property | Value |
|----------|-------|
| Resolution | 480 × 832 |
| Frame Count | 81 frames |
| Frame Rate | 20 FPS |
| Duration | ~4 seconds |
| Inference Steps | 30 (default) |
## Action Input Format
SCOPE accepts 10-DoF action inputs per frame via a Parquet file:
**Controller Buttons (6D binary):**
| Index | Column | Action |
|:-----:|--------|--------|
| 0 | `right_trigger` | Fire (RT) |
| 1 | `left_trigger` | Aim Down Sights (LT) |
| 2 | `south` | Jump (A) |
| 3 | `right_thumb` | Melee (R3) |
| 4 | `west` | Reload (X) |
| 5 | `north` | Weapon Switch (Y) |
**Dual Joystick (4D continuous):**
| Column | Axes | Function |
|--------|------|----------|
| `j_left` | [x, y] | Character movement (left stick) |
| `j_right` | [x, y] | Camera rotation (right stick) |
## Quick Start
### Requirements
- Python >= 3.10
- PyTorch >= 2.0 with CUDA support
- GPU: NVIDIA with >= 24 GB VRAM (single GPU inference with CPU offload)
### Installation
```bash
git clone https://github.com/z2tong/SCOPE.git
cd SCOPE
pip install -e .
```
### Download Weights
```bash
# Download all weights (SCOPE DiT + Text Encoder + VAE + Tokenizer) in one command
huggingface-cli download zizhaotong/SCOPE --local-dir ./SCOPE
```
### Inference
**Single image + action sequence:**
```bash
python inference.py \
--model_dir ./SCOPE \
--input_image input.png \
--action_path action.parquet \
--prompt "First-person shooter perspective in a modern city" \
--seed 42
```
**Batch processing (directory of images):**
```bash
python inference.py \
--model_dir ./SCOPE \
--input_image_dir ./images \
--action_path action.parquet \
--prompt "First-person view in a battlefield" \
--output_dir ./outputs
```
For full usage details and advanced options, see the [GitHub repository](https://github.com/z2tong/SCOPE).
## Repository Contents
This repo contains **all weights** needed for inference in a single download:
| File | Component | Size |
|------|-----------|------|
| `model-00001-of-00003.safetensors` | SCOPE DiT shard 1 | ~5.0 GB |
| `model-00002-of-00003.safetensors` | SCOPE DiT shard 2 | ~5.0 GB |
| `model-00003-of-00003.safetensors` | SCOPE DiT shard 3 | ~4.6 GB |
| `model.safetensors.index.json` | Shard index mapping | — |
| `models_t5_umt5-xxl-enc-bf16.pth` | Text Encoder (UMT5-XXL) | ~20 GB |
| `Wan2.2_VAE.pth` | Video VAE | ~700 MB |
| `google/umt5-xxl/` | Tokenizer | ~10 MB |
| `config.json` | Model architecture config | — |
> **Inference code** is available at [github.com/z2tong/SCOPE](https://github.com/z2tong/SCOPE).
## CrossFPS Dataset
SCOPE is trained on [**CrossFPS**](https://huggingface.co/collections/zizhaotong/crossfps), the first multi-game FPS dataset with frame-aligned action telemetry:
| Property | Value |
|----------|-------|
| Games | 7 diverse FPS titles |
| Total Clips | 69,000+ |
| Action Dimensions | 10-DoF (6 buttons + 4D joystick) |
| Annotation | Frame-aligned action telemetry |
| Curation | Gameplay-bias removal for general visual-to-action mapping |
## Citation
```bibtex
@misc{scope2026,
title={SCOPE: Simulating Cross-game Operations in Playable Environments for FPS World Models},
author={Zizhao Tong and Hongfeng Lai and Zeqing Wang and Zhaohu Xing and Kexu Cheng and Haoran Xu and Zhao Pu and Shangwen Zhu and Ruili Feng and Jian Zhao and Yan Zhang and Hao Tang and Yeying Jin and Ling Shao},
year={2026},
eprint={2605.23345},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2605.23345},
}
```
## Acknowledgements
We thank the [Wan Team](https://huggingface.co/Wan-AI) for open-sourcing Wan2.2 and the [DiffSynth](https://github.com/modelscope/DiffSynth-Studio) team for the inference framework.