WorldFoundry
English中文

SCOPE

Method overview: S CO P E architecture. A SCOPE module is inserted into each DiT block. Discrete inputs use cross-attention with visual queries to confine effects to in-scope regions. Continuous inputs use MLP fusion and temporal.

SCOPE: `image` and `prompt` in; `generated_video` and `scope.mp4` out; runner verified; start on `scope`.

Edit this pageGitHubMarkdown

SCOPE is an interactive game world model over a Wan backbone: image plus per-step actions in, game-world video out (paper arXiv 2605.23345). In this catalog both integration and runner are verified — the image-action-to-video demo runs through the Workspace infer path with staged zizhaotong/SCOPE weights, the in-tree example image, and an action parquet.

Official paper teaser: S CO P E executes complex multi-action controls and action-environment interactions (highlighted in red boxes ) across diverse, unseen first-person scenes without retraining.
Official paper teaser: S CO P E executes complex multi-action controls and action-environment interactions (highlighted in red boxes ) across diverse, unseen first-person scenes without retraining.
Method overview: S CO P E architecture. A SCOPE module is inserted into each DiT block. Discrete inputs use cross-attention with visual queries to confine effects to in-scope regions. Continuous inputs use MLP fusion and temporal.
Method overview: S CO P E architecture. A SCOPE module is inserted into each DiT block. Discrete inputs use cross-attention with visual queries to confine effects to in-scope regions. Continuous inputs use MLP fusion and temporal.

The paper is SCOPE: Simulating Cross-game Operations in Playable Environments for FPS World Models. Project page: z2tong.github.io/SCOPE. Code: z2tong/SCOPE.

Catalog aliases include scope-world-model and zizhaotong/SCOPE.

Build your run

Pipeline
SCOPEPipelineworldfoundry.pipelines.scope.pipeline_scope
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  scope \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Drive SCOPE as an action-conditioned world route. Launch worldfoundry-eval run scope — required image and prompt; out generated_video and scope.mp4.

Score on WorldModelBench.

Runtime contract

Required: image and prompt. Optional: video. Actions: action_parquet. Artifacts: generated_video and scope.mp4. Status: Verified. SCOPEPipeline fails fast unless execute=True is supplied, then invokes the in-tree native assembly. Weights stay external HF assets injected through model_dir/checkpoint_root or SCOPE_MODEL_DIR.

Running in WorldFoundry

Run in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage zizhaotong/SCOPE locally — the runtime does not download weights. Launch worldfoundry-eval run scope.