--- title: Game2Player emoji: ๐ŸŽฎ colorFrom: red colorTo: gray sdk: gradio sdk_version: 6.32.0 app_file: app.py short_description: Visual game-playing agent โ€” screenshot + task to next action python_version: "3.12" startup_duration_timeout: 30m --- # Game2Player โ€” Visual Game-Playing Agent **Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development** ([paper](https://arxiv.org/abs/2611.24154) ยท [code](https://github.com/game2player/Game2Player)) This Space runs the released **Game2Player-9B** checkpoint (a Qwen3.5-9B fine-tune) with transformers, replicating the authors' serving recipe โ€” bf16, โ‰ค1,048,576 pixels per screenshot, thinking disabled, temperature 1.0 โ€” and their PlayerBench prompt and action decoder, both ported 1:1 from the official repository. Give the model a **game screenshot** and a **task description** (plus the game's rules and allowed keys), and it returns the next keyboard or mouse action as a `computer_use` tool call, decoded to a canonical action and drawn on the screenshot: click coordinates, drag segments, key presses. The bundled examples are real tasks from **PlayerBench**, the paper's benchmark: screenshots captured from the benchmark's own games (`107_simon-sequence` and `105_memory-match`, Apache-2.0), with the exact game-rule, role, and control strings from the benchmark's task definitions. ## Usage 1. Pick an example, or upload your own screenshot and describe the task and rules. 2. Hit **Predict next action**. 3. The red crosshair on the output image is where the agent would click; the action summary line shows the decoded action (pixel and normalized 0โ€“1000 coordinates). The model is trained on a 1280ร—720 viewport; any input image is rescaled by the processor to the training pixel budget. Coordinates in the raw tool call are normalized 0โ€“1000 and mapped back to the viewport by the decoder.