Spaces:
Sleeping
Sleeping
| import gradio as gr | |
| import pandas as pd | |
| # Static data | |
| STATIC_DATA = [ | |
| ["VLM", "w/o WM", "–", "RGB", "72B", 50.27, 6.24], | |
| ["Image Gen.", "PathDreamer [36]", "Viewpoint", "RGB-D; Sem; Pano", "0.69B", 56.99, 5.28], | |
| ["Image Gen.", "SE3DS [11]", "Viewpoint", "RGB-D; Pano", "1.1B", 57.53, 5.29], | |
| ["Video Gen.", "NWM [25]", "Trajectory", "RGB", "1B", 57.35, 5.68], | |
| ["Video Gen.", "SVD [6]", "Image", "RGB", "1.5B", 57.71, 5.29], | |
| ["Video Gen.", "LTX-Video [5]", "Text", "RGB", "2B", 56.08, 5.37], | |
| ["Video Gen.", "Hunyuan [4]", "Text", "RGB", "13B", 57.71, 5.21], | |
| ["Video Gen.", "Wan2.1 [23]", "Text", "RGB", "14B", 58.26, 5.24], | |
| ["Video Gen.", "Cosmos [1]", "Text", "RGB", "2B", 52.27, 5.898], | |
| ["Video Gen.", "Runway", "Text", "–", "–", "–", "–"], | |
| ["Video Gen. Post-Train", "SVD† [6]", "Action", "RGB; Pano", "1.5B", 60.98, 5.02], | |
| ["Video Gen. Post-Train", "LTX† [5]", "Action", "RGB; Pano", "2B", 57.53, 5.49], | |
| ["Video Gen. Post-Train", "WAN2.1† [23]", "Action", "RGB; Pano", "14B", "XXX", "XXX"], | |
| ["Video Gen. Post-Train", "Cosmos† [1]", "Action", "RGB; Pano", "2B", 60.25, 5.08], | |
| ] | |
| COLUMNS = ["Model Type", "Method", "Control Type", "Input Type", "#Param.", "Acc. ↑", "Mean Traj. ↓"] | |
| def create_leaderboard(): | |
| df = pd.DataFrame(STATIC_DATA, columns=COLUMNS) | |
| return df | |
| # Create the Gradio interface | |
| with gr.Blocks(title="World-in-World: Building a Closed-Loop World Interface to Evaluate World Models", theme=gr.themes.Soft()) as demo: | |
| gr.HTML("<h1 style='text-align: center; margin-bottom: 1rem'>🏆 World-in-World: Building a Closed-Loop World Interface to Evaluate World Models</h1>") | |
| gr.Markdown(""" | |
| **Performance comparison across vision-language models, image generation, and video generation models.** | |
| 📊 **Metrics:** Acc. ↑ (Accuracy - higher is better) | Mean Traj. ↓ (Mean Trajectory error - lower is better) | |
| """) | |
| with gr.Tabs(): | |
| with gr.TabItem("📊 Leaderboard"): | |
| leaderboard_table = gr.DataFrame( | |
| value=create_leaderboard(), | |
| headers=COLUMNS, | |
| datatype=["str", "str", "str", "str", "str", "number", "number"], | |
| interactive=False, | |
| wrap=True | |
| ) | |
| with gr.TabItem("📝 About"): | |
| gr.Markdown(""" | |
| # World-in-World: Building a Closed-Loop World Interface to Evaluate World Models | |
| This leaderboard showcases performance metrics across different types of AI models in world modeling tasks: | |
| ## Model Categories | |
| - **VLM**: Vision-Language Models | |
| - **Image Gen.**: Image Generation Models | |
| - **Video Gen.**: Video Generation Models | |
| - **Video Gen. Post-Train**: Post-training specialized Video Generation Models | |
| ## Metrics Explained | |
| - **Acc. ↑**: Accuracy score (higher values indicate better performance) | |
| - **Mean Traj. ↓**: Mean trajectory error (lower values indicate better performance) | |
| ## Notes | |
| - † indicates post-training specialized models | |
| - XXX indicates results pending/unavailable | |
| - – indicates not applicable or not available | |
| *Results represent performance on world modeling evaluation benchmarks and may vary across different evaluation settings.* | |
| """) | |
| if __name__ == "__main__": | |
| demo.launch() |