Spaces:
Running on Zero
Running on Zero
metadata
title: Code-as-World VL Demo
emoji: 🌍
colorFrom: indigo
colorTo: red
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
short_description: Video-based quantitative physical reasoning VLM
python_version: '3.12'
startup_duration_timeout: 30m
license: apache-2.0
Code-as-World VL Demo
Code as Worlds introduces executable world representations for physical reasoning. This demo showcases the Code-as-World-VL vision-language model, which performs quantitative physical reasoning from videos — estimating object sizes, velocities, and distances from visual evidence.
Usage
- Upload a short video clip.
- Enter a physics measurement question (e.g. "What is the length of the object in cm?").
- Optionally provide prior information (e.g. "ruler calibre = 1 cm") that helps ground the measurement.
- Click Analyze and the model streams its numerical answer.
Models
| Label | Repo | Params |
|---|---|---|
| 9B (recommended) | MirroS-Lab/Code-as-World-VL-9B |
9B |
| 4B (faster) | MirroS-Lab/Code-as-World-VL-4B |
4B |
Example assets
Example videos are from the QuantiPhy validation set (CC-BY-4.0). Questions and priors are drawn from the same dataset.
Acknowledgements
- Model: MirroS-Lab (Apache 2.0)
- Example videos: QuantiPhy validation set (CC-BY-4.0)
- Base architecture: Qwen3.5