PhysInOne Benchmark Leaderboard

PhysInOne Benchmark — Visual Physics Learning and Reasoning in One Suite.
Created by vLAR Group / HK PolyU.

How to Participate

  1. Sign in with Hugging Face and create a dataset under your own account.
  2. Upload scene-level ZIP files with the expected structure.
  3. Open the page of the major task you want to evaluate, choose the subtask, and fill in display name and user_dataset (for example, username/repo).
  4. The system enqueues the job, and the backend worker evaluates it scene by scene asynchronously.
  5. Queue and submission history are visible only for your signed-in account.
Major Task

Video Generation

Generate a physically plausible video from task-provided conditioning input. Use this page for layout requirements, submission, queue status, and registered subtasks.

<scene_id>.zip
|-- CineCamera_0/
|   `-- rgb/
|       |-- 0075.jpg
|       `-- ...
|-- CineCamera_1/
|   `-- rgb/
`-- ...
  • Each subtask keeps its own primary metric.
  • Full and mini variants should be treated as separate subtasks.
  • Add subtask-specific notes in the registered task metadata when those tasks are wired in.

Submit

Sign in with Hugging Face and submit a dataset under your own account in `username/repo` format.

Display Name
Subtask
User Dataset

Video Generation submissions are not enabled yet. This page keeps the same submit layout for consistency.

Queue

This queue only shows submissions for the subtasks under this page.

Queue will appear here after this major task is connected to backend evaluation.

Registered Subtasks

No frontend-registered subtasks are available for Video Generation yet.