license: mit
pipeline_tag: video-text-to-text
library_name: transformers
PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models
This repository contains the PhyDetEx model, designed for detecting and explaining physically implausible content in videos generated by Text-to-Video (T2V) models. PhyDetEx introduces a lightweight fine-tuning approach, enabling Vision-Language Models (VLMs) to not only detect physically implausible events but also generate textual explanations on the violated physical principles.
This work was presented in the paper: PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models
- 📖 Paper: PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models
- 💻 Code: https://github.com/Zeqing-Wang/PhyDetEx
- 🤗 PID Dataset: https://huggingface.co/datasets/NNaptmn/PhyDetExDatasets
🔥 News
- [2025.12.01] 🔥 We release the PID Dataset and the PhyDetEx Model!
Introduction
PhyDetEx is a model designed for detecting physical implausible content. Additionally, to better address and test physical implausible content detection, we provide the PID Physical Implausibility Detection dataset.
🔧 How to Start
Download the PID Test split
Download PID_Test_split.zip from 🤗 PID Dataset, place it in the Data/PID_test directory, and organize it as follows:
PID_test/
pos/
video_xxx.mp4
......
neg/
video_xxx.mp4
......
anno_file.json
### Download the PhyDetEx
Download PhyDetEx from [🤗 PhyDetEx Model](https://huggingface.co/NNaptmn/PhyDetEx).
### Prepare the Environment
```bash
pip install -r requirements.txt
Please note that the version of transformers may affect specific metrics, so it is recommended to use the version specified in requirements.txt.
Set variables
In benchmark_on_pid_test_split.py, set the corresponding path for PhyDetEx, then run:
python benchmark_on_pid_test_split.py
The resulting ./res/res_on_pid_test.json will contain the F1 Score, Acc Plausible, and Acc Implausible.
Get the reasoning score
Deploy any LLM using lmdeploy. In the paper, we report results using LLaMa3 8B.
In infer_llm_score_for_pid_test_lmdeploy.py, set the corresponding port and evaluation file path, then run:
python infer_llm_score_for_pid_test_lmdeploy.py
🧪 Test on ImpossibleVideos
You can download and process the Physical Law-related data from Impossible-Videos. Alternatively, we recommend directly downloading our preprocessed data: 🤗 PID Dataset "ImpossibleVideos_Physical_Law_Only.zip", and placing it in Data/PID_test. The remaining steps are the same as for the PID Test.
Please note that the scripts for running ImpossibleVideos are benchmark_on_impossible_videos.py and infer_llm_score_for_impossible_video_lmdeploy.py.
🔧 Train the PhyDetEx
In the 🤗 PID Dataset, we also provide the PID Train Split. For training PhyDetEx, we use LLaMA-Factory.
Acknowledgement
We heavily borrow the data and code from ImpossibleVideos, and LLaMA-Factory. Thanks for sharing their code.
📌 Citation
If you find the code useful for your work, please star this repo and consider citing:
@article{wang2025phydetex,
title={PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models},
author={},
journal={arXiv preprint arXiv:2512.01843},
year={2025}
}