README.md

---
license: mit
language:
- en
base_model:
- Efficient-Large-Model/VILA1.5-13b
pipeline_tag: video-text-to-text
---

# LiFT: Leveraging Human Feedback for Text-to-Video Model Alignment

LiFT-Critic is a novel Video-Text-to-Text reward model for synthesized video evaluation.

## 🔧 Installation

1. Clone the github repository and navigate to LiFT folder
```bash
git clone https://github.com/CodeGoat24/LiFT.git
cd LiFT
```
2. Install packages
```
bash ./environment_setup.sh lift
```

## 🚀 Inference

### Run
Please download this public [LiFT-Critic-13b-lora](https://huggingface.co/Fudan-FUXI/LiFT-Critic-13b-lora) checkpoints. 

We provide some synthesized videos for quick inference in `./demo` directory.

```bash
python LiFT-Critic/test/run_critic_13b.py --model-path ./LiFT-Critic-13b-lora
```

# Citation

If you find Euclid useful for your research and applications, please cite using this BibTeX:
```bibtex
@article{LiFT,
  title={LiFT: Leveraging Human Feedback for Text-to-Video Model Alignment.},
  author={Wang, Yibin and Tan, Zhiyu, and Wang, Junyan and Yang, Xiaomeng and Jin, Cheng and Li, Hao},
  journal={arXiv preprint arXiv:2412.04814},
  year={2024}
}
```