language:
- en
license: apache-2.0
size_categories: - 10K<n<100K
task_categories: - text-to-video
pretty_name: VPO
tags: - Prompt Optimization
- Video Generation
Dataset Card for VPO
Dataset Description
- Repository: https://github.com/thu-coai/VPO
- Paper: https://huggingface.co/papers/2503.20491
Data Summary
To bridge the gap between training and inference in video generation models, we present VPO—a principle-driven framework designed to generate harmless, accurate, and helpful prompts for high-quality video generation.
We release an SFT dataset containing 10k samples constructed using gpt-4o. In addition, we provide the DPO datasets derived from CogVideoX-2B and CogVideoX-5B.
Please refer to our paper for further details.
Supported Tasks
The VPO dataset is designed for prompt optimization for video generation models.
The SFT subset is a good choice for enhancing foundational prompt optimizaiton capabilities. The DPO training data is designed for preference learning of a specific video model, combining both video-level and text-level feedback.
Language
The data in VPO are in English.
Data Structure
Data Fields
Our dataset is formatted for quick development with LLaMA-Factory.
Additional Information
Source Data
Our original prompts come from the VidProM dataset.