VideoVerse: Does Your T2V Generator Have World Model Capability to Synthesize Videos?

March 29, 2026 · View on GitHub

Official repository for the paper "VideoVerse: Does Your T2V Generator Have World Model Capability to Synthesize Videos?".

🌐 Webpage 📖 Paper 🤗 Huggingface Dataset 🏆 Leaderboard

🔥 News

[2026.03.18] 🔥 We have updated the full text and added a sub-question validation mechanism to Dynamic. Please check the Homepage and arXiv for the latest results.
[2025.10.17] 🔥 Sora2 achieves a new SOTA! Cool! The videos from Sora2 have been released.
[2025.10.17] 🔥 VideoVerse Dataset are now release!
[2025.10.05] 🔥 Evaluation code, prompt files, and evaluation questions are now released!

VideoVerse is a benchmark for evaluating current T2V (Text-to-Video) models from the perspective of world models. It covers both static and dynamic dimensions, with each dimension containing several world-model-level and basic-level evaluation aspects essential for T2V models.

🔧 How to Start

VideoVerse is organized for easy benchmarking of T2V models:

📑 Prompt Files

Our prompt file is located at prompt/prompts_of_VideoVerse.json, which contains all prompts and their corresponding evaluation information.

Generate Videos & Directory Organization

For each dictionary in the prompt file, generate a video using the "t2v_prompt" field under "t2v_following_prompt", and name the video using the corresponding key. The resulting video files should be organized as follows:

VIDEO_FOLDER/
  8f348e44-546c-4319-aefa-b860c02d9cbc.mp4
  dc4fa681-8b4a-413d-9571-29af7aa36c2e.mp4
  DICT_KEY.mp4

🧪 Evaluation with Gemini 2.5 Pro

As recommended in our paper, we suggest using Gemini 2.5 Pro for video evaluation. Here is how to integrate your Gemini 2.5 Pro API with our evaluation code:

Set variables:
According to the official Gemini 2.5 Pro API, you need to upload your video to a platform that provides a public URL, such as:
https://www.naptmn.cn/Homepage_of_VideoVerse/static/videos/banner_video.mp4
We recommend using GitHub as a free video hosting platform.
Run evaluation:
After obtaining the public URL for your video, replace the {BASEURL} variable in scripts/eval_with_Gemini_like_video_url.py.
Then, update the request_api_video function in the same script to use your own Gemini 2.5 Pro API endpoint.
Run the evaluation script:
```
python scripts/eval_with_Gemini_like_video_url.py
```
The results will be saved as eval_res.json in the eval_res directory.

For sub-question evaluation, you need to run
```
python scripts/eval_sub_question_eval_sub_question_with_Gemini_like_video_urlwith_Gemini_like_video_url.py
```
Summarize results:
After obtaining eval_res.json, update the eval_res_path in scripts/cal_acc.py and eval_sub_question_acc.py to point to your result file.
Run the script, and a new file ending with _report.json will be generated in the same directory, containing the evaluation summary

📊 Evaluation with Open-Source VLMs (e.g., QwenVL)

When using open-source VLMs, you do not need to upload videos to a public URL. Simply place them in the eval_videos directory.

Replace the model checkpoints:
Update the checkpoints path in scripts/eval_with_other_vlm.py.
Run with other VLMs:
```
python scripts/eval_with_other_vlm.py
```

📌 Citation

If you find the code useful for your work, please star this repo and consider citing:

@article{wang2025videoverse,
  title={VideoVerse: How Far is Your T2V Generator from a World Model?},
  author={Wang, Zeqing and Wei, Xinyu and Li, Bairui and Guo, Zhen and Zhang, Jinrui and Wei, Hongyang and Wang, Keze and Zhang, Lei},
  journal={arXiv preprint arXiv:2510.08398},
  year={2025}
}

🙋‍♂️ Questions?

Open an issue.

Enjoy using VideoVerse! 🚀🖼️🤖