wcxhimself/KIVI-Generated-Videos

KIVI Generated Videos

Generated videos and generation artifacts for KIVI: Knowledge-Intensive Video Generation.

• Paper
• Code

Overview

KIVI evaluates whether text-to-video models can produce factually accurate and practically useful videos from information-seeking prompts.

This dataset contains preserved outputs for:

• MiniMax H3
• Seedance 2.5

Each model was evaluated on the same 54 prompts across 18 categories.

Availability note: The paper evaluates nine models. Outputs for HappyHorse 1.0, Seedance 2.0, Wan 2.2, HunyuanVideo 1.5, Helios-Base, LongCat-Video, and LongLive 1.0 were not retained because an unexpected server restart occurred before backup. They are therefore not included.

Browse Videos

Final videos can be viewed individually:

• MiniMax H3
• Seedance 2.5

Each prompt directory contains:

• final_video.mp4: original-quality video
• final_video_compressed.mp4: compressed version for convenient viewing

Structure

├── videos/                    # Individually browsable final videos
│   ├── minimax-h3/
│   └── seedance-2.5/
├── archives/                  # Complete generation artifacts
│   ├── minimax-h3.tar.gz
│   └── seedance-2.5.tar.gz
├── common_outlines/           # Shared video outlines
└── common_script/             # Shared generation scripts

The complete archives additionally contain segment-level videos, prompts, configurations, and reference frames.

Statistics

Model	Prompts	Final outputs	Browsable MP4s	Archive size
MiniMax H3	54	54	108	3.05 GB
Seedance 2.5	54	54	108	3.01 GB

Each final output is provided in both original and compressed form. The dataset also includes 54 shared scripts, 54 primary outlines, and 14 additional segment-level outline files.

Download

Install the Hugging Face client, preferably in an isolated environment:

pip install -U huggingface_hub

Download the complete dataset:

hf download wcxhimself/KIVI-Generated-Videos \
  --repo-type dataset \
  --local-dir KIVI-Generated-Videos

Download one complete archive:

hf download wcxhimself/KIVI-Generated-Videos \
  archives/minimax-h3.tar.gz \
  --repo-type dataset \
  --local-dir .

Replace minimax-h3.tar.gz with seedance-2.5.tar.gz for the other model.

Extraction

tar -xzf ./archives/minimax-h3.tar.gz
tar -xzf ./archives/seedance-2.5.tar.gz

Checksums

minimax-h3.tar.gz
5760ACCF2ADA6E688E32A43512C865CCB33BF03ACAF2BDAE36B5EC684A608E30

seedance-2.5.tar.gz
A9C6C672AF0A27A3EEB0F17CA82FA5656A5196FBDF1973E2443D7689882F6A47

License and Usage

These files are released for research and evaluation purposes. Users must comply with the applicable terms and licenses of the corresponding model providers. The MIT license of the KIVI code repository does not automatically apply to third-party models, services, or generated outputs.

Citation

@misc{wang2026knowledgeintensivevideogeneration,
  title         = {Knowledge-Intensive Video Generation},
  author        = {Chenxu Wang and Mingda Chen},
  year          = {2026},
  eprint        = {2606.01285},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.01285}
}