GenState-AI

GenState-AI is a benchmark for fine-grained text-to-video retrieval over AI-generated videos. It contains 956 human-verified clips in four visual domains. Main videos are paired with text prompts; available hard negatives provide controlled temporal/state or semantic contrasts.

Directory Structure

video_aacl/
├── Toy/            # Toy subset
├── Home/           # Home subset
├── Manipulation/   # Manipulation subset
├── Cartoon/        # Cartoon subset
├── metadata.jsonl  # Global metadata index for all 956 videos

Each category directory follows a similar structure:

Toy/
├── video001_toy_main.mp4
├── video001_toy_hn_semantic.mp4
├── video001_toy_hn_temporal.mp4
├── ...
└── video_toy_main.jsonl

The video_{category}_main.jsonl file contains textual descriptions for the available main videos in that category.

---

File Naming Convention

All video files follow the naming format:

video{id}_{category}_{type}.mp4

where:

• category is one of:

  • toy
  • home
  • manipulation
  • cartoon

• type is one of:

  • main: target video associated with a text prompt
  • hn_semantic: semantic hard-negative video
  • hn_temporal: temporal hard-negative video

• id is a category-specific sequential identifier starting from 1.

Videos belonging to the same group share the same ID and category.

For example:

video001_toy_main.mp4
video001_toy_hn_semantic.mp4
video001_toy_hn_temporal.mp4

These videos belong to the same group:

video001_toy

Category directory names use an initial capital letter:

Toy/
Home/
Manipulation/
Cartoon/

while category names inside filenames and metadata remain lowercase.

---

Category-Level JSONL Files

Each category directory contains a:

video_{category}_main.jsonl

file describing the available main videos.

For example:

{
  "video_id": "video001_toy_main",
  "group_id": "video001_toy",
  "type": "main",
  "prompt": "on a wooden table, ..."
}

Fields

• video_id Unique identifier of the main video. It matches the corresponding video filename without the .mp4 extension.

• group_id Identifier shared by videos belonging to the same retrieval group.

  For example:

video001_toy_main
video001_toy_hn_semantic
video001_toy_hn_temporal

all share:

group_id = video001_toy

• type Video type. Entries in these category-level JSONL files use:

main

• prompt Natural-language description of the target video.

Only main videos are associated with prompts.

---

Global metadata.jsonl

The root-level:

metadata.jsonl

provides a global index covering all 956 video files.

Each line corresponds to one video.

Example:

{
  "file_name": "Cartoon/video001_cartoon_hn_semantic.mp4",
  "video_id": "video001_cartoon_hn_semantic",
  "group_id": "video001_cartoon",
  "category": "cartoon",
  "type": "hn_semantic"
}

For a main video:

{
  "file_name": "Cartoon/video001_cartoon_main.mp4",
  "video_id": "video001_cartoon_main",
  "group_id": "video001_cartoon",
  "category": "cartoon",
  "type": "main",
  "prompt": "cartoon style, ..."
}

Metadata Fields

• file_name Relative path to the corresponding video file.

• video_id Unique identifier of the video.

• group_id Identifier used to associate the target video with its hard-negative videos.

• category One of:

toy
home
manipulation
cartoon

• type One of:

main
hn_semantic
hn_temporal

• prompt Natural-language description of the target video.

  This field is available only for main videos.

---

Evaluation Protocol

The machine-readable evaluation filters are provided in evaluation_protocol.json. Recall@K uses all 195 main-query descriptions, with the corresponding category pool as candidates. Triplet metrics use the 154 query groups containing one main, one hn_temporal, and one hn_semantic video with the same group_id; the other 41 main queries remain in Recall@K and are excluded only from triplet metrics.

Citation and License

Please cite the associated paper:

Minghan Li, Tongna Chen, Tianrui Lv, Yishuai Zhang, Suchao An, and Guodong Zhou. 2026. “GenState-AI: State-Aware Dataset for Text-to-Video Retrieval on AI-Generated Videos.”

The dataset materials in this repository are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt them, including for commercial purposes, provided you give appropriate credit, link to the license, and indicate whether changes were made. Do not imply endorsement by the authors. When using the dataset, please cite the associated paper above. This dataset license does not grant rights to the Wan model or any third-party materials that are not part of this dataset.

Retrieval Structure

The dataset is designed for fine-grained text-to-video retrieval.

A typical retrieval group is organized as:

Text Prompt
    |
    v
video001_toy
    |
    ├── video001_toy_main.mp4          # Target / positive video
    ├── video001_toy_hn_semantic.mp4   # Semantic hard negative
    └── video001_toy_hn_temporal.mp4   # Temporal hard negative

The main video represents the target associated with the text prompt, while the hard-negative videos are designed to be visually or semantically similar but differ in important semantic or temporal details.

This structure enables evaluation of fine-grained video-text alignment and retrieval under challenging hard-negative settings.