VGA-Bench Data Files

Official evaluation toolkit: https://github.com/BestiVictory/VGA-Bench

This directory releases the VGA-Bench Prompt Suite, a subset of generated videos, and their expert annotations. This README explains what each data file contains and how the files are related.

The current release contains:

• 1,016 bilingual Chinese–English prompts;
• 8,349 generated videos;
• metadata for every released video;
• Aesthetic Quality, Aesthetic Tagging, and Generation Quality annotations.

The remaining videos and annotations will be released progressively.

Directory Structure

.
├── README.md
├── videos_name.csv
├── prompts/
│   ├── 120-base_prompts.csv
│   ├── 200-aesthetic-prompts.csv
│   ├── 220-tag-prompts.csv
│   └── 476-GQprompts.csv
├── videos/
├── anno/
│   ├── Aes_result.csv
│   ├── Tag_results.json
│   └── Gen_results/
└── codes_cvpr/                     # See the GitHub link in the Code section

The directory consists of five main parts:

Part	Description
prompts/	The complete VGA-Bench Prompt Suite
videos/	Generated videos included in the current release
videos_name.csv	Master index connecting videos, models, prompts, and annotations
anno/	Aesthetic Quality, Aesthetic Tagging, and Generation Quality expert annotations
codes_cvpr/	See the code repository: xxx

1. Prompt Suite: prompts/

The prompts/ directory contains 1,016 prompts divided into four groups.

File	Count	Description
120-base_prompts.csv	120	Basic generation quality, including clarity, noise, exposure, stability, and distortion
200-aesthetic-prompts.csv	200	Aesthetic quality, including overall aesthetics, composition, lighting, color, depth of field, and human presentation
220-tag-prompts.csv	220	Aesthetic tags, including composition, light sources, shot size, depth of field, saturation, brightness, color temperature, and contrast
476-GQprompts.csv	476	Fine-grained generation quality, including people, objects, actions, scenes, motion, interaction, and physical plausibility

All four files provide both Chinese and English prompts, although their column names differ slightly.

120-base_prompts.csv

Field	Description
id	Prompt identifier
维度	Main evaluation dimension targeted by the prompt
prompt	Chinese prompt
en_prompt	English prompt

200-aesthetic-prompts.csv

Field	Description
编号	Prompt identifier
各组合维度	Aesthetic dimension or dimension combination targeted by the prompt
Prompt	Chinese prompt
en_prompt	English prompt

220-tag-prompts.csv

In addition to the Chinese and English prompts, this file lists the target attributes explicitly requested by each prompt, such as composition, number and position of light sources, light quality and color, shot size, depth of field, saturation, brightness, color temperature, and contrast.

An empty cell means that the corresponding attribute is not explicitly constrained by that prompt. It does not mean that the video's expert annotation lacks that attribute.

476-GQprompts.csv

Field	Description
ind	Prompt identifier
prompt	English prompt
prompt_cn	Chinese prompt
types	Evaluation types or dimensions involved in the prompt
tokens	Number of prompt tokens

2. Videos: videos/

The videos/ directory contains the 8,349 generated videos included in the current release. The following formats are used:

• .mp4
• .webm
• .gif
• .webp

All formats are treated as video samples. Do not remove the filename extension, because the complete filename is the key used to connect the video with its metadata and annotations.

3. Video Index: videos_name.csv

videos_name.csv describes which model and prompt were used to generate each released video.

Field	Description
name	Video filename, including its extension
modelname	Name of the video-generation model
en_prompt	English generation prompt
dimension	Source Prompt Suite group

The dimension field has four possible values:

Value	Corresponding file
120-base_prompts	prompts/120-base_prompts.csv
200-aesthetic-prompts	prompts/200-aesthetic-prompts.csv
220-tag-prompts	prompts/220-tag-prompts.csv
476-GQprompts	prompts/476-GQprompts.csv

In the current release, the 8,349 rows in videos_name.csv correspond one-to-one with the 8,349 files under videos/.

4. Aesthetic Quality: anno/Aes_result.csv

This file contains video-level Aesthetic Quality scores.

Field	Description
score	Overall aesthetic-quality score
composition	Composition
shotsize	Shot size
lighting	Lighting
visualtone	Visual tone
color	Color
depthoffield	Depth of field
expression	Human expression
costume	Costume
makeup	Makeup
video_id	Corresponding video filename
modelname	Video-generation model

The same video_id may occur in multiple rows when different annotation sources or annotation rounds are retained. Records should therefore be joined and organized by video_id.

For every video included in this annotation set, the released Aesthetic Quality data contains the complete set of aesthetic attributes rather than only the dimensions specified by the video's original prompt.

5. Aesthetic Tagging: anno/Tag_results.json

This file contains photographic and cinematic attributes in a multimodal dialogue format:

{
  "messages": [
    {"role": "user", "content": "<video_0>What is the composition of this video?"},
    {"role": "assistant", "content": "... The video uses a centered composition."}
  ],
  "videos": ["example.mp4"]
}

Field	Description
messages	Questions about the video attributes and the corresponding expert answers
videos	Associated video filename(s)

Every video included in this file is annotated with the complete Aesthetic Tagging attribute set, such as composition, lighting, shot size, depth of field, saturation, brightness, color temperature, and contrast. The annotations are not limited to attributes explicitly specified by the original prompt.

6. Generation Quality: anno/Gen_results/

anno/Gen_results/ contains the Generation Quality expert annotations aligned with the current public video release. It contains 31 evaluation-dimension files.

Each JSON file corresponds to one evaluation question or dimension. The filename itself describes the dimension, for example:

视频是否清晰.json
动作是否真实合理.json
物体是否和文本一致.json
静止的视频内容是否稳定不随时间变动.json

Each annotation record has the following structure:

{
  "prompt": "English prompt",
  "prompt_cn": "Chinese prompt",
  "model": "model name",
  "score": 3.0,
  "video_id": "example.webm"
}

Field	Description
prompt	English generation prompt
prompt_cn	Chinese generation prompt
model	Video-generation model
score	Dimension-specific expert score; different dimensions use different discrete score ranges
video_id	Corresponding video filename

Generation Quality does not use one universal 1–5 scale. The available values and their meanings are defined separately for each question in codes_cvpr/VGQA/prompt_qo_deduped.csv. Depending on the dimension, the maximum valid quality score may be 2, 3, 4, or 5.

Some questions also include -1 as a special option indicating that the target content is absent, the prompt does not specify the relevant content, the question is not applicable, or the question should be invalidated. It is not a universal lowest-quality score and should not be treated as an ordinary quality level.

Use the following files as the source of truth for score bounds, option meanings, and normalization:

• codes_cvpr/VGQA/prompt_qo_deduped.csv: defines the questions and available score options for each prompt;
• codes_cvpr/VGQA/prompt_gqa_mayi.txt: defines the VGQA question-answering task and output format;
• codes_cvpr/VGQA/vgqa_analyze.py: computes raw and normalized scores using the option range of each question.

Generation Quality uses applicable-dimension annotation. A video is included in a dimension file only when that dimension is meaningful for its prompt and content.

If a video does not appear in a particular JSON file, it means that the dimension was not annotated or is not applicable to that video. It should not be interpreted as a score of zero.

All Generation Quality records in this directory can be joined to the name field in videos_name.csv through video_id. The 31 files contain 13,440 dimension-level annotations covering 6,793 Generation Quality videos.

7. Code

The VGA-Bench training, inference, and official evaluation code is available at: xxx.

Replace xxx with the actual GitHub repository URL before publication.

The released Aesthetic Quality and Aesthetic Tagging data provides the complete attribute set for every covered video. The official code, however, performs inference, scoring, and aggregation strictly according to the Prompt Suite–dimension correspondence defined in the paper.

8. How the Files Are Connected

The main relationships between the files are:

prompts/*.csv
      │ generation text and target evaluation dimensions
      ▼
videos_name.csv
      │ name = video filename
      ├──────────────► videos/name
      ├──────────────► video_id in anno/Aes_result.csv
      ├──────────────► videos in anno/Tag_results.json
      └──────────────► video_id in anno/Gen_results/*.json

Recommended workflow:

1. Read videos_name.csv to obtain the list of currently released videos.
2. Use name to locate the corresponding file under videos/.
3. Match name with video_id in the Aesthetic Quality annotations.
4. Match name with videos in the Aesthetic Tagging annotations.
5. Match name with video_id in the Generation Quality annotations.
6. Keep the filename extension and use UTF-8 when reading Chinese fields and filenames.

9. Progressive Release

The current version includes the complete Prompt Suite and the first batch of 8,349 videos with their annotations. The remaining videos and corresponding annotations will be released progressively.

For reproducibility, record the release tag or commit hash used in your experiments.