anon-submission-7k2/vidscribe-benchmark

Anonymous Video Text Rendering Benchmark

Overview

This repository contains an anonymized benchmark release for evaluating text rendering and temporal behavior in video generation systems. It includes ground-truth metadata and the reference media required by the benchmark.

This release does not contain generated model outputs, model weights, evaluation metrics, evaluation results, source code, annotation histories, logs, or annotator information.

Dataset composition

Task	Description	Samples	Included media
I2V	First-frame image-conditioned video generation	229	229 reference images
T2V	Text-to-video generation	250	Prompts and ground truth only
V2V	Source-video-conditioned video generation	185	185 source videos
R2V	Reference-image-conditioned video generation	139	139 reference images
Total		803	553 media files

Repository structure

benchmark/
├── I2V/
│   ├── images/
│   ├── metadata_I2V_GT.json
│   └── metadata_I2V_GT_round2.json
├── T2V/
│   ├── metadata_T2V_GT.json
│   └── metadata_T2V_GT_round2.json
├── V2V/
│   ├── videos/
│   ├── metadata_V2V_GT.json
│   └── metadata_V2V_GT_round2.json
└── R2V/
    ├── images/
    ├── metadata_R2V_GT.json
    ├── metadata_R2V_GT_round2.json
    └── metadata_R2V_GT_round3.json

Each metadata file is a JSON array. Media references use paths relative to benchmark/.

Metadata

Records may contain the following groups of fields:

• Identity and task: sample_id, task, prompt
• Text ground truth: gt_text, gt_text_after, gt_text_switched, gt_text_visible
• Media references: reference_image, source_video
• Text properties: language, font, color, scale, position, hierarchy, carrier, and background separability
• Temporal properties: motion, deformation, occlusion, content dynamics, and temporal evidence
• Benchmark dimensions: F1 through F12

Some fields are task-specific and may be null when they do not apply.

Example:

import json
from pathlib import Path

root = Path("benchmark")
records = json.loads(
    (root / "I2V" / "metadata_I2V_GT.json").read_text(encoding="utf-8")
)

sample = records[0]
reference = root / sample["reference_image"]
print(sample["sample_id"], reference)

Privacy processing

The public release was prepared with the following safeguards:

• Author, annotator, institution, local path, and development-history files were excluded.
• Image EXIF and embedded text metadata, including AIGC provenance chunks, were removed.
• MP4 creator, title, encoder, handler, and other optional metadata were removed.
• Face regions in the included V2V source videos were masked before release.
• Only benchmark ground truth and required reference media are included.

Source media can contain incidental scene text. Such content is part of the benchmark input and may include names, usernames, addresses, or email-like strings visible in the original scene. Users are responsible for appropriate handling and downstream use.

Provenance and licensing

This repository uses license: other because the release contains material with mixed provenance and cannot be accurately represented by a single repository-wide license.

Some I2V reference images are derived from or associated with public research datasets, including TextOCR and LSVT. Their filenames retain source-family prefixes such as real_textocr_ and real_lsvt_. The complete upstream datasets are not included in this repository.

Users must review and comply with the original terms, attribution requirements, and usage restrictions of all upstream sources. Availability in this repository does not transfer ownership or grant rights beyond those provided by the corresponding source licenses. The remaining annotations and benchmark organization are provided for research evaluation during anonymous review.

If a rights holder identifies material that should not be redistributed, please open a repository discussion with the relevant sample identifier so that it can be reviewed or removed.

Intended use

The dataset is intended for:

• Research evaluation of generated text accuracy and visual quality
• Analysis of temporal consistency and text preservation
• Controlled comparison of text-to-video and conditioned video-generation systems

It is not intended for identity recognition, surveillance, personal-data extraction, or production decisions affecting individuals.

Limitations

• The benchmark does not cover every language, writing system, visual style, or video domain.
• Some samples are derived from public web or research data and may reflect source-specific biases.
• Face masking reduces identity exposure but does not guarantee that every contextual identifier has been removed from scene content.
• No evaluation scores or model rankings are included in this release.

Citation

Citation information is withheld during anonymous review and will be added after the review process.