ClearText-Video

A Large-Scale Text-Centric Video Dataset Bridging
Video Restoration and Scene-Text Enhancement

ECCV 2026

Project Page · arXiv Paper · Dataset · Quick Start · Results

---

Overview

ClearText-Video (CTVid) is a large-scale, scene-text-aware benchmark for studying text-centric video understanding under controlled quality variation. It pairs every high-quality source with content-matched degraded and restored variants, connecting Text-Centric Video Restoration with Multi-Quality VideoQA.

The central question is simple: does a video that looks better also preserve the textual evidence a multimodal model needs?

At a glance

Videos	Frames	Human-verified annotations	Spatial & temporal QA	Languages
4,639	550K+	1.6M	220K+	Chinese + English

• 4,327 training videos and 312 testing videos.
• Quality-controlled HQ, DQ-Low_res, DQ-Blur, RQ-DOVE, RQ-MIMO, and RQ-S3DIFF conditions in the full benchmark.
• Evaluation of 18 restoration methods and 16 multimodal LLMs.
• Two rounds of annotation and correction by 16 annotators, starting from 6.4K+ candidate videos.

Important

The current public Hugging Face release contains GT, blur, and downsample_x4. DOVE, MIMO and S3DIFF are full-paper evaluation conditions and are not part of the current public download.

Authors

Jinlong Li^†, Jiaming Ding, Dingfu Lu^‡, Malcolm Hsiu^‡, Chuang Ke, Kangning Yang, Bochen Guan, Lan Fu, Jie Cai, Huiming Sun, Zibo Meng

OPPO US AI Center · University of Wisconsin–Madison · University of California San Diego

_{^* Equal contribution · ^† Corresponding author · ^‡ Work done during internships at OPPO US AI Center}

Dataset

Quality regimes

Regime	Conditions	Purpose
HQ	Original high-quality video	Reference textual evidence
DQ	Low resolution, locally variant blur	Controlled robustness testing
RQ	DOVE, MIMO, S3DIFF	Test whether restoration preserves or changes evidence

Benchmark tasks

1. Text-Centric Video Restoration — image super-resolution, video super-resolution and video deblurring, evaluated with visual-quality and text-fidelity metrics.
2. Spatial VideoQA — text recognition and grounding within frame-level context, evaluated with Accuracy (Acc), Unbiased Accuracy (UAcc), Overconfidence (OC) and Answer Abstention (Abs).
3. Temporal VideoQA — text presence, localization, motion, scale and boundary reasoning across video frames.

Dataset split and annotation details

Split	Videos	Notes
Train	4,327	Large-scale training supervision
Test	312	74% offline / 26% online; balanced Chinese and English coverage by source

Question difficulty is distributed as 46% easy, 37% medium, and 17% hard.

Benchmark results

The final paper shows that visual enhancement does not guarantee textual fidelity or downstream reasoning gains.

• Across 16 MLLMs, low resolution reduces spatial QA accuracy by 3.14 points from HQ, while blur reduces it by 6.05 points.
• Gemini-2.5-pro reaches the best spatial HQ accuracy at 71.67%.
• Claude-Sonnet-4.6 ranks first across all five temporal quality conditions, reaching 60.02% on HQ and 60.37% on RQ-DOVE.
• Qwen2.5-VL-7B-SFT is the strongest open-source model by accuracy across all six spatial quality conditions, improving 6.31–10.11 points over its base model.

Best spatial accuracy by condition

Condition	Best model	Accuracy (%)
HQ	Gemini-2.5-pro	71.67
DQ-Low_res	Gemini-2.5-pro	65.00
DQ-Blur	Claude-Sonnet-4.6	60.00
RQ-DOVE	Gemini-2.5-flash	66.67
RQ-MIMO	Claude-Sonnet-4.6	70.00
RQ-S3DIFF	Gemini-2.5-pro	70.00

Quick start

1. Install

git clone https://github.com/jinlong17/CTVid-Bench.git
cd CTVid-Bench
pip install -r requirements.txt

2. Download the public test variants

python tools/download_data.py \
  --split test \
  --variants GT blur downsample_x4

See DATA.md for the Hugging Face layout, paths and manual download options.

3. Run spatial VideoQA

python evaluation/spatial/run_eval.py \
  --dataset GT \
  --config configs/default.yaml \
  --output_dir outputs/spatial/

python evaluation/spatial/metrics.py \
  --results_dir outputs/spatial/ \
  --dataset GT

4. Run temporal VideoQA

python evaluation/temporal/run_eval.py \
  --dataset GT \
  --config configs/default.yaml \
  --output_dir outputs/temporal/

python evaluation/temporal/metrics.py \
  --results_dir outputs/temporal/ \
  --dataset GT

Repository layout

CTVid-Bench/
├── configs/default.yaml
├── docs/
│   ├── review_spatial_qa.md
│   └── review_temporal_qa.md
├── evaluation/
│   ├── spatial/
│   └── temporal/
├── project_page/                 # GitHub Pages source
├── tools/download_data.py
├── DATA.md                       # Public data release guide
└── requirements.txt

Citation

@inproceedings{li2026cleartextvideo,
  title     = {ClearText-Video: A Large-Scale Text-Centric Video Dataset
               Bridging Video Restoration and Scene-Text Enhancement},
  author    = {Li, Jinlong and Ding, Jiaming and Lu, Dingfu and Hsiu, Malcolm
               and Ke, Chuang and Yang, Kangning and Guan, Bochen and Fu, Lan
               and Cai, Jie and Sun, Huiming and Meng, Zibo},
  booktitle = {European Conference on Computer Vision},
  eprint    = {2608.28784},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url       = {https://arxiv.org/abs/2608.28784},
  year      = {2026}
}

License

• Code: MIT License
• Public data: CC BY 4.0

---

_{Paper available as arXiv:2608.28784 · submitted 28 August 2026.}