CTVid-Bench: Clear Text Video Restoration Benchmark

CTVid-Bench is a benchmark and data release for evaluating text-centric video restoration and multimodal reasoning under different visual quality conditions. This repository contains a public testing benchmark and a separated training source release.

Dataset Structure

Testing Benchmark

• spatial/VQA_img/{GT,blur,downsample_x4}/: image frames for spatial ImageQA.
• spatial/VQA_json/{GT,blur,downsample_x4}_VQA_testing/: spatial QA v2 annotations.
• temporal/images/{GT,blur,downsample_x4}/: extracted temporal frames and frame-level metadata.
• temporal/videos_vqa/{GT,blur,downsample_x4}/: temporal video clips and temporal QA v2 annotations.

Testing covers three public quality variants:

• GT: high-quality reference frames and videos.
• blur: degraded blurry inputs.
• downsample_x4: low-resolution degraded inputs.

Private research methods such as DOVE, MIMO-UNetPlus, and VSR_S3Diff are not included in this public release.

Training Source

• training/source/<collection>/<clip>/frame_*.png: original training frames.
• training/source/<collection>/<clip>/frame_*.json: paired frame-level OCR annotations.

Current training source release:

• 19 collections
• 3,252 clip folders
• 51,924 frame images
• 51,924 paired frame JSON annotations
• 103,848 training source files
• about 119.6 GiB

The training source upload excludes OCR visualization images, historical QA summary files, and intermediate local generation artifacts.

Current Release Status

Component	Status
Testing spatial ImageQA	Available
Testing temporal VideoQA	Available
Testing media	Available
Training source frames	Available
Training frame-level OCR JSON	Available
Refined training ImageQA	In progress, not uploaded yet

Refined training QA will be added later under:

training/qa/spatial/VQA_json/GT_VQA_training/

Usage

Download the Testing Benchmark

huggingface-cli download jinlong17/CTVid-Bench \
  --repo-type dataset \
  --include "spatial/*" "temporal/*" \
  --local-dir ./CTVid-Bench

Download the Training Source

huggingface-cli download jinlong17/CTVid-Bench \
  --repo-type dataset \
  --include "training/source/*" "training/README.md" "training/training_manifest.json" \
  --local-dir ./CTVid-Bench

Use for Evaluation

1. Download the testing benchmark.
2. Run your restoration or multimodal model on the selected quality variant.
3. Evaluate spatial ImageQA using spatial/VQA_json/.
4. Evaluate temporal VideoQA using temporal/videos_vqa/.

Evaluation Dimensions

• Spatial text understanding: recognizing and grounding scene text in image regions.
• Temporal text understanding: tracking text visibility, motion, size, and dynamics across clips.
• Restoration robustness: measuring how degraded or restored visual quality affects text-centric reasoning.

Task Categories

• Fill-in questions
• Fill-in-the-blank questions
• Multiple-choice questions
• True/false questions
• Temporal localization and motion questions
• Temporal visibility and summary questions

Data Boundaries

This Hugging Face release is a public release view. It does not include private method outputs, local backup folders, generated intermediate folders, OCR visualization images, or unfinished training QA refinement outputs.

For exact file counts and release metadata, see:

• dataset_manifest.json
• training/training_manifest.json

Citation

Coming soon.