CTVid-Bench

CTVid-Bench is an open-source testing benchmark for clear text video restoration. This release packages the public evaluation media for three methods (GT, blur, downsample_x4) together with the latest QA v2 annotations.

Release Scope

This folder is the Hugging Face staging source for the public testing benchmark. Media files were staged on 2026-06-23; QA-v2 JSONs were refreshed on 2026-07-07 from test_qa_v2_release_20260707/.

Included:

• spatial/VQA_img/{GT,blur,downsample_x4}/ - spatial QA image frames.
• spatial/VQA_json/{GT,blur,downsample_x4}_VQA_testing/ - latest spatial QA v2 annotations.
• temporal/images/{GT,blur,downsample_x4}/ - temporal extracted frames and frame-level JSON metadata.
• temporal/videos_vqa/{GT,blur,downsample_x4}/ - temporal video clips plus latest temporal QA v2 under each method's vqa/ folder.

Excluded:

• training data.
• the full private 6-method research testing set.
• private methods DOVE, MIMO-UNetPlus, and VSR_S3Diff.
• backup/generated/optimized intermediate folders.
• legacy QA JSON from open_source/source/spatial/VQA_json.

Size and Counts

Disk usage by du -sh:

• Release root: 74G
• spatial/VQA_img: 19G
• spatial/VQA_json: 33M
• temporal/images: 55G
• temporal/videos_vqa: 664M

Total files including release docs: 84,929

{
  "json": 36382,
  "md": 1,
  "mp4": 485,
  "png": 48061
}

QA v2 counts:

• Spatial QA v2: 389 JSON files per method, 1,167 total.
• Temporal QA v2: 97 JSON files per method, 291 total.

See dataset_manifest.json for detailed per-method file counts.

Path Policy

The source QA v2 files contained local absolute provenance paths. In this release folder, frame_path and video_path fields have been normalized to release-relative paths such as:

• spatial/VQA_img/GT/CHN/C0011/frame_001.png
• temporal/videos_vqa/GT/videos/CHN_C0011.mp4

Upload Command

hf auth login
hf repos create jinlong17/CTVid-Bench --type dataset --exist-ok
HF_XET_HIGH_PERFORMANCE=1 hf upload-large-folder jinlong17/CTVid-Bench /nfs/AI/VideoEnhancement/Jinlong/flash_file/VSR_test/CTVid-Bench-Research/hf_release/CTVid-Bench-v1 --repo-type dataset --num-workers 8

Uploaded on 2026-07-10. The upload log is:

/nfs/AI/VideoEnhancement/Jinlong/flash_file/VSR_test/CTVid-Bench-Research/hf_release/upload_qa_v2_refresh_20260710_172628.log

Remote verification passed for dataset_manifest.json, temporal/videos_vqa/GT/vqa/EN_C0130_video_questions.json, and spatial/VQA_json/GT_VQA_testing/CHN_test_C0011.json.

Notes

The media files in this staging directory were assembled with hardlinks from the canonical open-source source tree to avoid duplicating the 74GB media payload locally. Hugging Face upload reads them as regular files.