RA-Bench-Extended

RA-Bench-Extended is the expanded media and reproducibility archive associated with RA-Bench, a benchmark for AI-generated video detection that uses real videos as anchors.

The original RA-Bench repository remains the frozen paper benchmark. This repository adds generation variants that are useful for controlled analysis but are not part of the main benchmark average.

Node archive scope

The first release phase contains:

• the frozen RA-Bench main media;
• RA-Bench-HumanProof indices, which reference media in the main split rather than duplicating files;
• RA-Bench-LastMile media;
• seeds 0, 42, and 123 for Wan2.2 dynamic, Wan2.2-Lightning, LTX, and OmniWeaving;
• seeds 0, 42, and 123 for the auxiliary fixed-duration Wan2.2 control.

The canonical node07/node09 archive contains 46,266 unique media entries. Of these, 45,555 are redistributable media files and 711 are retained only as local provenance because their licenses do not permit direct redistribution.

Component	Canonical media	Public media	URL/private provenance
RA-Bench main	17,886	17,375	511
Additional all-seed/control media	20,130	20,130	0
RA-Bench-LastMile T1--T5	8,250	8,050	200
Total	46,266	45,555	711

RA-Bench main contains 1,830 real anchors and 16,056 generated videos:

Source	Family	Videos
Real anchors	Real	1,830
Wan2.2 dynamic	Open-source	1,830
Wan2.2-Lightning	Open-source	1,830
LTX	Open-source	1,830
OmniWeaving	Open-source	1,830
HappyHorse	Closed-source	1,787
Runway	Closed-source	1,805
Kling v3	Closed-source	1,790
Hailuo 2.3 Fast	Closed-source	1,830
Seedance2.0	Closed-source	1,524

The complete first-frame I2V seed pool contains 1,830 videos for every source--seed cell. It covers Wan2.2 dynamic, Wan2.2-Lightning, LTX, OmniWeaving, and the fixed-duration Wan2.2 control at seeds 0, 42, and 123. The four main generators' seed-0 videos already appear in media/main/, so media/all_seeds/ contributes 20,130 additional files without duplication.

RA-Bench-LastMile contains 1,650 transformed videos at each dissemination stage T1--T5, for 8,250 canonical transformed videos. T0 refers to the corresponding canonical main/control media and is not duplicated.

Media that cannot be redistributed is not uploaded. Its metadata remains in the rights-aware release tables so that users can reconstruct eligible real anchors from their original URLs. The public archive contains no reviewer credentials, authentication material, or private review records.

RA-Bench-HumanProof consists of 633 indices into RA-Bench main. It adds no duplicate media; its membership is provided in metadata/ra_bench_humanproof.{csv,jsonl}.

The later HKUST phase will add the controlled T2V and first+last-frame I2V generation settings after a separate inventory and validation pass.

Layout

media/
  main/
  lastmile/
  all_seeds/<source>/seed_<seed>/
metadata/
  ra_bench_main.{csv,jsonl}
  ra_bench_humanproof.{csv,jsonl}
  ra_bench_lastmile.{csv,jsonl}
  real_rights_release.csv
  node_public_media_index.{csv,jsonl}
  node_archive_summary.json
  public_media_checksums.csv
  public_media_checksums_summary.json
tools/
  download_rabench_extended.py
  verify_media_index.py
docs/
  NODE_VIDEO_CONTENT_BREAKDOWN.md

Download and verification

Install huggingface_hub, then download either the full public archive or one component:

python tools/download_rabench_extended.py \
  --local-dir /path/to/RA-Bench-Extended \
  --component all

After downloading, verify every media file against the public size index:

python tools/verify_media_index.py /path/to/RA-Bench-Extended

Release status

The node07/node09 phase is complete. The public repository contains exactly 45,555 media files totaling 186,004,326,480 bytes. Repository paths and file sizes match the frozen public index with zero missing, extra, or mismatched files, and every media object has a published SHA-256 checksum. The checksum table is available under metadata/public_media_checksums.csv.

The controlled HKUST T2V and first+last-frame I2V phase remains pending and is not included in these frozen counts.