xju-arlab/vbench-repair

VBench Repair

A dimension-first archive of VBench 1.0 original sampled videos, the 16 original human-preference annotation files, and versioned counterfactual experiments.

Publication has been verified against the pinned data revision in provenance/dataset-verification.json. The archive contains 27,720 dimension-local original entries (19,400 distinct source paths), 6,930 annotation rows, 41,580 normalized unordered comparisons, and 27 versioned counterfactual experiment archives. The upstream reference issues below remain documented and unresolved.

Layout

dimensions/<dimension>/origin/videos/ contains byte-preserved MP4/GIF originals. human_preference/ contains unchanged official JSON, normalized unordered pairs, and explicit reference issues. counterfactual/<version>/ contains tar shards, member hashes, construction records and experiment-specific documentation. All paths in origin and pair manifests are relative to this repository root.

Source limitations

The official Drive Background_Consistency.json contains 1,710 video references that are missing or conflict with their generator keys. It is preserved unchanged. The corresponding official Scene video suite is included as a candidate source suite, but no broken human label is silently re-associated. Unresolved pairs are marked unusable and have null media paths. The other 15 annotation files match the pinned official video paths exactly. This is an archive of the released materials, not a claim that upstream annotations are free of errors.

The generated counterfactuals include historical and diagnostic experiments. Their recorded validity limitations and failed controls remain part of the data. A stored intervention does not imply that its repair method passed all tests.

Provenance and use

Videos: Vchitect/VBench_sampled_video, revision 8c89bb0218a12323d04821d11ec03c780d0dfb9c. Annotations: the official Google Drive release linked by VBench. Original source terms are retained; this reorganization does not assign a new license to third-party videos. The original annotation README is preserved.

Downloads/publication attempt https://hf-mirror.com first. The mirror returned HTTP 308 redirects/403 or rejected authenticated upload, so the publisher falls back to the official endpoint when explicitly enabled. Large originals use exact server-side LFS copies where available; Git-backed videos are downloaded, checked against the source blob, and uploaded using binary storage.

Selected trained weights are in xju-arlab/vbench-model. Dynamic publishes the current user-accepted aligned-v1 head and its frozen V-JEPA 2.1 ViT-B backbone in the model repository. This supersedes the earlier record-only decision; old joint/anchored failures remain documented.

Coverage

Dimension	Original files	Annotation rows	Pairs with exact media references	CF experiments
aesthetic_quality	1,860	465	2,790/2,790	0
appearance_style	1,800	450	2,700/2,700	0
background_consistency	1,720	430	5/2,580	1
color	1,700	425	2,550/2,550	1
dynamic_degree	1,440	360	2,160/2,160	11
human_action	2,000	500	3,000/3,000	2
imaging_quality	1,860	465	2,790/2,790	0
motion_smoothness	1,440	360	2,160/2,160	1
multiple_objects	1,640	410	2,460/2,460	2
object_class	1,580	395	2,370/2,370	1
overall_consistency	1,860	465	2,790/2,790	0
scene	1,720	430	2,580/2,580	2
spatial_relationship	2,160	540	3,240/3,240	2
subject_consistency	1,440	360	2,160/2,160	4
temporal_flickering	1,500	375	2,250/2,250	0
temporal_style	2,000	500	3,000/3,000	0

Original counts include shared source videos in separate dimension directories. Pair usability here means both media references resolve exactly; reverse-label consistency is recorded separately in reciprocal_labels. Counterfactual archive members include originals, controls, generated clips, frame arrays, previews and records, so their member counts must not be interpreted as unique derived-video counts.

Download and integrity

HF_ENDPOINT=https://hf-mirror.com hf download xju-arlab/vbench-repair \
  --repo-type dataset --include 'dimensions/scene/**' --local-dir vbench-repair

If the mirror fails, retry with HF_ENDPOINT=https://huggingface.co. Original videos are directly accessible at the repository-relative paths in origin/manifest.jsonl and human_preference/pairs.jsonl. For one counterfactual experiment, download every data-*.tar shard and extract them into one empty directory for that experiment. files.jsonl maps each source-relative member to its archive and SHA-256. Keep different experiment versions in separate extraction directories.

Verification checks every original file's remote size/hash, raw annotation byte identity, all shard hashes and index coverage, and downloaded original/complete-shard samples. Historical absolute machine paths are provenance; use the archive-relative indexes for access.

Reproduction supplements

Frozen six-dimension training inputs and compact table replay bundles are now available. See reproduction assets and the code runbook. The latest fresh restore verifies 222 metadata/evidence/source files and 79 archive members; the nine-dimension table replay matches 24,408 score cells, preserving eligibility and missing-score states. This verifies formula replay and score reaggregation; full fresh GPU inference, retraining and all appendix/ablation replays have not been certified. The original media verification above remains pinned to its recorded data revision.

Current nine-dimension paper: Dynamic aligned-v1

Dynamic Degree is the ninth repaired dimension in the current paper. The accepted model is aligned-v1, fixed step 300 (not dev-selected), tested on 450 sources, two 8px local texture-jitter seeds per source and 450 encoding controls. Origin: 0.680000 → 0.857778; Repair: 0.629251 → 0.623035. Controls are checked separately, not pooled into the CF mean. DEV17/23 and the failed original strict scale gate remain recorded; batch stability is not per-video invariance.

New training records are in dimensions/dynamic_degree/training/aligned-v1/. The compact raw-score replay bundle is dimensions/dynamic_degree/reproduction/aligned450-v1/evidence.tar.gz. Existing counterfactual/vjepa-expansion450-v1/ media are reused byte-for-byte. Both weights are in the public Dynamic model directory; full code and execution instructions remain in the private GitHub repository.

Container reset recovery

Project-specific dependency runtimes, upstream model assets and verification media are preserved in reproduction/runtime/h100-20260923. The release manifest pins eight archives and their resumable chunks. The private code provides the current nine-repair evaluator and all 16 official VBench dimensions. Representative H100 checks and the independent H200 official checks passed; further H200 nine-repair testing was stopped at the user’s request. See the linked validation record for the exact scope.

Dynamic external generalization: LASIESTA and BMC (2026-09-23)

Full reports, protocols and byte-exact numeric evidence add 43 LASIESTA and 84 BMC clips from 16 source recordings. The aligned-v1 weights and the 450-source paper experiment are unchanged. LASIESTA static/moving: Origin 0.000000→0.875000 / 0.592593→0.981481; Repair 0.125877→0.163079 / 0.589715→0.610019. BMC all: Origin 0.261905→1.000000; Repair 0.282374→0.292002. BMC labels are agent review, not official motion labels; native LASIESTA timing is unverified. These are exploratory mean-robustness results, not per-clip invariance or 127 independent scenes. Only numeric evidence is released; no original/derived external media are redistributed. Earlier archive counts above remain unchanged.