xju-arlab/counterfactual-vbench

VBench-CF: counterfactual probes for VBench 1.0

VBench-CF is a metamorphic test set for VBench 1.0 metrics. Rather than arguing about which of two videos is better, each clip comes from a deterministic transformation whose effect on the score is known in advance. If a metric violates that relation, the violation is a property of the metric, not of anybody's taste.

Every family contains an unmodified control plus one or more transformed levels derived from it. The transformation is applied to the video; the prompt and the dimension contract stay fixed.

Families

Dimension	Transformation	Expected relation	Clips
Spatial Relationship	Mirror the axis named by the ordered relation, prompt unchanged.	original > flip.	80
Scene	2x2 grid mixing a wrong-scene donor with target-scene quadrants at 0-100%.	Monotone increasing in target coverage.	125
Human Action	Byte-identical copies named with a correct, wrong and neutral action.	Invariance: only the filename changes.	75
Dynamic Degree	Resample one trajectory to 8/6/4/2 fps with duration held fixed.	Invariance: every rung should score the same.	160
Motion Smoothness	Duplicate, skip and locally reverse short segments; frame count and rate fixed.	Monotone decreasing in perturbation severity.	125
Subject Consistency	One fixed subject-region corruption placed at the start, middle and end.	clean > corrupted, and the three positions should tie.	100
Multiple Objects	Alpha-blend the tracked weaker target towards its local mean at 0-100%.	Monotone decreasing as the weak target disappears.	150

Overall Consistency (condition substitution) is deliberately absent: it requires human-authored prompt conditions and cannot be generated from metadata alone.

Layout

manifest.jsonl                     one row per derived clip, full provenance
metadata/bases.jsonl               the selected bases and why they qualify
metadata/verification.json         integrity, property and replay results
metadata/build_summary.json        counts, code SHA, rejected bases
<dimension>/source/<base_id>.*     the unmodified source clip
<dimension>/interventions/<family>/<base_id>__<level>.mp4

manifest.jsonl fields include dimension, family, base_id, derived_id, level, expected_relation (-1/0/+1 against the family reference), expected_rank, split, prompt_en, generator, input_sha256, output_sha256, fps, frame_count, duration_s, transformation_parameters and code_sha.

Reading the levels

expected_relation is stated against the family's reference level:

• directional_flip: reference original, levels horizontal_flip, original, vertical_flip
• environment_coverage: reference coverage_100, levels coverage_000, coverage_025, coverage_050, coverage_075, coverage_100
• filename_invariance: reference filename_correct, levels filename_correct, filename_neutral, filename_wrong
• fps_resampling: reference fps8, levels fps2, fps4, fps6, fps8
• temporal_jerk: reference jerk_0_original, levels jerk_0_original, jerk_1_duplicate, jerk_2_duplicate_skip, jerk_3_local_reverse, jerk_4_multiple
• temporal_relocation: reference clean, levels clean, corrupt_end, corrupt_middle, corrupt_start
• weakest_object_visibility: reference occlusion_000, levels conjunction_control, occlusion_000, occlusion_025, occlusion_050, occlusion_075, occlusion_100

Construction guarantees

• Deterministic. No random sampling and no generated frames. Level 0 of every family is re-encoded with the same encoder settings as its siblings, so the only difference between levels is the transformation itself.
• Replayable. Every clip is re-derived from its recorded parameters and byte-compared during verification; the detector box is stored as a parameter so verification needs no GPU.
• Score-independent. Bases are chosen from the frozen prompt split by stable hash. No Official or Audit score is consulted at any point, so the set cannot be biased towards a desired outcome.
• Split-safe. The frozen prompt split is reused and is prompt-disjoint, so a prompt never appears in both dev and test.

Counts

• bases: 205
• derived clips: 815
• built from code SHA 66c4a99dba05aceaebe80276ffbffc607c3d2e40

Dimension	dev clips	test clips
dynamics_degree	40	120
human_action	15	60
motion_smoothness	25	100
multiplt_object	30	120
scene	25	100
spatial_relationship	20	60
subject_consistency	20	80

Source and licence

Derived from the public VBench 1.0 human-preference video package (Vchitect/VBench, sampled_videos, upstream commit fd18b3d055cb0fc6f066ca90fe2c3c8cbb698490). Original clips are redistributed unaltered under <dimension>/source/ so that each counterfactual can be compared against its own base.

The upstream release terms govern reuse of these clips and of anything derived from them. This dataset is a research artefact for auditing metric behaviour; it is not a human-preference benchmark and carries no new human labels.