README.md

metadata

license: apache-2.0
language:
  - en
task_categories:
  - text-to-video
pretty_name: VBench and VBench 2.0 Prompt Suites
configs:
  - config_name: vbench
    data_files:
      - split: test
        path: data/vbench/test-*.parquet
  - config_name: vbench2
    data_files:
      - split: test
        path: data/vbench2/test-*.parquet
  - config_name: vbench2_camera_motion
    data_files:
      - split: test
        path: data/vbench2_camera_motion/test-*.parquet
  - config_name: vbench2_complex_landscape
    data_files:
      - split: test
        path: data/vbench2_complex_landscape/test-*.parquet
  - config_name: vbench2_complex_plot
    data_files:
      - split: test
        path: data/vbench2_complex_plot/test-*.parquet
  - config_name: vbench2_composition
    data_files:
      - split: test
        path: data/vbench2_composition/test-*.parquet
  - config_name: vbench2_diversity
    data_files:
      - split: test
        path: data/vbench2_diversity/test-*.parquet
  - config_name: vbench2_dynamic_attribute
    data_files:
      - split: test
        path: data/vbench2_dynamic_attribute/test-*.parquet
  - config_name: vbench2_dynamic_spatial_relationship
    data_files:
      - split: test
        path: data/vbench2_dynamic_spatial_relationship/test-*.parquet
  - config_name: vbench2_human_anatomy
    data_files:
      - split: test
        path: data/vbench2_human_anatomy/test-*.parquet
  - config_name: vbench2_human_clothes
    data_files:
      - split: test
        path: data/vbench2_human_clothes/test-*.parquet
  - config_name: vbench2_human_identity
    data_files:
      - split: test
        path: data/vbench2_human_identity/test-*.parquet
  - config_name: vbench2_human_interaction
    data_files:
      - split: test
        path: data/vbench2_human_interaction/test-*.parquet
  - config_name: vbench2_instance_preservation
    data_files:
      - split: test
        path: data/vbench2_instance_preservation/test-*.parquet
  - config_name: vbench2_material
    data_files:
      - split: test
        path: data/vbench2_material/test-*.parquet
  - config_name: vbench2_mechanics
    data_files:
      - split: test
        path: data/vbench2_mechanics/test-*.parquet
  - config_name: vbench2_motion_order_understanding
    data_files:
      - split: test
        path: data/vbench2_motion_order_understanding/test-*.parquet
  - config_name: vbench2_motion_rationality
    data_files:
      - split: test
        path: data/vbench2_motion_rationality/test-*.parquet
  - config_name: vbench2_multi-view_consistency
    data_files:
      - split: test
        path: data/vbench2_multi-view_consistency/test-*.parquet
  - config_name: vbench2_thermotics
    data_files:
      - split: test
        path: data/vbench2_thermotics/test-*.parquet
  - config_name: vbench_aesthetic_quality
    data_files:
      - split: test
        path: data/vbench_aesthetic_quality/test-*.parquet
  - config_name: vbench_appearance_style
    data_files:
      - split: test
        path: data/vbench_appearance_style/test-*.parquet
  - config_name: vbench_background_consistency
    data_files:
      - split: test
        path: data/vbench_background_consistency/test-*.parquet
  - config_name: vbench_color
    data_files:
      - split: test
        path: data/vbench_color/test-*.parquet
  - config_name: vbench_dynamic_degree
    data_files:
      - split: test
        path: data/vbench_dynamic_degree/test-*.parquet
  - config_name: vbench_human_action
    data_files:
      - split: test
        path: data/vbench_human_action/test-*.parquet
  - config_name: vbench_imaging_quality
    data_files:
      - split: test
        path: data/vbench_imaging_quality/test-*.parquet
  - config_name: vbench_motion_smoothness
    data_files:
      - split: test
        path: data/vbench_motion_smoothness/test-*.parquet
  - config_name: vbench_multiple_objects
    data_files:
      - split: test
        path: data/vbench_multiple_objects/test-*.parquet
  - config_name: vbench_object_class
    data_files:
      - split: test
        path: data/vbench_object_class/test-*.parquet
  - config_name: vbench_overall_consistency
    data_files:
      - split: test
        path: data/vbench_overall_consistency/test-*.parquet
  - config_name: vbench_scene
    data_files:
      - split: test
        path: data/vbench_scene/test-*.parquet
  - config_name: vbench_spatial_relationship
    data_files:
      - split: test
        path: data/vbench_spatial_relationship/test-*.parquet
  - config_name: vbench_subject_consistency
    data_files:
      - split: test
        path: data/vbench_subject_consistency/test-*.parquet
  - config_name: vbench_temporal_flickering
    data_files:
      - split: test
        path: data/vbench_temporal_flickering/test-*.parquet
  - config_name: vbench_temporal_style
    data_files:
      - split: test
        path: data/vbench_temporal_style/test-*.parquet

VBench prompt suites for lmms-eval

This is a Hugging Face-native conversion of the official VBench and VBench 2.0 prompt registries. The source is pinned to 1ee42dada7a2f7cfaf4290e8a02d087f6f8ee425, the VBench revision immediately preceding the official Wan2.2 sample release.

Each row represents one video-generation request. Rows are expanded according to the official protocol: VBench uses 5 samples per prompt and 25 for temporal_flickering; VBench 2.0 uses 3 samples per prompt and 20 for Diversity. Exact duplicate prompt strings are merged in the combined vbench and vbench2 configs so a generated video can be reused across dimensions.

The seed column is a deterministic lmms-eval seed derived from suite, prompt, and sample index. VBench does not prescribe exact seed values, and the official Wan2.2 seeds were not published; these seeds provide reproducibility but do not claim bit-for-bit equivalence with official samples.

Config sizes

Config	Rows
vbench	6,220
vbench2	3,209
vbench2_camera_motion	324
vbench2_complex_landscape	90
vbench2_complex_plot	180
vbench2_composition	159
vbench2_diversity	200
vbench2_dynamic_attribute	273
vbench2_dynamic_spatial_relationship	207
vbench2_human_anatomy	360
vbench2_human_clothes	225
vbench2_human_identity	138
vbench2_human_interaction	300
vbench2_instance_preservation	171
vbench2_material	150
vbench2_mechanics	171
vbench2_motion_order_understanding	297
vbench2_motion_rationality	174
vbench2_multi-view_consistency	291
vbench2_thermotics	150
vbench_aesthetic_quality	465
vbench_appearance_style	450
vbench_background_consistency	430
vbench_color	425
vbench_dynamic_degree	360
vbench_human_action	500
vbench_imaging_quality	465
vbench_motion_smoothness	360
vbench_multiple_objects	410
vbench_object_class	395
vbench_overall_consistency	465
vbench_scene	430
vbench_spatial_relationship	420
vbench_subject_consistency	360
vbench_temporal_flickering	1,875
vbench_temporal_style	500

Columns

• prompt: text passed to the video generator.
• dimensions / official_dimensions: normalized lmms-eval names and upstream names.
• sample_index, num_samples, seed: sampling metadata.
• auxiliary_info: JSON-encoded upstream metadata used by VBench scorers.
• source_record_ids: indices into the pinned upstream full-info registry.

The unmodified source registries and the upstream Apache-2.0 license are included under source/ and LICENSE.