CVGBench v1.0

CVGBench is a real-video benchmark for counterfactual video generation. Each item pairs a source video with a minimal edit request and the chain of visible consequences that the edit should cause. Multiple-choice questions then check whether a generated video shows those consequences.

This is the release for the paper A Real-Video Benchmark for Counterfactual Video Generation with Causal-Chain Diagnostics (WACV 2027).

Contents

• 1,000 items, with IDs 0001 to 1000.
• 7,400 multiple-choice questions, each with four options (A–D) and one correct answer.
• One causal chain per item, of depth 2 to 6 (200 items at each depth).

Depth	Item IDs	Questions per item
2	0001–0200	7
3	0201–0400	7
4	0401–0600	7
5	0601–0800	8
6	0801–1000	8

Each item's edit belongs to one of 9 families, and each family has 2 labels (55 or 56 items per label):

Family	Labels
collision	drive_object_into_obstacle_or_support, redirect_swung_object_into_contact
deformation_elasticity	overbend_object_past_elastic_limit, press_soft_material_into_deeper_deformation
material_flow_ejection	create_open_path_for_material_escape, loosen_closure_before_contents_release
motion_friction	reduce_surface_grip_during_motion, swap_contact_surface_for_new_friction
rotation	flip_container_to_new_orientation, turn_opening_or_face_away
spatial_occlusion	cover_visible_item_with_occluder, move_item_behind_visible_occluder
stability_balance	prevent_destabilizing_action_keep_stack_upright, shift_support_or_center_of_mass_for_balance
state_change	change_material_or_closure_state_before_action, toggle_ready_or_locked_state_before_motion
trajectory_incline	redirect_motion_onto_sloped_route, tilt_support_into_incline_path

Files

CVGBench-v1.0/
  README.md
  LICENSE
  items/<ID>.json          edit request and questions
  answer_keys/<ID>.json    correct option for each question
  causal_graphs/<ID>.json  causal chain

The three files of an item share the same <ID> and item_id. All files are UTF-8 JSON.

items/<ID>.json

Field	Meaning
schema_version	"CVGBench item v1.0"
item_id	Item ID, e.g. "0001"
source_clip_id	ID of the source video in Something-Something V2 (SSv2)
causal_depth	Number of events in the causal chain (2–6)
intervention	{"family": ..., "label": ...} (see the table above)
edit_prompt	The edit request given to the video generator
questions	List of questions: question_id (Q1–Q8), text, options (id A–D and text), and question_uid (e.g. "0001.Q1")

answer_keys/<ID>.json

Field	Meaning
schema_version	"CVGBench answer key v1.0"
item_id	Item ID
answers	One entry per question: question_id and correct_option_id (A–D)

causal_graphs/<ID>.json

Field	Meaning
schema_version	"CVGBench causal graph v1.0"
item_id, source_clip_id, causal_depth, intervention, edit_prompt	Same as in the item file
causal_graph.nodes	The events in order: id (N1 … Nd) and description
causal_graph.edges	One edge between each pair of consecutive events: source, target and relation (causes, leads_to, enables or requires)

Each chain is a single path, N1 → N2 → … → Nd, with no branches. causal_depth equals the number of events.

Questions

Question numbers are fixed roles:

Question	Checks
Q1	The main outcome of the edit
Q2	A later consequence
Q3	A consequence deep in the chain (depth 5–6 items only)
Q4	Local physical realism
Q5	The order of events
Q6	Whole-event visual coherence
Q7	Temporal coherence
Q8	Preservation of content the edit should not change

Depth 2–4 items have no Q3, so their questions are Q1, Q2 and Q4–Q8. This is intended, not missing data.

Videos

This release contains no video files. source_clip_id is the video ID in Something-Something V2 (SSv2). Download the source videos from the official SSv2 release, under its own terms. Generated videos are not included.

Evaluation

1. The video generator under test receives the source video and edit_prompt, and produces a generated video.
2. An evaluator watches the source video and the generated video, and answers each question from what the generated video shows. The evaluator receives the edit prompt, the causal chain and the questions, but not the answer key or the identity of the model.
3. An answer is correct when it matches correct_option_id. An unanswered question counts as incorrect.

Scores:

• Criterion pass rate: the share of questions answered correctly.
• All-pass: an output passes only when every one of its questions is answered correctly.

The paper reports human-corrected all-pass as its primary metric, and describes the full protocol.

Citation

@inproceedings{chou2027cvgbench,
  title     = {A Real-Video Benchmark for Counterfactual Video Generation with Causal-Chain Diagnostics},
  author    = {Chou, Shawn and Shen, Yiqing and Unberath, Mathias},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}

License

The annotations (edit prompts, questions, options, answer keys and causal graphs) are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0); see LICENSE. The SSv2 source videos are not included and remain under their own terms.

Contact

Shawn Chou, Yiqing Shen and Mathias Unberath, Johns Hopkins University: yzhou218@jhu.edu, yshen92@jhu.edu, unberath@jhu.edu

Version

v1.0 (2026): first release.