A Plausible Future Is Not a Causal Future: Can Video Models Generate How the World Unfolds?

Jialiang Li; Zhengning Xu; Zongyang Ma; Shaobo Cui

cs.CVOct 11, 2026

A Plausible Future Is Not a Causal Future: Can Video Models Generate How the World Unfolds?

Authors: Jialiang Li, Zhengning Xu, Zongyang Ma, Shaobo Cui

Organizations: DeepDelta Lab, School of Artificial Intelligence, Shanghai Jiao Tong University · Xi’an Jiaotong-Liverpool University

Abstract

Large-scale video generation models are increasingly viewed as a promising path toward world models, yet their ability to capture causally coherent physical dynamics remains poorly understood. Existing evaluations largely emphasize visual quality, prompt adherence, or broad physical plausibility, leaving open whether a model can faithfully realize a complete causal cascade, in which an initiating event propagates through causally dependent interactions to a terminal outcome. We introduce CausalCascade, a simulator-grounded benchmark centered on 50 event-rich Blender scenes with structured event graphs, graph-grounded diagnostic VQA, three levels of future-event specification, and 2,832 controlled counterfactual variants, and use it to evaluate nine image-to-video models with complementary motion-, event-, and counterfactual-level diagnostics. Across the evaluated models, generated motion is more reliably localized spatially than realized with correct spatiotemporal evolution, while richer future-event specification improves event recovery but leaves physical realization limited. Counterfactual evaluation further reveals weak selective responsiveness to physical interventions, while most audited failures are foundational, arising at scene consistency or event occurrence. Together, these results position causal cascade generation as a process-level test of world modeling and highlight the need to jointly recover the intended future, faithfully realize its physical evolution, selectively respond to physical interventions, and maintain a coherent physical world over time. Code and dataset are available at https://github.com/SJTU-DeepDeltaLab/CausalCascade .

Figures & tables

Figure 1: Overview of CausalCascade. (A) Base scene construction and controlled object-centric interventions. (B) Simulator-grounded causal cascade generation with factual and counterfactual rollouts. (C) Complementary motion-, event-, and counterfactual evaluations.

Benchmark	Video Generator Eval.	Reference Rollout	Controlled Intervention	Paired CF Eval.	Localized CF Effect	Ref.-based Motion Metric	Structured Event Graph	Event-level Diagnostics
WorldBench ( Upadhyay et al., 2026 )	✓	✓	✗	✗	✗	✓	✗	✗
CLEVRER-Humans ( Mao et al., 2022 )	✗	✗	✗	✗	✗	✗	✓	✓
CoPhy ( Baradel et al., 2020 )	✗	✓	✓	✓	✗	$\circ$	✗	✗
CRONOS ( Begiristain et al., 2026 )	✓	✓	✓	$\circ$	$\circ$	✓	✗	$\circ$
What-If World ( Cai et al., 2026 )	✓	✗	✓	✓	$\circ$	✗	✗	$\circ$
CausalCascade (Ours)	✓	✓	✓	✓	✓	✓	✓	✓

Table 1: Comparison with representative physical, causal, and counterfactual benchmarks. ✓,

\circ

, and ✗ denote explicit, partial, and absent coverage, respectively.

Figure 2: Benchmark statistics. Left: Distributions of structural statistics across 50 base scenes. Right: Distribution of intervention types across 2,832 counterfactual variants.

Figure 3: Main results for RQ1. Left: Performance of current I2V models under the Level 1 prompt. Right: Model-average performance as future-event specification becomes more explicit.

Figure 4: Wan2.2-I2V-A14B rollouts for one scene under three future-event specification levels. Numbers in parentheses denote VQA Overall scores. Richer future-event specification improves event-structure recovery without comparable gains in faithful physical realization.

Figure 5: Temporal alignment diagnostics. Top: Model-average agreement under strict timing, global affine alignment, and constrained DTW. Bottom: Representative alignment paths over the pairwise motion-IoU matrix.

Figure 6: Counterfactual motion response across intervention types. Each marker denotes one I2V model. Higher values on both axes indicate stronger selective causal responsiveness.

Figure 7: Earliest primary failure distribution. Angles show earliest-failure shares, while radial extents show mean dimension-level VQA scores across the audited subset. Dimensions are grouped into foundational and process failures; the right panel reports exact values.

Figure 8: Representative manifestations of foundational and process failures. Examples of cascade initiation and entity consistency illustrate foundational failures, while examples of causal propagation and temporal progression illustrate process failures.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Construction pipeline of each scene.

Diagnostic dimension	# Questions
Event occurrence	152
Temporal ordering	179
Causal pathway	226
Outcome and termination	115
Scene consistency	200
Total	872

Appendix

Table 2: Distribution of the diagnostic questions.

Tool	Inputs and sampling	Evidence supplied to the judge
observe	Generated or reference video	A $5\!\times\!4$ contact sheet of 20 uniformly spaced frames, including the first and final frames. Cells are rendered at 336 pixels before the sheet is capped at 1680 pixels, and each cell carries its frame index and physical timestamp.
zoom	Video, interval, optional box	A high-resolution crop of up to 12 requested times; without explicit times, approximately ten samples cover the interval. The box accepts normalized or pixel coordinates.
temporal_pair	Video, exact times, optional box	A two-column before/after comparison from one rollout, used for subtle changes in identity, support, orientation, and terminal state.
paired_zoom	Interval or exact times	Generated and reference crops at the same nominal times, arranged in two rows; eight times are used by default. It identifies the intended object or endpoint but does not compute a verdict.
flow_info	Generated or reference video	A global motion timeline from OpenCV DIS optical flow (medium preset) at $0.5\times$ spatial scale, including the 95th-percentile magnitude and candidate active windows.
segments	Generated or reference video	Active intervals whose flow exceeds the within-video 0.30 quantile for at least 0.25 seconds, used to select focused inspection windows.

Appendix

Table 3: Tools exposed to the semantic judge.

Subdimension	Fixed violation signals	Exact-option projection
Entity persistence	Persistent multi-component mask, reliable sparse-re-detection birth, or terminal non-boundary death.	Neither signal — consistent; death only — unexplained disappearance; birth only — unexplained duplication; both — combined.
Identity and geometry integrity	A mask is split in at least two sampled frames, or a propagated identity is lost away from the boundary.	No signal maps to the first (preserved) option; any violation maps to the designated integrity-violation option.
Trajectory continuity	At least one compensated track satisfies the state-revisit test, or one semantic object has simultaneous multi-component paths.	No signal maps to spatial continuity; any violation maps to the multiple-discontinuity option. Jump and reversal statistics remain available for diagnosis.
Static scene/camera stability	Broad-motion signal $[\operatorname{median}(\mathrm{flow})>0.10]$ and anchored-region signal $[Q_{0.95}(\Delta_{\mathrm{centre}})>2.0]$ .	Neither — stable; anchored only — structure change; broad motion only — viewpoint change; both — combined.

Appendix

Table 4: Frozen decision rules and answer-option projection for scene consistency.

Diagnostic dimension	Agreement	Accuracy
Event occurrence	17/20	85%
Causal pathway	19/20	95%
Temporal ordering	16/20	80%
Outcome and termination	20/20	100%
Scene consistency	17/20	85%
Overall	89/100	89%

Appendix

Table 5: Agreement with independent human annotations on a 100-question audit set.

Figure 13: Interface used for manual screening of edited rollouts. The base and edited rollouts are shown together for comparison, and synchronized playback is supported. Annotators assign each candidate to High quality , Full dataset , or Discard .

Model	Checkpoint	Resolution	Frames	FPS	Steps	Guidance / schedule
CogVideoX-Fun-V1.5	CogVideoX-Fun-V1.5-5b-InP	$1360{\times}768$	85	8	50	guidance 6.0
Cosmos3-Nano	Cosmos3-Nano	$832{\times}480$	189	24	35	shift 10.0; guidance 6.0
HunyuanVideo-I2V	HunyuanVideo-I2V	$960{\times}544$	129	24	50	shift 7.0; embedded CFG 6.0
LTX-Video-13B	LTX-Video-13B	$832{\times}480$	121	16	$30,30^{\dagger}$	scheduled CFG/STG †
MAGI-1-4.5B	MAGI-1-4.5B-base	$832{\times}480$	120	24	64	multi-stage text/prev.-chunk guidance
SkyReels-V2-I2V	SkyReels-V2-I2V-14B-720P	$1280{\times}720$	121	24	50	shift 3.0; guidance 5.0

Appendix

Table 6: Inference configurations of the models under test in our experiments. We retain each model’s released generation configuration rather than imposing a common resolution, frame rate, or sampling schedule. Guidance parameters follow their model-specific definitions and are therefore not necessarily directly comparable across architectures.

Model	Spatial IoU $\uparrow$	Strict ST-IoU $\uparrow$	Weighted Spatial IoU $\uparrow$	mMSE $\downarrow$	SSIM $\uparrow$	VQA Overall $\uparrow$
CogVideoX-Fun-V1.5	0.147	0.037	0.067	0.0539	0.771	0.095
Cosmos3-Nano	0.252	0.043	0.137	0.0579	0.903	0.357
HunyuanVideo-I2V	0.178	0.027	0.085	0.0645	0.608	0.174
LTX-Video-13B	0.182	0.037	0.087	0.0583	0.727	0.189
MAGI-1-4.5B	0.128	0.031	0.072	0.0673	0.697	0.153
SkyReels-V2-I2V	0.193	0.047	0.105	0.0597	0.738	0.177

Appendix

Table 7: Complete performance of current I2V models under the Level 1 prompt and strict evaluation protocol. Best results are shown in bold and second-best results are underlined.

	Spatial IoU $\uparrow$			Strict ST-IoU $\uparrow$			VQA Overall $\uparrow$
Model	L1	L2	L3	L1	L2	L3	L1	L2	L3
CogVideoX-Fun-V1.5	0.147	0.174	0.191	0.037	0.042	0.062	0.095	0.128	0.186
Cosmos3-Nano	0.252	0.273	0.387	0.043	0.061	0.080	0.357	0.358	0.568
HunyuanVideo-I2V	0.178	0.202	0.172	0.027	0.035	0.034	0.174	0.171	0.167
LTX-Video-13B	0.182	0.180	0.217	0.037	0.042	0.057	0.189	0.213	0.227
MAGI-1-4.5B	0.128	0.142	0.154	0.031	0.038	0.045	0.153	0.171	0.229

Appendix

Table 8: Complete effect of future-event specification on causal cascade generation. We report Spatial IoU, Strict ST-IoU, and VQA Overall under the three specification levels. The best level for each model and metric is shown in bold.

Model	CF Spatial Effect IoU $\uparrow$	CF Weighted Effect IoU $\uparrow$	CF Preserve Spatial IoU $\uparrow$	CF Preserve Weighted IoU $\uparrow$
CogVideoX-Fun-V1.5	0.0261	0.0461	0.1190	0.0507
Cosmos3-Nano	0.0410	0.0488	0.1983	0.1223
HunyuanVideo-I2V	0.0365	0.0389	0.1577	0.0895
LTX-Video-13B	0.0324	0.0362	0.1471	0.0800
MAGI-1-4.5B	0.0353	0.0372	0.1067	0.0829
SkyReels-V2-I2V	0.0321	0.0468	0.1687	0.0946

Appendix

Table 10: Complete counterfactual results for delete interventions. All values are scene-macro averages. Higher is better for all metrics. Best results are shown in bold and second-best results are underlined.

Model	CF Spatial Effect IoU $\uparrow$	CF Weighted Effect IoU $\uparrow$	CF Preserve Spatial IoU $\uparrow$	CF Preserve Weighted IoU $\uparrow$
CogVideoX-Fun-V1.5	0.0278	0.0349	0.1235	0.0537
Cosmos3-Nano	0.0451	0.0576	0.1879	0.1360
HunyuanVideo-I2V	0.0367	0.0385	0.1720	0.0979
LTX-Video-13B	0.0457	0.0400	0.1274	0.0704
MAGI-1-4.5B	0.0231	0.0283	0.1096	0.0665
SkyReels-V2-I2V	0.0243	0.0389	0.1729	0.0935

Appendix

Table 11: Complete counterfactual results for move interventions. All values are scene-macro averages. Higher is better for all metrics. Best results are shown in bold and second-best results are underlined.

Model	CF Spatial Effect IoU $\uparrow$	CF Weighted Effect IoU $\uparrow$	CF Preserve Spatial IoU $\uparrow$	CF Preserve Weighted IoU $\uparrow$
CogVideoX-Fun-V1.5	0.0180	0.0278	0.1136	0.0547
Cosmos3-Nano	0.0482	0.0525	0.1879	0.1215
HunyuanVideo-I2V	0.0256	0.0264	0.1535	0.0749
LTX-Video-13B	0.0333	0.0367	0.1654	0.0860
MAGI-1-4.5B	0.0193	0.0263	0.1024	0.0566
SkyReels-V2-I2V	0.0308	0.0357	0.1661	0.0856

Appendix

Table 12: Complete counterfactual results for scale interventions. All values are scene-macro averages. Higher is better for all metrics. Best results are shown in bold and second-best results are underlined.

Model	CF Spatial Effect IoU $\uparrow$	CF Weighted Effect IoU $\uparrow$	CF Preserve Spatial IoU $\uparrow$	CF Preserve Weighted IoU $\uparrow$
CogVideoX-Fun-V1.5	0.0229	0.0344	0.1182	0.0494
Cosmos3-Nano	0.0341	0.0397	0.1847	0.1099
HunyuanVideo-I2V	0.0360	0.0387	0.1596	0.0950
LTX-Video-13B	0.0108	0.0222	0.1502	0.0855
MAGI-1-4.5B	0.0250	0.0331	0.1020	0.0636
SkyReels-V2-I2V	0.0278	0.0425	0.1745	0.0922

Appendix

Table 13: Complete counterfactual results for mass interventions. All values are scene-macro averages. Higher is better for all metrics. Best results are shown in bold and second-best results are underlined.

Panel of Figure 14 — Figure 14: Representative cases of intervention neglect despite active dynamics. Top: For Wan2.2-I2V-A14B, changing the barrel mass substantially changes the reference trajectory, while the generated factual and counterfactual rollouts remain dynamic but evolve similarly. Bottom: For Wan2.2-I2V-A14B, moving a critical domino interrupts the reference counterfactual cascade before the fork, but the generated cascade continues through downstream branches in both rollouts. Reference and generated factual-counterfactual rollouts are shown at matched normalized times.

Panel of Figure 15 — Figure 15: Generated differences that do not realize the reference intervention consequence. Top: For Cosmos3-Nano, the factual and counterfactual generations differ, but their difference is dominated by identity and geometry drift rather than the required interruption of the domino pathway. Bottom: For Wan2.2-I2V-A14B, the scale intervention produces a localized generated response, yet the counterfactual rollout does not reproduce the reference support-loss mechanism or collapsed terminal state. These examples show that pairwise change alone does not establish a correct causal response.

Panel of Figure 16 — Figure 16: Representative failures to preserve intervention-unaffected dynamics. Top: Shortening one support should locally alter the platform support relation, but the generated pair exhibits large changes in platform scale and geometry that extend beyond the intervention-affected dynamics. Bottom: Deleting one support should primarily alter the platform response, but the generated pair introduces numerous additional balls whose motion is unrelated to the intervention and disrupts dynamics that should remain shared. Together, the cases show that differences between generated factual and counterfactual rollouts are not reliably confined to the consequences of the controlled intervention.

Panel of Figure 17 — Figure 17: Motion-map diagnostics for representative counterfactual responses. Left: For the scale intervention, the generated maps retain partial spatial correspondence with the reference response but introduce substantially more diffuse preserved, suppressed, and induced motion; the correct support-loss process is not recovered. Right: For the delete intervention, the generated pair introduces broad preserved and suppressed activity while recovering little of the concentrated induced-motion pattern present in the reference pair. Columns show preserved, intervention-suppressed, and intervention-induced motion, while rows compare spatial and weighted maps for the reference and generated factual-counterfactual pairs.

Failure manifestation	All cases	Level 1	Level 2	Level 3
Spontaneous motion	51/90 (56.7%)	20/36 (55.6%)	16/27 (59.3%)	15/27 (55.6%)
Incorrect physical response	47/90 (52.2%)	20/36 (55.6%)	17/27 (63.0%)	10/27 (37.0%)
Identity or attribute corruption	26/90 (28.9%)	10/36 (27.8%)	8/27 (29.6%)	8/27 (29.6%)
Motion recurrence or state replay	23/90 (25.6%)	11/36 (30.6%)	6/27 (22.2%)	6/27 (22.2%)
Topology corruption	21/90 (23.3%)	8/36 (22.2%)	5/27 (18.5%)	8/27 (29.6%)
Structural-role misinterpretation	15/90 (16.7%)	7/36 (19.4%)	4/27 (14.8%)	4/27 (14.8%)

Appendix

Table 14: Multilabel observable failures coded from the complete qualitative descriptions.

Model	Topology	Unsupported motion	Identity error	Physical response	Duplication	Structural role	Viewpoint drift	State replay
CogVideoX-Fun-V1.5	10	6	3	–	–	–	–	–
Cosmos3-Nano	–	5	–	7	5	–	–	–
HunyuanVideo-I2V	–	9	4	4	–	–	–	–
LTX-Video-13B	–	3	–	8	–	5	–	–
MAGI-1-4.5B	–	–	5	3	–	–	5	–
SkyReels-V2-I2V	3	9	–	5	–	–	–	–

Appendix

Table 15: Model-specific failure profiles on each model’s three most frequent error manifestations.

Earliest primary failure	Cases	Proportion
Event occurrence	43	47.8%
Causal pathway	8	8.9%
Temporal ordering	4	4.4%
Outcome and termination	2	2.2%
Scene consistency	33	36.7%

Appendix

Table 16: Location of the earliest primary failure. Each rollout receives exactly one label.

Causal Video Generation

See all 21 papers

CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models
May 22, 2026León Begiristain, Olaf Dünkel, Adam KortylewskiCounterfactual Evaluation Physical Consistency in Video Generation
University of Freiburg Freiburg im Breisgau, Germany · Max Planck Institute for Informatics Saarland Informatics Campus, Germany · CISPA Helmholtz Center for Information Security Saarbrücken, Germany
Foresight at the Event Boundary: Evaluating Physical Prediction in Video World Models
Sep 29, 2026Estela Monserrat Arriaga Santana, Julian Rosas Scull, Ehécatl Sacamch'en Núñez Rico +1Physical Consistency in Video Generation World Model Evaluation
National Autonomous University of Mexico · National Autonomous University of Mexico, Mexico · University of Texas at El Paso +1
Chain of Event-Centric Causal Thought for Physically Plausible Video Generation
Mar 10, 2026Zixuan Wang, Yixin Hu, Haolan Wang +4Causal Video Generation Video Generation
Sichuan University · The University of Adelaide · Hong Kong Polytechnic University +1

May 22, 2026J/K move · Enter open · S save

CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models

León Begiristain, Olaf Dünkel, Adam Kortylewski

University of Freiburg Freiburg im Breisgau, Germany · Max Planck Institute for Informatics Saarland Informatics Campus, Germany · CISPA Helmholtz Center for Information Security Saarbrücken, Germany

Counterfactual Evaluation Physical Consistency in Video Generation Video Prediction Causal Reasoning in Videos Video Generation Evaluation

Open paper AskPDF arXiv

Video prediction is increasingly viewed as a path toward generalizable world models, yet it remains unclear whether these systems learn underlying causal structure or merely exploit superficial visual correlations for future prediction. We introduce CRONOS, an intervention-based benchmark to evaluate counterfactual physical consistency: whether a model's predictions of physical events respond appropriately to controlled changes in the visual input, such as variations of camera viewpoint, scene, object category, and object appearance. Built in a photorealistic Unreal Engine environment, CRONOS enables controlled, high-fidelity generation of videos across diverse scenes and dynamics, including sampling multiple physically plausible outcomes of the same physical event. In contrast to previous benchmarks, that evaluate video prediction on scenes CRONOS systematically intervenes on four factors -camera viewpoint, scene, object category, and object appearance- while keeping the underlying physical event type, such as a fall, collision, or occlusion, fixed. Our evaluation of recent open-source video generators reveals substantial failures in counterfactual physical consistency: prediction quality for the same physical event type is affected by appearance, environment, and particularly viewpoint changes. CRONOS provides a controlled and reproducible testbed for diagnosing how generated video quality changes across interventions, establishing a concrete target for developing models that perform consistently across changes of multiple conditions. The dataset and evaluation code are available at our project page.

Abstract

Causal Video Generation

Explore similar work

CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models

Foresight at the Event Boundary: Evaluating Physical Prediction in Video World Models

Chain of Event-Centric Causal Thought for Physically Plausible Video Generation