WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World
1. Paper overview
In one sentence: WorldLens exposes the gap between convincing driving imagery and functional reliability by evaluating reconstruction, control, perception and human judgment alongside generation quality. e02e04e05e14e15e20e22e26
| At a glance | What to know |
|---|---|
| Research problem | Source description A convincing driving video may still contain unstable geometry or cues that mislead a planner. WorldLens asks which aspects survive when generated imagery is reconstructed, interpreted by pretrained perception systems, or fed back into a driving simulator. These tests measure different properties and should not be collapsed into a single realism claim. e02e03e12 |
| Core mechanism | Source description The benchmark organizes 24 dimensions into five aspects, combining learned visual proxies, 4D reconstruction, simulated control, downstream perception and human ratings. e02 |
| A key reported result | Closed-loop route following: RLGF: RC 13.51%, ADS 10.59%; open-loop PDMS 78.45%. Route Completion and Arena Driving Score; both higher are better. DriveArena-aligned five-sequence simulation on singapore-onenorth and boston-seaport, with planner decisions feeding back into generation. MagicDrive-V2: RC 12.31%, ADS 9.50%, open-loop PDMS 78.91%; MagicDrive: RC 6.89%, ADS 4.82%. RLGF leads reported closed-loop scores, while MagicDrive-V2 leads open-loop PDMS. Low route fractions expose a serious simulated-control gap, but no physical deployment or uncertainty estimate is established. e11e12e13e15 |
| Reading caution | Source description Driving-only scope, style/regional annotation bias, inherited critic weaknesses and evolving notions of physical realism are acknowledged limitations. e23 |
Core contributions
- Source description
The benchmark organizes 24 dimensions into five aspects, combining learned visual proxies, 4D reconstruction, simulated control, downstream perception and human ratings. e02
Figure 3. Human supervision is collected with several synchronized representations of the same scene. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the upper-left Video panel, then compare it with the Semantic panel beside it. The lower panels show depth and 3D boxes, giving annotators additional representations when judging scene structure and traffic behavior. The task panel selects H.4, Behavioral Safety, in this example. The bottom rating field and text area indicate the two kinds of annotation collected: a numerical judgment and a rationale. Section 4.1 explains that two independent groups review the videos and revisit disagreements. The crop depicts the annotation procedure; its empty rating field does not report an actual evaluation outcome. e04e02
What it supports. WorldLens-26K contains 26,808 scoring records, each pairing a judgment with an explanation. The interface shows why these labels carry more context than a simple preference between two RGB clips. Reader analysis: the additional representations may help reviewers identify failures, but they also make the labels dependent on the quality of those representations.
Where the evidence stops. The static interface does not establish temporal behavior, annotation agreement or the number of unique clips. Its semantic colors and depth display are not accompanied by quantitative legends, so no class-color mapping or metric depth is inferred.
2. Motivation
2.1 The problem and the proposed response
A convincing driving video may still contain unstable geometry or cues that mislead a planner. WorldLens asks which aspects survive when generated imagery is reconstructed, interpreted by pretrained perception systems, or fed back into a driving simulator. These tests measure different properties and should not be collapsed into a single realism claim. e02e03e12
2.2 What this reading follows
A synthetic driving clip can look convincing while becoming difficult to reconstruct or misleading a planner. WorldLens turns that concern into a benchmark spanning five aspects and 24 dimensions. It evaluates existing generators with separate visual, geometric and functional tools, then adds human score-and-rationale annotations and a trained vision-language critic. Read the results as a collection of tests with different targets: reconstructing a scene, retaining perception cues and finishing a simulated route are distinct achievements. The original visuals below explain these distinctions, while the report preserves the supplied v2 paper's inconsistent values and separates qualitative critic examples from measured agreement. e02e04e05e14e15e20e22e26
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Not assigned |
| Architecture | Not assigned |
| Prediction paradigm | Not assigned |
| Quadrant | Not assigned |
This table preserves the labels recorded at reading time. The current major category is Benchmarks & simulators. View the current classification.
3.1 Evidence-based assessment
Classification assessment not applicable
The recorded catalog fields are all unassigned. WorldLens is a benchmark plus a rating critic, not a proposed joint future/action generator or inverse-dynamics controller. Separate pretrained planners execute simulated actions, while the critic emits judgments. A One Model/quadrant assignment would conflate evaluated systems with the evaluator. e02e05e12
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 First identify what each reconstruction metric compares
Begin with a generated multi-camera clip and fit an OmniRe 4D Gaussian field to it. Original-pose LPIPS asks whether that fitted field can reproduce the generated input. It does not directly ask whether the input matches a real scene. For geometric discrepancy, WorldLens reconstructs a paired real clip separately and compares rendered depths inside road-and-vehicle masks. The reference is consequently another fitted scene, rather than direct physical depth measurements. Novel-view tests then render both fields along matched shifted trajectories; MUSIQ judges image quality and FVD compares video distributions. Reader interpretation: separating these comparisons explains why low photometric error and low geometric error need not select the same model, and why reconstruction settings must stay fixed before making claims about generator quality. e07e08e09e14e29
5.2 Then follow one action through the simulation loop
The open-loop displacement test passes generated and real observations through the same pretrained planner and compares their predicted waypoints. Those predictions are measurements. In the closed-loop test, the planner's output instead updates the simulated ego state, and that changed state conditions the next generated observation. The agent then plans again. This creates an opportunity for errors to accumulate through observation, planning and generation. Route Completion tracks how much distance survives this process; ADS also penalizes poor driving quality through its product with PDMS. The product is computed within each episode before averaging. Reader interpretation: Table 2's low route completion motivates testing the complete feedback system, but the cross-model ranking alone cannot separate generator drift from planner or simulator limitations. e10e11e12e13e15
5.3 Finally distinguish a learned judge from a learned driver
The human annotation interface offers RGB, semantics, depth and boxes, while the critic architecture describes generated video plus rubric text as inputs. Human records supervise an attention-only LoRA adaptation of the stated Qwen3-VL-8B backbone; the visual encoder, projector and other named components remain frozen. The output is an evaluative score and rationale, not an action or predicted world state. Although the architecture illustration groups multiple criteria, the printed prompt specifies one target dimension and one score/reason object per call. Reader analysis: learning to imitate these labels may automate part of evaluation, but selected explanations on unseen videos do not quantify reliability. The low human medians make comparison with an untuned critic and a constant-score predictor particularly informative before using the critic as a reward source. e04e05e06e21e22e28
5.4 Training and inference
During training
Reconstruction is per-clip optimization: six cameras, 16 frames per camera at 12 Hz, 544×304 pixels and 30,000 OmniRe optimization steps. Initialization, losses and scheduling are delegated to OmniRe defaults rather than fully enumerated. e07
The stated critic backbone is Qwen3-VL-8B. Attention-only LoRA uses rank 16 and dropout 0.05; vision encoder, projector, embeddings and MLPs remain frozen. Supervised tuning runs three epochs at learning rate 1e-4 with cosine decay, 0.1 warmup ratio, bfloat16 and eight A100 GPUs. The backbone citation is inconsistent. e05e28
During inference
The critic projects video features into language-token space alongside rubric text, then autoregressively emits scores and reasons. The printed prompt specifies one dimension per call, with half-point ratings from 1 to 10; the architecture illustration's multiple outputs do not establish a different executable interface. e05e06
The simulator uses five sequences on two maps, 10 Hz traffic and 2 Hz control. Three reference frames condition 448×800 generation, resized to 224×400 for the agent. These frequencies describe the simulation protocol, not demonstrated real-time inference throughput. e11
5.5 Implementation flow
- Measure visual signals
Use instance classifiers and ReID features for subjects, DINO/CLIP features for consistency, depth and semantic predictions for scene stability, I3D FVD for distributional similarity, and LoFTR for cross-view correspondence. These are learned proxies whose own domain sensitivity affects interpretation. e03e24e25
- Turn videos into a reconstruction test
Fit a 4D Gaussian field to each generated clip. Re-render original poses for appearance error; separately reconstruct the paired real clip and compare masked rendered depths. Evaluate center-interpolated, S-curve and lateral-offset views using MUSIQ and FVD. Reconstructability therefore depends on the reconstruction pipeline as well as the generator. e07e08e09
- Separate prediction from executed feedback
Displacement Error compares UniAD predictions from generated and real videos over one second on 150 nuScenes validation scenes. Open-loop PDMS scores proposed behavior without updating the ego trajectory from those predictions. Closed-loop control updates the simulator state and conditions subsequent generation; RC and ADS measure the resulting episodes. e10e11e12e13
- Probe perception without retraining
Apply real-data-pretrained camera perception models directly to generated inputs: BEVFusion for map segmentation and detection, ADA-Track for tracking, and SparseOcc for occupancy. Compare with corresponding annotations. This measures how useful generated observations are to existing models, not whether synthetic-data training improves real-world deployment. e16e17e18e19
- Collect supervision for a critic
Ten annotators work in two independent groups, consulting synchronized RGB, semantic masks, depth and 3D boxes. Disagreements trigger re-evaluation. Ratings cover realism, physical plausibility, 3D/4D consistency and behavioral safety, with separate overall/vehicle/pedestrian realism rubrics. e02e04e06
6. Experiments & results
WorldLens evaluates driving video generators through appearance, reconstructability, planner behavior, perception and human judgment. Its strongest lesson is that favorable image metrics coexist with poor closed-loop route completion. A separate LoRA-trained critic learns score-and-rationale outputs from human annotations; its generalization evidence remains qualitative.
The paper supplies cross-model tables and qualitative diagnostics, but no controlled ablation isolating a generator design choice or the critic's LoRA adaptation. Figure 5 therefore occupies the diagnostic role rather than establishing a causal mechanism. Agent alignment is supported by selected examples rather than a quantitative agreement table. The illustrated method panel uses the readable annotation interface, while the critic architecture is explained from Section 12.1. e04e05e14e20e22e29
6.1 Read the original evidence
Table 1. The strongest appearance, geometry and novel-view scores belong to different models. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read this table by column, following each upward or downward arrow; the entries do not share a common scale. The turquoise block contains eight generation dimensions, including distributional FVD in G.7. The yellow block tests reconstruction: R.1 is original-pose LPIPS, R.2 is masked rendered-depth AbsRel, R.3 is novel-view MUSIQ, and R.4 is novel-view FVD. The lower reference row is labeled Empirical Max by the authors, with several missing entries. It is an empirical comparator rather than a mathematical upper bound. Also keep the percent formatting of novel-view quality distinct from a percentage of successful episodes. e03e07e08e09e14e25
What it supports. DiST-4D has the lowest generation FVD, 58.08, compared with OpenDWM's 90.42. OpenDWM narrowly leads original-view LPIPS, 0.065 versus 0.066, while DriveDreamer-2 leads geometric AbsRel at 0.073. This supports evaluating multiple properties separately; the table alone does not establish which architectural choice caused a model's advantage.
Where the evidence stops. G.8 matches appendix Table 11's VC Match row, not its Total row. The prose also says DiST-4D trails OpenDWM in view consistency, although the printed values are 389.78 versus 211.18. These source discrepancies are preserved.
Table 2. High open-loop adherence coexists with very limited closed-loop route completion. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read A.1 as a comparison between planner trajectories induced by generated and real observations. A.2 is open-loop PDMS, where predicted actions do not drive the ego trajectory. A.3 measures the fraction of route distance reached after closing the feedback loop. A.4 is ADS, computed from episode-level route completion and driving quality. Panacea's dashes indicate missing reports rather than zero performance. The appendix places the simulation on five sequences across two map environments. Do not multiply A.2 and A.3 to reconstruct A.4: open-loop PDMS is a different evaluation, and the closed-loop definition averages products within episodes. e10e11e12e13e15e26
What it supports. MagicDrive-V2 leads open-loop PDMS at 78.91%, but completes 12.31% of route distance and scores 9.50% ADS. RLGF has slightly lower open-loop PDMS, 78.45%, but reaches 13.51% RC and 10.59% ADS. The ranking change and low route fractions show why short-horizon scoring cannot replace feedback-based simulation.
Where the evidence stops. Appendix Table 16 changes three A.1 values: DreamForge 0.55 to 0.57, DrivingSphere 0.54 to 0.55, and MagicDrive-V2 0.53 to 0.54. The unexplained 0.51 empirical reference also complicates the displacement definition. No uncertainty or real-vehicle result is supplied.
Table 3. Generated observations preserve some perception cues better than others. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each column tests a pretrained perception system on generated observations against the corresponding task annotations. Map Seg. reports mIoU, 3D Obj. Det. reports NDS, 3D Obj. Trk. reports AMOTA, and Occ. Pred. reports mean RayIoU. The appendix names BEVFusion for mapping and detection, ADA-Track for tracking, and SparseOcc for occupancy. These models are applied to synthetic inputs rather than retrained on them. Compare each score with its own empirical reference in the bottom row. A common percent sign does not make these four metrics interchangeable, and the generator roster differs from that of the action-following table. e16e17e18e19e20e15e27
What it supports. DiST-4D leads map segmentation, detection and tracking at 35.55%, 33.22% and 15.30%. DriveDreamer-2 leads occupancy at 26.82%, versus DiST-4D's 26.10%. All are below the listed empirical references. The benchmark therefore supports task-specific evaluation rather than transferring a model's favorable appearance ranking to every perception task.
Where the evidence stops. The neighboring prose claims a 30–40% advantage over next-best models; the listed map, detection and tracking gaps do not support it. These comparisons also do not establish the benefit of training a perception system on synthetic data.
Table 27. The highest mean rating still belongs to a distribution concentrated near the low end of the rubric. Original paper, p. 59 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the shaded mean row, then inspect the median and quartiles before interpreting the maxima. DiST-4D's mean is the highest, but every model's median is 2.0. The standard-deviation row describes variation among reported scores; it is not a standard error or an agreement statistic between annotators. Section 11.4 asks raters to examine motion continuity, object contact, occlusion ordering and lighting stability. The empirical reference of 10 follows the rubric's explicit ground-truth category. Keep that assigned reference distinct from a blinded experiment demonstrating that people consistently give every real recording a perfect rating. e04e21e23
What it supports. DiST-4D averages 2.583 with standard deviation 1.187; OpenDWM averages 2.312 with standard deviation 0.674. The common median of 2.0 makes the low typical rating clearer than the occasional high maximum. Reader analysis: this concentration motivates testing whether an automatic critic improves meaningfully over a simple low-score baseline.
Where the evidence stops. The table does not provide per-model sample counts or confidence intervals, so the significance of small mean differences remains unknown. Human-perceived physical plausibility is a rubric-based judgment, not direct measurement of forces or a certificate of driving safety.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Generation distributional fidelity Table 1's six-generator comparison against real driving video features; the precise common clip list and sampling uncertainty are not given. | DiST-4D: 58.08 I3D Fréchet Video Distance, lower is better | OpenDWM: 90.42; MagicDrive: 222.00. This is distributional appearance/motion evidence, not a direct physical-validity or control-success score. e03e14 |
| 4D reconstruction and novel-view evaluation OmniRe per-clip reconstruction, original poses and four novel-view conditions; road/vehicle-masked depth comparison with real-video reconstructions. | OpenDWM LPIPS: 0.065; DriveDreamer-2 AbsRel: 0.073; DiST-4D novel-view MUSIQ: 43.09% as printed, FVD: 192.39. LPIPS and AbsRel (lower); novel-view MUSIQ (higher) and FVD (lower) | DiST-4D LPIPS: 0.066 and AbsRel: 0.080; OpenDWM novel-view MUSIQ: 39.54%, FVD: 287.73. Different models lead different metrics. MUSIQ's printed percent is not a success rate; depth targets themselves come from reconstruction. No uncertainty intervals are reported. e07e08e09e14 |
| Closed-loop route following DriveArena-aligned five-sequence simulation on singapore-onenorth and boston-seaport, with planner decisions feeding back into generation. | RLGF: RC 13.51%, ADS 10.59%; open-loop PDMS 78.45%. Route Completion and Arena Driving Score; both higher are better | MagicDrive-V2: RC 12.31%, ADS 9.50%, open-loop PDMS 78.91%; MagicDrive: RC 6.89%, ADS 4.82%. RLGF leads reported closed-loop scores, while MagicDrive-V2 leads open-loop PDMS. Low route fractions expose a serious simulated-control gap, but no physical deployment or uncertainty estimate is established. e11e12e13e15 |
| Perception on generated observations Real-data-pretrained camera models evaluated on synthetic inputs against corresponding nuScenes task annotations; not synthetic-data retraining. | DiST-4D: 35.55%, 33.22%, 15.30%, 26.10%, respectively. Map mIoU, detection NDS, tracking AMOTA, occupancy mean RayIoU; higher is better | DriveDreamer-2: 33.62%, 30.90%, 13.30%, 26.82%. Empirical reference row: 40.64%, 44.72%, 36.30%, 37.05%. DiST-4D leads three tasks; DriveDreamer-2 leads occupancy. The reference row is an empirical comparator, not a proven ceiling. These methods differ from the action-following model roster. e15e16e17e18e19e20 |
| Human-rated physical plausibility Human review of generated driving clips under the Section 11.4 rubric; higher scores indicate greater perceived plausibility. | DiST-4D: mean 2.583, standard deviation 1.187; median 2.0. Rating on a 1–10 scale; mean and reported standard deviation | OpenDWM: mean 2.312, standard deviation 0.674; all six model medians are 2.0. Even the highest mean remains low. Standard deviations describe score dispersion, not confidence intervals or inter-annotator agreement; the rubric explicitly assigns ground truth a score of 10. e04e21 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. Changing the viewpoint exposes reconstruction defects that are less obvious in the generated frame. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each model column downward. The first row shows the generated scene; the caption identifies the second as a reconstruction rendered at a lateral offset, and the third as a depth map. Compare the truck outline, road surface and background across the first two rows, using the small insets to locate floating or fragmented appearance. MagicDrive and DreamForge show conspicuous reconstruction artifacts. The numerical overlays belong to this selected example. They should not be substituted for the dataset averages in Table 1, and the depth colors should be read qualitatively because no metric colorbar accompanies them. e29e07e08e09e14
What it supports. The example makes reconstructability concrete: a plausible-looking input can yield a damaged scene when viewed from a shifted camera. The authors use these artifacts to motivate geometry and temporal-consistency evaluation. Reader analysis: the observation diagnoses a failure of the generator-plus-reconstruction pipeline; it cannot uniquely attribute that failure to the generator's training objective.
Where the evidence stops. This is a comparison between different models, not a controlled ablation. The static crop cannot verify temporal smoothness, and both the fitted geometry and reference depths depend on the reconstruction method. Individual overlay metrics are not aggregate performance estimates.
7. Analysis & limitations
7.1 What the evidence leaves open
Driving-only scope, style/regional annotation bias, inherited critic weaknesses and evolving notions of physical realism are acknowledged limitations. e23
Cross-model comparisons do not isolate geometry supervision, training-data diversity or temporal conditioning. The paper's causal design recommendations exceed what controlled ablations establish. Claimed critic generalization is illustrated with selected cases, without aggregate agreement statistics or an untuned-backbone comparison. e14e20e22
Source inconsistencies remain unresolved: main/appendix depth and semantic extractors differ; Table 1's G.8 uses VC Match rather than Table 11's Total; three displacement values change in Table 16. The view-consistency ranking and geometric-error multiplier in the prose also contradict tables. This report retains location-specific values without silently repairing the source. e24e25e26e27
7.2 Questions for discussion
- How much of the closed-loop gap comes from generation drift versus the planner's sensitivity to synthetic observations?
- Does the critic improve agreement on unseen generators and rare high-quality clips beyond predicting the dominant low rating?
- How stable are reconstruction rankings under different Gaussian-field fitting budgets and masks?
8. Reproducibility audit
8.1 Requirements and known gaps
A faithful replication needs paired calibrated clips, task labels, exact generator/planner checkpoints, OmniRe settings, masks and route definitions. The paper specifies useful camera and simulation settings but not a complete common evaluation manifest, repeat counts or self-contained reconstruction loss configuration. e07e08e10e11e16
Critic replication additionally needs annotation records, scene/generator-disjoint splits and exact checkpoint identity. Batch size, video sampling limits and a held-out quantitative validation protocol are not specified in the training description. The proposed checks in the illustrated edition test feedback sensitivity and improvement over an untuned critic; neither was run here. e04e05e06e22e28
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test whether failure accumulates after closing feedback
Reader-proposed check, not performed: run MagicDrive-V2 and RLGF on the same five simulation sequences with identical planner checkpoint, initial ego states, image preprocessing and paired random seeds. Within each generator, compare the prescribed open-loop setting with closed-loop ego feedback, retaining the stated three-frame reference history. Log first-step planner disagreement, later visual drift, collision/off-road termination times, per-episode PDMS, RC and ADS; repeat seeds and report route-level variation. Compute ADS from each closed-loop episode, never from the open-loop summary column. If errors are already large before feedback begins, the evidence favors an initial observation/planner mismatch. If initially comparable runs deteriorate mainly after repeated feedback, it supports accumulated instability. Neither observation alone identifies a particular training mechanism. e10e11e12e13e15
Check 2: Test whether LoRA adds human agreement beyond a low-score prior
Reader-proposed check, not performed: first resolve the exact backbone checkpoint, then compare the frozen backbone, the reported LoRA critic and a predictor fixed at the training-set median. Hold out entire generators and source scenes, using the same RGB clips, dimension-specific prompt and decoding settings for both language models. Collect new judgments with the paper's two-group procedure, retain disagreements, and evaluate each criterion separately. Report absolute error and rank correlation, stratify by human score range, and estimate uncertainty by resampling scenes rather than correlated frames. A useful critic should outperform both baselines on unseen generators and uncommon high-score clips, not merely reproduce the dominant rating near 2. Failure to do so would weaken the claimed benefit of preference adaptation. e04e05e06e21e22e28
8.3 Reading coverage
Visual audit: The title/byline, metric definitions, reconstruction settings, planning protocols, perception implementations, critic training hardware and prompt, numerical comparison tables, human rubric and limitations were visually inspected. Six final original crops were individually viewed with headers and comparison rows retained. Figure 3 supplies the method visual; Figure 5 is a diagnostic, not a controlled ablation. Figure 35's visible arrows and frozen-module markers were cross-checked with Section 12.1, and its multiple illustrated outputs were distinguished from the single-dimension prompt on p. 67. Some embedded diagram labels on uncropped pages are absent in the renderings; the readable text supplies the retained mechanism claims. No unseen labels or continuous video behavior are inferred. Other appendix example images and external media remain outside the visual pass.
PDF pages inspected for this edition: 1, 4, 5, 6, 7, 8, 9, 10, 11, 19, 23, 27, 29, 31, 33, 35, 37, 38, 39, 41, 43, 45, 47, 49, 51, 53, 59, 65, 66, 67, 68, 69, 70, 72. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract and title/byline
- 1 Introduction
- 2 Related Work
- 3 WorldLens: A Full-Spectrum Benchmark (3.1–3.5)
- 4 Human Annotation & Evaluation Agent (4.1–4.3)
- 5 Experiments (5.1–5.3)
- 6 Conclusion
- Appendix contents
- 7 Generation (7.1–7.8, definitions, equations, implementations, examples and results)
- 8 Reconstruction (8.1–8.4)
- 9 Action-Following (9.1–9.4)
- 10 Downstream Task (10.1–10.4)
- 11 Human Preference (11.1–11.6 and all rubrics)
- 12 Evaluation Agent (12.1–12.4 and printed prompt)
- 13 Broader Impact & Limitations
- 14 Public Resource Used
- References [1]–[148]
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- All 17 supplied text chunks covering the 80-page PDF were read. Visual inspection covers the pages declared in the illustrated edition, not every example image in the appendix.
- The inspected artifact is arXiv v2 dated 1 June 2026, later than the catalog submission date of 11 December 2025. No earlier edition or revision history was supplied, so changes from v1 and the catalog's CVPR 2026 venue label were not independently verified.
- The title page credits WorldBench Team but gives no individual institutional affiliations; none are inferred.
- The supplied extraction does not reconstruct images. Original PDF pages and six final crops were inspected separately. Some embedded labels in Figures 2, 4, 7, 8 and 35–37 are absent from the page renderings; no claim depends on recovering these labels. Figure 35's mechanism is verified from the readable Section 12.1 text.
- Separate supplemental material availability has not been fully verified; none was supplied.
- Linked code, datasets, model weights, project pages and external references were not opened. No experiment or installation was performed.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, complete byline and arXiv margin
The title and all 22 named authors match the supplied catalog. The artifact is arXiv:2512.10958v2 [cs.CV], 1 June 2026. It credits WorldBench Team without individual institutional affiliations.
Go to primary source ↓e02PDF pp. 4–7, Figure 2 caption and Sections 3.1–3.5
WorldLens evaluates Generation, Reconstruction, Action-Following, Downstream Task and Human Preference, organized into eight generation dimensions and four dimensions in each other aspect.
Go to primary source ↓e03PDF p. 4, G.1–G.5; p. 5, G.6–G.8
Generation probes instance confidence, ReID identity stability, DINO consistency, depth continuity, CLIP temporal consistency, semantic stability, I3D FVD and LoFTR cross-view correspondence.
Go to primary source ↓e04PDF p. 7, Sections 4.1–4.2 and Figure 3
Ten annotators in two independent groups assess synchronized video, semantics, depth and 3D boxes, with re-evaluation of disagreements. WorldLens-26K contains 26,808 score-and-rationale records; this is not stated to be the number of unique videos.
Go to primary source ↓e05PDF p. 65, Sections 12.1–12.3; p. 66, Figure 35
The critic is described as Qwen3-VL-8B with attention-only LoRA; the vision encoder, projector, embeddings and MLPs remain frozen. Training uses rank 16, dropout 0.05, three epochs, learning rate 1e-4, cosine decay, warmup ratio 0.1, eight A100 GPUs and bfloat16.
Go to primary source ↓e06PDF p. 67, complete prompt scheme, Evaluation Target, Score Range and Instruction
The printed prompt requests one target dimension at a time and a JSON object containing score and reason. Scores range from 1 to 10 in steps of 0.5, whereas the architecture illustration presents multiple output dimensions.
Go to primary source ↓e07PDF p. 29, Sections 8.1.1–8.1.3, Eq. (9), Table 12
Each nuScenes-style six-camera clip has 16 frames per camera at 12 Hz and 544×304 resolution. OmniRe reconstruction optimizes one 4D Gaussian field for 30,000 steps and evaluates original-view LPIPS, PSNR and SSIM. Eq. (9) averages LPIPS; its following sentence says higher is better, contrary to Table 12's downward arrow and Section 3.2.
Go to primary source ↓e08PDF p. 31, Sections 8.2.1–8.2.3, Eq. (10), Table 13
Generated and real videos are separately reconstructed with identical poses and settings. Rendered depths are compared within road-and-vehicle masks obtained from ground-truth images using Grounded SAM 2. AbsRel is 0.073 for DriveDreamer-2, 0.080 for DiST-4D and 0.088 for OpenDWM.
Go to primary source ↓e09PDF p. 33, Sections 8.3.1–8.3.3 and Table 14; p. 35, Sections 8.4.1–8.4.3 and Table 15
Novel views use center interpolation, an S-curve and left/right lateral offsets of 1 m. MUSIQ assesses frame quality; I3D FVD compares novel-view videos from generated and real reconstructions. DiST-4D averages 43.09% MUSIQ as printed and 192.39 FVD.
Go to primary source ↓e10PDF p. 37, Sections 9.1.1–9.1.3 and Eq. (15)
Displacement Error compares UniAD waypoints inferred from generated versus corresponding real videos on the 150-scene nuScenes validation set at a one-second planning horizon. It compares two planner predictions, not a prediction directly against the recorded future trajectory.
Go to primary source ↓e11PDF p. 39, Sections 9.2.1–9.2.3, Eq. (16), Table 17
Open-loop PDMS combines no-collision and drivable-area penalties with progress, time-to-collision and comfort. Five validation sequences span singapore-onenorth and boston-seaport; traffic runs at 10 Hz and control at 2 Hz. Generation uses three reference frames, produces 448×800 images, and resizes them to 224×400 for the driving agent.
Go to primary source ↓e12PDF p. 41, Sections 9.3.1–9.3.3, Eq. (17), Table 18
Closed-loop planner controls update the simulated ego state, which conditions the next generated frame. Route completion is traveled distance divided by planned distance. Episodes end on completion or safety failure; the definition also mentions timeout.
Go to primary source ↓e13PDF p. 43, Sections 9.4.1–9.4.3, Eq. (18), Table 19
Arena Driving Score is episode-level RC multiplied by PDMS, then averaged across episodes. Table 19 reports closed-loop component scores, with rounded totals of 0.095 for MagicDrive-V2 and 0.106 for RLGF.
Go to primary source ↓e14PDF p. 8, Table 1, all model rows, G.7 and R.1–R.4 columns
DiST-4D reports generation FVD 58.08 versus OpenDWM 90.42 and MagicDrive 222.00. Original-view LPIPS is 0.066, 0.065 and 0.140 respectively. The lowest geometric AbsRel belongs to DriveDreamer-2; DiST-4D leads novel-view quality and discrepancy.
Go to primary source ↓e15PDF p. 9, Table 2, MagicDrive, MagicDrive-V2 and RLGF rows
MagicDrive-V2 has open-loop PDMS 78.91%, RC 12.31% and ADS 9.50%; RLGF has 78.45%, 13.51% and 10.59%; MagicDrive has 71.23%, 6.89% and 4.82%. Panacea has no reported A.2–A.4 values.
Go to primary source ↓e16PDF p. 45, Section 10 introduction, Sections 10.1.1–10.1.3 and Table 20
Real-data-pretrained perception models are applied directly to generated videos. Map segmentation uses the stated camera-only BEVFusion setup and evaluates mIoU against corresponding BEV labels.
Go to primary source ↓e17PDF p. 47, Sections 10.2.1–10.2.3 and Table 21, NDS row
The pretrained detector is evaluated against nuScenes ground-truth boxes. NDS is 0.3322 for DiST-4D, 0.3090 for DriveDreamer-2, 0.2196 for OpenDWM and 0.4472 in the empirical reference column.
Go to primary source ↓e18PDF p. 49, Sections 10.3.1–10.3.3 and Table 22, AMOTA row
Camera-only ADA-Track runs directly on generated multi-view videos. AMOTA is 15.30% for DiST-4D, 13.30% for DriveDreamer-2 and 6.90% for OpenDWM, versus the 36.30% empirical reference.
Go to primary source ↓e19PDF p. 51, Sections 10.4.1–10.4.3, Eq. (22), Table 23
Frozen SparseOcc occupancy is assessed by class and frontmost-ray depth agreement, averaged over 1, 2 and 4 m tolerances. Mean RayIoU is 26.82% for DriveDreamer-2 and 26.10% for DiST-4D, versus 37.05% in the empirical reference column.
Go to primary source ↓e20PDF p. 9, Table 3, all model rows and D.1–D.4 columns
DiST-4D reports 35.55% mIoU, 33.22% NDS, 15.30% AMOTA and 26.10% occupancy score. DriveDreamer-2 reports 33.62%, 30.90%, 13.30% and 26.82%, leading occupancy rather than DiST-4D.
Go to primary source ↓e21PDF p. 59, Section 11.4, scoring rubric and Table 27
Physical plausibility ratings examine continuity, contact, occlusion and illumination. DiST-4D's mean is 2.583 with standard deviation 1.187; OpenDWM's is 2.312 with standard deviation 0.674. All six model medians are 2.0; the rubric assigns 10 to ground truth.
Go to primary source ↓e22PDF p. 10, Section 5.2, Human-Agent Alignments; p. 11, Figure 8; pp. 65–66, Section 12.4; pp. 68–69, Figures 36–37
The authors claim agreement and generalization on unseen Gen3C, Cosmos-Drive and CARLA videos. Evidence presented consists of selected scores and explanatory examples, without an aggregate agent-versus-human correlation, error table or controlled agent ablation.
Go to primary source ↓e23PDF p. 70, Section 13.3, Potential Limitations
The authors identify driving-only scope, annotator style/regional biases, inherited language-model and supervision limitations, and the open-ended nature of physical realism metrics.
Go to primary source ↓e24PDF p. 4, G.4 versus p. 19, Section 7.4.3; p. 5, G.6 versus p. 23, Section 7.6.3
Main text names Depth Anything V2 and SegFormer, whereas appendix implementation sections name Video DepthAnything and OpenSeeD. The supplied source does not resolve which configurations produced all reported scores.
Go to primary source ↓e25PDF p. 8, Table 1, G.8 column and Section 5.1 Generation paragraph; p. 27, Eq. (8) and Table 11
Table 1 G.8 reproduces Table 11's VC Match values, not its Total values: DiST-4D 389.78 versus OpenDWM 211.18, while Table 11 Total is 0.4574 versus 0.3720. The prose incorrectly says DiST-4D underperforms OpenDWM in view consistency.
Go to primary source ↓e26PDF p. 9, Table 2, Displacement Error; p. 38, Table 16, Total row; p. 37, Eq. (15)
Main/appendix displacement values differ for DreamForge (0.55/0.57), DrivingSphere (0.54/0.55) and MagicDrive-V2 (0.53/0.54). Both give an empirical reference of 0.51, whose nonzero value is unexplained under a literal same-planner, real-versus-real interpretation of Eq. (15).
Go to primary source ↓e27PDF p. 8, Table 1, R.2; p. 9, reconstruction prose and downstream prose beside Table 3
The claim that MagicDrive's geometric error is over twice OpenDWM's conflicts with 0.115 versus 0.088. The claimed 30–40% downstream advantage over the next-best models is not supported by Table 3's map, detection and tracking gaps.
Go to primary source ↓e28PDF p. 65, Sections 12.1 and 12.3; p. 72, reference [2] and Qwen3-VL resource entry
The method repeatedly names Qwen3-VL-8B, but its cited reference [2] is titled Qwen2.5-vl technical report. The resource list separately names Qwen3-VL; an exact checkpoint revision is not supplied.
Go to primary source ↓e29PDF p. 9, Figure 5, model headers, three rows, inset metrics and caption
The diagnostic compares reference and generated truck scenes, lateral-offset renderings and depth maps. MagicDrive and DreamForge show marked floating/distorted structure. Insets report individual-example LPIPS and AbsRel, not dataset averages; no metric depth colorbar is provided.
Go to primary source ↓8.5 Primary sources
WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World ↗
PDF · 26,757 extracted words
Source fingerprint
d9e19bd6b0718d9172e0f63e25508708d020d99e922a1bc33e7382dd8f6a295b