acceptodds
Under review as a conference paper at ICLR 2027

GeoSCope: Grounded, Stable and Controllable Driving World Models

65 views

Abstract

Generative driving world models produce realistic video but lack grounding in a real location. Conditioned on their own predictions, they modify road layout and surrounding structures, causing long-horizon rollouts to diverge from the recorded scene. A policy evaluated in such rollouts is effectively tested on a different scene than the one recorded. We propose GeoSCope, a method that grounds driving world models in the geometry of a real route using only a single monocular recording. A dynamics-aware SLAM pipeline reconstructs a static 3D map of the route, which is rendered at the camera pose of each frame prior to its generation and used to condition a generative world model. Although the rendered signal is sparse and imperfect, the model learns to exploit it selectively. Experiments with three world models on four datasets show that this anchoring substantially reduces long-horizon drift and preserves scene structure without suppressing dynamic objects. Since the map can be rendered along arbitrary trajectories for conditioning, the same model supports counterfactual rollouts, turning existing video-based world models into grounded, stable, and controllable neural simulators. With geometric conditioning trained only on nuPlan, the approach transfers zero-shot to Waymo, MARS and nuScenes, and sets a new state of the art on the Ann-Arbor-City-Bench multi-traversal benchmark. Code and weights will be released upon acceptance.

Primary area:
generative models
Keywords:
driving world models, conditioned video generation, geometric conditioning, autonomous driving

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uEqrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GeoSCope, a framework designed to improve the long-horizon stability and controllability of generative driving world models. The method first reconstructs a static 3D point cloud from a single monocular driving recording using a dynamics-aware SLAM pipeline. During generation, the reconstructed scene is rendered at the target camera poses and used as geometric conditioning for pretrained world models, helping preserve static scene structure while allowing the models to generate dynamic content. The proposed framework is evaluated on three driving world models across four datasets, demonstrating improvements in long-horizon geometric consistency and generalization to unseen datasets.

Critical strengths
  1. This paper focus on an important problem in driving world models, where scene geometry can gradually drift during long video generation. Using reconstructed 3D geometry to keep the generated scenes consistent is a simple and effective idea that improves the reliability of long-horizon generation.

  2. The paper provides clear explanations of the method and implementation details, including the SLAM reconstruction, point cloud filtering, and geometric conditioning mechanisms. The appendix also includes detailed configurations and training settings, making the approach easier to understand or reproduce.

Critical weaknesses

1. The claimed controllability and interactive simulation capabilities are not sufficiently supported. The method relies on rendering geometric references from a pre-reconstructed scene for each target frame, which limits its effectiveness when the vehicle moves into previously unobserved areas.

For example, if the original recording follows a left turn, it is unclear how the model would handle a counterfactual right turn where no reconstructed geometry is available. The current experiments only consider small lateral offsets, trajectory wiggles, and speed changes near the original route, without evaluating such meaningful route deviations. This limitation is not sufficiently discussed and significantly reduces the practical contribution of the proposed framework as an interactive driving world model.

Overall, I think this work is closer to scene completion within a reconstructed 3D environment than to a truly controllable and interactive driving world model.

2. The main evaluation relies on a highly favorable same-route reconstruction setting. Most of the main experiments rely on 3D point clouds reconstructed from the complete video of the same route being evaluated, including frames from the prediction horizon.

This provides the model with substantial additional information about the target scene, including both geometry and appearance, which is unavailable to the baselines. Therefore, the improvements in generation quality and consistency metrics are somewhat expected, and it is unclear how much of the gain comes from the proposed model design rather than simply having access to more information.

Combined with the limitations discussed in Weakness 1, this raises concerns about both the practical applicability of the method and the significance of its reported improvements.

Questions for authors
  1. How does the proposed method handle counterfactual trajectories that go beyond the reconstructed region? (in Weakness 1)

  2. How should the performance improvements be interpreted given the same-route reconstruction setting?(in Weakness 2)

  3. Could the authors provide a fairer comparison that isolates the effect of geometric conditioning from additional fine-tuning? According to Section 3.5, all learned GeoSCope variants fine-tune the entire pretrained backbone together with the newly introduced conditioning modules.. However, the main experiments compare these models against the original pretrained checkpoints without equivalent additional fine-tuning.

  4. Why does GeoSCope lead to worse video quality on Cosmos 2.5?

3 comments

We thank the reviewer for the detailed and helpful review.

W1 + Q1: Controllability and trajectories beyond the reconstructed region. We agree that the map only provides geometry where the original drive observed the scene, and we now state this scope explicitly in the introduction: GeoSCope anchors generation within and near the recorded route, and beyond it the model falls back to its learned prior. To quantify what happens when the ego leaves the map, we built maps from only the first L = 1, 2, 4 or 8 s of each clip and froze them, so the rollout drives out of the map (new Appendix J.9). Inside the map, performance follows the full-map curve. Outside it, performance degrades smoothly toward the base model without falling below it (Table R5).

Table R5. nuPlan-Turns 172-clip subset, 10–20 s window (outside every frozen map).

ModelMap fromLPIPS ↓FVD ↓KP matches ↑
Orbis v1–0.54665163
+ GeoSCopefirst 1 s0.52153472
+ GeoSCopefirst 2 s0.51054179
+ GeoSCopefirst 4 s0.50152884
+ GeoSCopefirst 8 s0.48949393
+ GeoSCopefull clip0.383338204

W2 + Q2: Same-route setting. The same-route setting is an intended use case: re-simulating a recorded drive with counterfactual ego motion, as reconstruction-based simulators and recent industrial closed-loop simulators (e.g. GAIA-4; Wayve, 2026) do. Here the map is an input, comparable to an HD map for a map-conditioned planner. Our contribution is to show that a sparse map from a single monocular pass is sufficient, and that it need not come from the evaluated video:

(i) On Ann-Arbor-City-Bench (Table 3), the map is reconstructed from a different drive on a different day, so the map contains no frame of the evaluated video. GeoSCope, never trained on multi-traversal data, outperforms the Seoul World Model (FVD 487 vs. 640, mPSNR 18.6 vs. 15.2 dB).

(ii) In Table R5, the map is built only from frames before second L, yet GeoSCope still outperforms the base model beyond it.

(iii) The model does not copy the map: the rendered map alone scores far worse (FID 201 vs. 39, FVD 2298 vs. 487; Table 3). In new cross-traversal pairs where the map drive and the target drive differ in weather or season, GeoSCope keeps the target's appearance and improves LPIPS and keypoints in every condition (reply to Reviewer rwS6, Table R2).

(iv) The gains do not come from the additional fine-tuning on nuPlan either (Q3).

Q3: Fine-tuning control. To isolate the conditioning signal, we fine-tuned Orbis v1 with exactly the GeoSCope recipe but without the geometric reference.

Table R6. Average over the four datasets (protocol of Table 2 with short / long = 0–8 s / 8–16 s).

ModelLPIPS ↓FVD ↓KP matches ↑Depth AbsRel ↓
Orbis v1 (released)0.417 / 0.508323.9 / 539.4217.3 / 104.80.214 / 0.307
Orbis v1 fine-tuned, no geo-reference0.431 / 0.518342.1 / 573.8199.5 / 88.50.225 / 0.318
GeoSCope (Orbis v1)0.345 / 0.411227.1 / 325.0267.8 / 184.50.157 / 0.217

Averaged over the four datasets, the control is not better than the released checkpoint, while GeoSCope, trained with the same data and budget, is best on every metric on every dataset. Without the geo-reference, fine-tuning helps in-domain (nuPlan long FVD 440 → 270) but hurts transfer, drifting toward the training domain. We have added this control to Table 2.

Q4: Video quality on Cosmos 2.5. We agree that the Cosmos variant had worse FVD than the base model, although it improved LPIPS, keypoints and depth. We found two problems:

(i) Rollouts drift in appearance: contrast and saturation rise chunk by chunk, because classifier-free guidance on the map signal compounds over the autoregressive rollout.

(ii) The gates were unstable during training due to their initialization.

We fixed (i) by lowering the guidance weight from 3 to 1.5 and (ii) by initialising the gate open with a zero-initialised output projection. Averaged over the four datasets, FVD drops by about 30% compared with the submitted variant and is 10–12% lower than the base model. LPIPS and FVD slope improve as well, and long-horizon keypoints remain more than twice the base model's (Table R7). We have updated Tables 2 and 15.

Table R7. Cosmos 2.5, average over the four datasets (Table 2 protocol with short / long = 0–8 / 8–16 s; all models at 480p).

ModelLPIPS ↓FVD ↓FVD slope ↓KP matches ↑
Cosmos 2.50.433 / 0.514336.6 / 503.943.5300.7 / 136.4
+ GeoSCope (submitted)0.358 / 0.421418.7 / 623.357.7421.6 / 328.3
+ GeoSCope (retrained)0.341 / 0.412285.0 / 455.142.3381.0 / 278.7

If these clarifications and additional results address your concerns, we would be grateful if you would consider raising your score.

Reviewer_uEqrOpenReview ↗

Response to authors

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns.

However, I still have concerns about the practical value of the proposed method. As the authors acknowledge, GeoSCope is mainly effective within the reconstructed region, and its performance approaches that of the base model once the vehicle moves beyond it. Given this limitation, I do not think the rebuttal sufficiently addresses my concerns about the claimed controllability and interactive simulation capabilities. Moreover, reconstructing a 3D point cloud introduces additional computational and deployment overhead, which further limits the practical applicability of the proposed world model.

Considering these limitations, along with the concerns raised by other reviewers, I have decided to maintain my original rating of 2.

Thank you for engaging with our response. We briefly clarify two points for the record.

Scope. GeoSCope targets re-simulation of recorded drives, e.g. evaluating a policy with counterfactual ego motion on a logged route. This is the same setting as reconstruction-based simulators (e.g. NAVSIM v2) and industrial closed-loop simulators (e.g. GAIA-4), and reconstruction-based methods typically break down outright beyond the reconstructed region whereas GeoSCope falls back to the base world model. GeoSCope substantially reduces drift within this setting and never performs worse than the base model beyond it (Table R5), so grounding is a strict improvement rather than a trade-off.

Overhead. Map reconstruction is a one-time offline step per route from a single monocular recording, reused across all rollouts. Rendering the point cloud at each pose is cheap compared with the world model's denoising steps. Compared with methods that require dense geo-registered imagery, this is a lightweight requirement.

Reviewer_UzuEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the structural drift problem in generative driving world models by combining static SLAM-based scene reconstruction with dynamic video generation, where the reconstructed static scene serves as a reference to guide the world model toward more consistent long-horizon generation. Extensive experiments across multiple datasets, including zero-shot evaluations and ablation studies, demonstrate the effectiveness of the proposed approach. Overall, this work presents a framework-level integration of static SLAM reconstruction and generative driving world models, which is valuable for counterfactual simulation in pre-recorded environments. However, dynamic modeling and long-horizon prediction still largely rely on the capabilities of the underlying pretrained world models.

Critical strengths

-The motivation is clear and well-aligned with contributions. The authors target the structural drift, a common limitation of generative driving world models. Using static SLAM reconstruction as a geometric reference is a straightforward and intuitive solution. -The proposed approach addresses two important aspects: improving long-horizon generation consistency and enabling controllable counterfactual trajectory generation for driving world models. -The authors conduct experiments across four datasets, including zero-shot evaluations, and further assess the method on Ann-Arbor-City-Bench. Extensive ablation studies investigate different design choices, while qualitative visualizations and supplementary videos demonstrate the generation results.

Critical weaknesses

-The framework depends on pretrained world models heavily. As the method primarily integrates static SLAM reconstruction with existing generative world models, the quality of generated videos, particularly the modeling of dynamic objects, still largely depends on the underlying world model. Consequently, the main improvements appear to focus on static scene consistency and drift reduction, with relatively limited attention to dynamic modeling. -The work pays limited attention to counterfactual dynamics modeling, as counterfactual generation still heavily relies on the pretrained world model to predict dynamic agents. Since the static reconstruction provides no explicit dynamic information, the proposed framework does not directly address the challenges of dynamic consistency under alternative ego trajectories. This raises concerns about its reliability as a counterfactual driving simulator.

Questions for authors

-Counterfactual generation appears to rely on real context frames and the pretrained world model to initialize and predict dynamic agents. How does it handle scenarios where the initial ego pose differs from the recorded trajectory, such that no corresponding real context observations are available? How can dynamic agents be generated and maintained consistently under these alternative initial poses and subsequent trajectories? -Section 5.4 and Figure 3 only evaluate whether the number of dynamic objects is preserved. However, object count alone is insufficient to demonstrate reliable dynamic modeling. Could the authors provide quantitative evaluations of dynamic properties, such as object trajectory accuracy, velocity, acceleration, or interaction consistency, particularly under counterfactual generation?

1 comment

We thank the reviewer for the positive assessment and the precise questions on dynamics.

W1: Dependence on the pretrained world model. This is a deliberate design choice: GeoSCope anchors only the static scene and leaves dynamics to the world model. This keeps the method backbone-agnostic (three world models in the paper) and lets it benefit directly from better world models. Modelling dynamics explicitly in 3D would require separate models to lift, track and control other agents, which is exactly what limits reconstruction-based simulators to labelled data.

Anchoring the static scene does not degrade dynamics. Instead it improves where agents are placed in the scene (Table R3). Vehicle position error is roughly halved and close to the measurement floor (1.19–1.23 m vs. 1.02 m), moving vehicles stay in their lanes more often (93–95% vs. 81–85%), and the share of vehicles with no counterpart in the log drops to about a third (12–13% vs. 34–36%).

Interaction dynamics such as crossing traffic come from the base model and are not improved, which we have added to the limitations (Sec. 6).

W2 + Q2: Dynamic properties, including under counterfactuals. We agree that object count alone is insufficient and added an agent-level evaluation against the nuPlan lidar ground truth: on 400 nuPlan-Turns clips we detect vehicles in the generated frames, lift them to metric ground-plane positions using the recorded ego pose and camera calibration, track them over 4 s, and compare them with the logged 3D tracks and the HD map. Errors are measured in the world frame of the logged drive, so they reflect both where the model places agents and how well the rollout follows the logged ego path. We restrict the comparison to the first 4 s, because beyond that the generated traffic legitimately diverges from the log.

Table R3. Vehicle tracks in the first 4 s of 400 nuPlan-Turns rollouts. The same pipeline on the recorded video gives the measurement floor. Invented: tracks with no lidar vehicle within 2.5 m.

ModelPosition error (m) ↓Speed error, moving (m/s) ↓Moving vehicles in lane ↑… on sidewalk ↓Invented vehicles ↓
Recorded video (floor)1.020.7297%0%7%
Orbis v12.391.6081%4%36%
+ GeoSCope1.231.1893%0%12%
Orbis v22.281.5085%5%34%
+ GeoSCope1.190.9695%1%13%

Counterfactual rollouts have no ground truth, so we measure world-frame consistency instead. With the ego path shifted by about 2 m on all 400 clips, no tracked vehicle overlaps the ego in any of the 1,200 rollouts, parked vehicles stay anchored in the world (0.45–0.46 vs. 0.51 on the logged path), and moving vehicles stay in their lanes (82–90% vs. 93%).

How other agents react to a changed ego trajectory comes from the world-model prior, and we do not evaluate reactive behaviour. Existing reactivity metrics operate on agent states, which a video world model does not expose, and we are not aware of an equivalent for pixel-space rollouts.

Q1: Initial pose off the recorded trajectory. GeoSCope needs a few context frames to initialize appearance and dynamic agents. Two settings in our evaluation start without context recorded at the map's poses:

(i) In the cross-traversal benchmarks (Ann-Arbor-City-Bench and the weather pairs in our reply to Reviewer rwS6), the context comes from a different drive on a different day than the map.

(ii) To test a start without any recorded frame, we replaced all context frames with noise, so the model only sees the map at the start pose (Table R4). The short horizon is worse than with recorded context, as appearance and traffic must be invented, but on par with or better than the base model with recorded context. Over the long horizon, it approaches GeoSCope with recorded context. The generated vehicles are plausible but do not correspond to the real traffic, which only context frames can provide. For off-log starts, real context is thus needed to initialize dynamic agents. Box-conditioned driving video models such as MagicDrive-V2 (Gao et al., 2025) show that agents can also be specified explicitly, which could be combined with GeoSCope in future work.

Table R4. 400 nuPlan-Turns clips (protocol of Table 2, short / long = 0–8 s / 8–16 s).

ModelContextLPIPS ↓FVD ↓KP matches ↑
Orbis v2recorded0.440 / 0.536222.3 / 333.8178.7 / 63.6
+ GeoSCopenone (map only)0.405 / 0.383222.7 / 211.1222.7 / 239.2
+ GeoSCoperecorded0.304 / 0.360131.9 / 183.3298.1 / 249.4

We have added both analyses, the agent-level evaluation and the no-context start, to the appendix. If these clarifications and additional results address your questions, we would be grateful if you would consider raising your score.

Reviewer_rwS6OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoSCope, a framework that grounds driving world models with the help of geometry in a single monocular recording. The pipeline includes a dynamics-aware SLAM approach and future frame(s) generation conditioned on context and geo-reference frames.

Critical strengths
  • Extensive experiments have been conducted on three world models and four datasets.
  • The proposed geometric grounding substantially improves long-horizon structural consistency.
Critical weaknesses
  • Limited novelty. The pipeline largely relies on the established SLAM pipeline and world models.
  • Further evaluation under weather or lighting differences between the monocular recording and the context frames would be helpful.
Questions for authors
  • What would happen if the given monocular recording has a different weather or lighting condition as the context frames?
2 comments

We thank the reviewer for the constructive feedback and for recognising the breadth of the evaluation and the gains in long-horizon structural consistency.

W1: Novelty. GeoSCope deliberately reuses an off-the-shelf SLAM system and existing world models, which is what makes it model-agnostic (three backbones, Sec. 3.5). The contribution is the interface between the two: we anchor a generative world model with a static map reconstructed from a single monocular drive and rendered at the pose of each upcoming frame. This keeps the scene consistent over long horizons and enables counterfactual rollouts, while dynamics continue to come from the base world model.

Our experiments show this is not a plug-in combination. Training-free latent blending only reduces long-horizon FVD from 501.3 to 433.5, whereas learned conditioning reaches 188.4 (Table 6). A model trained on aligned renders drops to FVD 728.4 under 0–3 m localization error, worse than the unconditioned base model, and our pose-perturbed training recovers 299.3 (Table 5).

Retrieval-based approaches such as CityRAG and the Seoul World Model (SWM) instead require dense geo-registered imagery. Although GeoSCope is never trained on multi-traversal data, both variants outperform SWM on SWM's own cross-traversal benchmark in FVD, mLPIPS, mPSNR and FID (Table R1, values from Table 3 of the main paper).

Table R1. Comparison with the best Seoul World Model (SWM) variant per metric on Ann-Arbor-City-Bench (values from Table 3 of the main paper). GeoSCope is never trained on multi-traversal data. Change relative to the best SWM value in parentheses.

MetricOrbis v1 + GeoSCopeOrbis v2 + GeoSCopeBest SWM
FVD ↓486.69 (−24.0%)443.16 (−30.8%)640.17 (TF)
mLPIPS ↓0.313 (−34.9%)0.326 (−32.2%)0.481 (TF)
mPSNR ↑18.61 (+3.43 dB)17.03 (+1.85 dB)15.18 (TF)
FID ↓39.27 (−10.7%)39.34 (−10.5%)43.97 (SF)
Img. Q. ↑0.600.650.66 (TF)

W2 + Q1: Weather and lighting differences between the map recording and the context. Our main cross-traversal results (Sec. 5.2, Ann-Arbor-City-Bench) already reconstruct the map from one traversal and condition on a second traversal recorded on a different day, so appearance never matches exactly. To test this further, we mined additional cross-traversal pairs in which the map and the target drive differ in weather or season, and therefore also in illumination: a sunny map for a rainy drive, a rainy map for a sunny drive, and an autumn map for a winter drive.

GeoSCope follows the map's geometry (road layout, buildings, turns) while keeping the weather, lighting and appearance of the target drives context frames. For example, a sunny map does not make a rainy rollout sunny. Quantitatively (Table R2), GeoSCope improves LPIPS and keypoint matches in every condition compared to the Orbis v2 base model, and keeps 1.8× as many keypoint correspondences with the target drive in the second half of the rollout (203 vs. 111). FID improves overall and in two of three conditions. FVD is lower in the second half (1185 vs. 1576), where the base model drifts, and higher in the first half (856 vs. 760), where it has not yet drifted. With 27 clips, FVD is noisy and should be read as a trend only. Representative rollouts are shown in the new Appendix J.8 (Weather and Lighting Differences) of the revised paper.

Table R2. Cross-traversal pairs of daytime drives whose map and target drive differ in weather, season and the resulting illumination (27 MARS pairs, Orbis v2 backbone; number of pairs in parentheses). All metrics are computed against the target drive. LPIPS and keypoint (KP) matches are given as full 18 s / second half (9–18 s). FVD is pooled over all pairs (first / second half) and omitted per condition because the subsets are too small.

ConditionMethodLPIPS ↓KP matches ↑FID ↓FVD ↓
Sunny map → rainy drive (12)Orbis v20.428 / 0.499170 / 9464.5–
+ GeoSCope0.404 / 0.442234 / 20371.5–
Rainy map → sunny drive (9)Orbis v20.461 / 0.524161 / 9782.5–
+ GeoSCope0.426 / 0.458206 / 17769.5–
Autumn map → winter drive (6)Orbis v20.329 / 0.450275 / 16767.0–
+ GeoSCope0.321 / 0.399310 / 24362.8–
All (27)Orbis v20.417 / 0.497190 / 11148.3760 / 1576
+ GeoSCope0.393 / 0.438242 / 20345.9856 / 1185

If these clarifications and additional results address your questions, we would be grateful if you would consider raising your score.

Thank you for the detailed response and additional experiments.

The new cross-traversal experiments under different weather and seasonal conditions adequately address my concern regarding appearance discrepancies between the reconstructed map and the target context. In particular, the improvements in long-horizon perceptual similarity and keypoint correspondences provide useful evidence of the robustness of the proposed geometric grounding.

The additional ablations also help clarify the contribution of the learned geometric conditioning and pose-perturbed training. However, I still have some reservations regarding the methodological novelty, as the framework primarily integrates established reconstruction and generative modeling components.

Overall, the rebuttal has addressed my main concern about robustness and strengthened the experimental evidence. I will therefore raise my rating accordingly.

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback.

GeoSCope grounds driving world models in a 3D map reconstructed from a single monocular drive. Across three backbones and four datasets, sparse map renders reduce long-horizon drift while preserving dynamic objects, and support counterfactual rollouts near the recorded route. Trained only on nuPlan, the geometric conditioning transfers zero-shot to Waymo, MARS, and nuScenes and outperforms the Seoul World Model on the Ann-Arbor-City-Bench cross-traversal benchmark in FVD, mLPIPS, mPSNR and FID.

The rebuttal and revised paper add the following evidence:

(i) Weather and lighting (rwS6, W2+Q1). Across 27 challenging cross-traversal pairs whose map and target drive differ in weather or season, GeoSCope preserves the target drive’s appearance and improves LPIPS and keypoint matches in every condition (Table R2, Appendix J.8).

(ii) Learned grounding, not a plug-in combination (rwS6, W1). GeoSCope needs only a sparse map from a single monocular pass, while retrieval-based approaches such as SWM require dense geo-registered imagery. Making this sparse signal useful requires learned conditioning. Training-free blending only reaches 433.5 FVD vs. 188.4 with learned conditioning (Table 6). Under 0–3 m localization error, aligned-only training degrades to 728.4. Our pose-perturbed training recovers 299.3 (Table 5).

(iii) Traffic and off-log starts (UzuE, W1+W2+Q1+Q2). Against nuPlan lidar tracks, GeoSCope roughly halves vehicle position error (Table R3, Appendix J.10). We added shifted-path consistency tests and map-only starts. Recorded context is needed to recover actual initial traffic (Table R4, Appendix J.11).

(iv) Independent maps and coverage limits (uEqr, W1+Q1+W2+Q2). When the ego leaves a frozen partial map, performance degrades gradually toward the base model without falling below it. We clarified this scope in the introduction (Table R5, Appendix J.9). With a map from a different drive on Ann-Arbor-City-Bench, GeoSCope reaches an FVD of 487, compared with 640 for the Seoul World Model. GeoSCope does not simply copy the map: the map render alone scores FID 201 vs. 39 for GeoSCope (Tables R1 and Table 3).

(v) Fine-tuning control and Cosmos 2.5 (uEqr, Q3+Q4). Averaged over four datasets, matched Orbis v1 fine-tuning without geometry worsens long-horizon FVD from 539.4 to 573.8. With geometry, FVD falls to 325.0 (Table R6, Table 2). After we lowered the guidance weight and fixed the gate initialization, the retrained Cosmos 2.5 variant improves long-horizon FVD over its base model from 503.9 to 455.1 (Table R7, Tables 2 and 15).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.