acceptodds
Under review as a conference paper at ICLR 2027

World in World: Explore the World with World Models

Abstract

Recent advances in video world models enable interactive generation beyond fixed-length synthesis, yet extending pretrained models to new forms of control typically requires control-specific architectures or additional training. We introduce World in World (WiW), a training-free visual-evidence interface that flexibly extends the controllability of frozen causal video world models. WiW leverages native self-attention, which already consumes clean visual states, as a shared interface for heterogeneous external controls without modifying the pretrained backbone. It converts source observations, target-view projections, rendered geometry, and generated history into clean visual evidence annotated with camera, temporal, and spatial-validity information. These sources provide appearance references, spatial guidance, completion cues for newly visible regions, and long-range context for revisiting generated states. We introduce correspondence-guided attention routing (CGAR) to direct queries toward geometrically corresponding visual tokens, and evidence-wise attention CFG (EWA) to regulate the influence of different evidence sources within native self-attention without additional network function evaluations for guidance. We instantiate WiW on a frozen LingBot-World 2.0 backbone and evaluate camera-controlled video rerendering on DAVIS and OpenVid-1M across diverse viewpoint changes. Beyond rerendering, WiW enables exploration within a given video's dynamic world and supports bullet-time generation, video stabilization, video editing, and cross-generation K/V sharing, demonstrating a unified and versatile framework for training-free world-model control.

Primary area:
generative models
Keywords:
Video world model, Camera control, 4D generation, Training-free

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DkGEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

World in World proposes a training-free method to make full use of pretrained world models’ visual and motion prior, extending its controllability. It introduces a unified visual evidence interface to guide generation through native self-attention, with additional mechanisms to improve spatial alignment and control the influence of different evidence.

Critical strengths

WiW solved a very interesting problem. For world models, visual evidence comes from diverse sources. Most existing methods inject new visual evidence in a modular way, but this is obviously not unified enough and not elegant enough. WiW proposes a clever idea: the causal self-attention of the existing video world model itself can read the historical clean visual states. Then, other visual evidence can also be transformed into a similar form and injected into the model through the existing attention mechanism. In this way, there is no need to design and train new control modules for different visual evidence, and at the same time, the existing prior knowledge of the pre-trained model can be fully utilized. I think this idea is very concise and elegant, and it's also training-free. This is my favorite part of this paper.

In addition, the author also conducted relatively thorough experiments. On DAVIS and OpenVid-1M, WiW has achieved good results in both video quality and camera control accuracy, especially in the case of significant changes in viewing angles, it can still generate relatively consistent videos. The article also showcases different applications such as bullet time and video editing, which to some extent demonstrates the flexibility of this unified interface. I think these experiments have demonstrated the potential of this idea quite well.

Critical weaknesses

In my opinion, one of the main problems with WIW is its reliance on external geometric estimates. While the idea of injecting visual evidence with self-attention is elegant, practical implementation still requires additional depth estimation, camera pose estimation, point tracking, and even human geometric reconstruction. This weakens the simplicity of the method to some extent, and also makes the whole pipeline heavy. More importantly, the performance of the method seems to depend heavily on this geometric information. As you can see from Table 2, there is a significant performance degradation with target-view warping removed, but a relatively small performance degradation with CGAR removed. This makes me wonder how much of the improvement comes from the author's attention mechanism and how much from the strong constraints provided by the external geometric information?

In addition, I am concerned about the robustness of the method to geometric estimation errors. Geometric estimation may be inaccurate under large-scale view changes or complex dynamic scenarios, and these erroneous geometric information will be injected into the model as visual evidence, which may affect the subsequent generation and long-range consistency. Although the authors compare different depth estimators in the Appendix and also present a 220-frame long-horizon revisit, it is mainly a qualitative experiment. It would be more convincing if some quantitative experiments on geometric estimation errors and the stability of long-range rollout could be added.

Questions for authors

I wonder whether the errors introduced by the external geometric estimation module will gradually accumulate in the long-term rollout, thereby leading to increasingly obvious scene drift. Although the author presented the qualitative results of 220 frames, I would prefer to see some systematic quantitative analyses, such as how the geometric and visual consistencies change as the rollout length increases, especially in cases where there are estimation errors in depth, camera pose, or point tracking.

3 comments
AuthorsOpenReview ↗

Response to Reviewer DkGE

We sincerely thank Reviewer DkGE for recognizing our unified evidence interface.

1. External geometry and overhead. We agree WiW uses external geometry, common in rerendering: geometry provides the spatial reference needed to project observed source content and its time-varying motion into new camera views. Notably, WiW supports diverse visual evidence via the same frozen attention interface, selectively used by task rather than requiring all geometry at once (e.g., human geometry is mainly for 180°/360° rotations to guide unseen surfaces). For efficiency, Appendix A reports that for matched frames, depth/pose estimation (0.084 s/frame) and point tracking (0.098 s/frame) total just 13.1% of WiW's inference time (1.389 s/frame, Table A2). Our newly measured LHM++ and SMPL-X take 0.108 and 0.049 s/frame/person (by subject count). Preprocessing runs once per source video and is reused across target paths, adding only modest one-time overhead.

2. Contribution of geometry and attention. Target-view warping is indeed vital, providing pixel-level references for how the scene evolves over time under the target camera; without it, generation becomes unconstrained (Fig. A3(a)). Besides providing the interface for warped videos to enter the frozen model, WiW improves how the model reads and uses this evidence. With identical geometry, EWA boosts PSNR from 21.52 to 22.98 dB on DAVIS (Table 2) and from 19.59 to 20.60 dB under large rotations (Table A1). CGAR routes attention to corresponding unwarped source features so generated subjects follow source appearance and pose at that event time. To isolate this, we added ablations with identical source K/V against ordinary attention and routing-only (Table R1 in Common Response), and tested 50 high-motion pairs (Table R2 in Common Response), showing larger CGAR gains. These results separate geometry's spatial constraints from the extra gains of our attention mechanisms.

3. Robustness to geometric errors and long-rollout stability. To test sensitivity to depth estimators, we first compared DepthCrafter, ViGeo, and VDA on the 200 DAVIS pairs with fixed references (Table R7), verifying WiW's robustness. For long rollouts, depth, poses, and tracks are preprocessed once over the full source video rather than re-estimated chunk by chunk; thus, while these priors may have estimation errors, this largely avoids autoregressive error accumulation in the control signals. To test robustness to imperfect geometry and if errors still propagate via autoregressive generation, we built 50 long sequences (180-220 frames: rotation, translation, orbit) from 25 OpenVid-1M videos. We perturbed depth (randomly dropping 5%/10% of values), camera poses (randomly picking 5%/10% of poses for a 10% multiplicative rotation/translation perturbation), and point tracks (randomly dropping 5%/10% of tracks) and compared WiW with streaming 4D InSpatio-World (ISW) under matched settings. Table R8 shows WiW degrades more gradually as errors increase and outperforms ISW in visual fidelity and camera control. We also measured fidelity and camera errors at early (T₁), mid (T₂), and late (T₃) rollout stages with clean and perturbed geometry against ISW (Table R9). In Tables R8 and R9, even with input errors, WiW keeps more stable visual and camera consistency over long sequences, resisting autoregressive error accumulation. We hope these new results fully address your concerns.

Table R7. Depth estimators on 200 DAVIS pairs.

MethodVBench↑Rot(°)↓Trans↓PSNR↑SSIM↑LPIPS↓
DepthCrafter83.90581.78560.055723.020.7230.136
ViGeo83.89711.79810.056122.560.7200.138
VDA (default)83.90421.78230.053322.980.7270.133

Table R8. Robustness under geometric perturbations.

Perturb.MethodPSNR↑LPIPS↓SSIM↑Rot(°)↓Trans↓
NoneOurs24.390.03610.88741.1040.0075
NoneISW22.530.06410.80681.3900.0081
Depth 5%Ours22.980.04650.85691.1040.0083
Depth 10%Ours21.540.06140.81031.3120.0089
Depth 10%ISW18.530.12490.76841.7050.0103
Pose 5%Ours24.200.03670.88701.1080.0087
Pose 10%Ours23.970.03790.87521.3110.0098
Pose 10%ISW22.010.07010.78711.6330.0121
Track 5%Ours24.360.03670.88611.1100.0079
Track 10%Ours24.300.03710.88551.1100.0078

Table R9. Long-rollout consistency across stages T₁-T₃.

MethodT₁ LPIPS↓T₁ Rot(°)↓T₂ LPIPS↓T₂ Rot(°)↓T₃ LPIPS↓T₃ Rot(°)↓
Ours0.03600.6970.03560.9680.03681.646
ISW0.05250.7060.06111.1930.07852.272
Ours (Depth 10%)0.04660.7380.05821.1530.07932.045
ISW (Depth 10%)0.08141.0110.12211.5100.17142.593
Ours (Pose 10%)0.03660.6980.03721.1130.03992.121
ISW (Pose 10%)0.05830.8140.06801.3670.08392.719
Ours (Track 10%)0.03630.6970.03580.9800.03731.653

Thank you for your detailed reply. The new results have resolved most of my concerns. I will keep my score.

AuthorsOpenReview ↗

Thank You for Your Constructive Feedback

We greatly appreciate the time and effort you have invested in carefully reviewing our work. We sincerely appreciate your recognition of the novelty of our work and your constructive suggestions, which have helped us improve the paper. We are glad that our additional results have addressed most of your concerns. Please feel free to let us know if you have any further questions or concerns.

Reviewer_2qifOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

World in World (WiW) introduces a training-free visual-evidence interface for controlling a frozen LingBot-World 2.0 backbone, integrating source observations, projected views, rendered geometry, and generation history through clean K/V features. Its CGAR and EWA mechanisms support geometric evidence routing and evidence weighting, with experiments primarily demonstrating improvements in camera-controlled video rerendering.

Critical strengths
  • This paper proposed an unified, training-free control. The shared evidence interface integrates heterogeneous visual conditions and historical context without backbone fine-tuning or learned control-specific adapters.

  • Comparisons against six baselines and same-backbone ablations support the system’s practical utility, with the complete system improving PSNR from 20.92 to 22.98 over target-view warping alone.

Critical weaknesses

W1: The experiments do not sufficiently substantiate the claimed capability for persistent exploration of dynamic worlds

Quantitative evaluation primarily measures source-conditioned rerendering, with reconstruction metrics excluding newly exposed regions. Qualitative revisits demonstrate static appearance preservation but do not establish consistency as events and scene states evolve. A systematic evaluation of dynamic revisits is needed to substantiate this central claim.

W2: The incremental benefits of the specific CGAR and EWA designs remain insufficiently established

The CGAR ablation removes the dedicated unwarped source-video branch (§4.3), thereby evaluating the contribution of the entire branch rather than isolating the effect of geometric routing. It therefore remains unclear whether the routing mechanism in §3.5 improves upon ordinary attention to the same source K/V features when all other settings are held fixed.

EWA yields substantial improvements over the configuration without additional guidance (Tables 2 and A1). However, this ablation sets all guidance strengths to zero (Appendix E, “Ablation settings”), leaving EWA’s advantage over simpler, comparably tuned per-channel evidence amplification untested. A controlled comparison using the same evidence, schedule, and output-norm control would help isolate the benefit of EWA’s specific correction.

Since CGAR and EWA are presented as primary technical contributions (§1, third contribution), these missing comparisons limit the evidence for their incremental value over simpler mechanisms using the same visual evidence.

Questions for authors

pls refer to weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2qif

We sincerely thank Reviewer 2qif for the constructive feedback.

1. Quantitative evaluation of dynamic revisits. We agree our prior revisit evaluation focused on static structures. Exploring a dynamic scene on an out-and-back path as event time advances requires revisit consistency in both newly generated (source-unobserved) and source-observed regions. In newly generated regions, historical K/V retrieval preserves synthesized scene appearance and layout beyond the rolling cache. In source-observed regions, aligned warps and CGAR-routed source K/V provide current-time references; we disable historical retrieval here via visibility masks so earlier states do not interfere with ongoing motion, while rendered geometry and EWA help guide newly exposed surfaces of moving subjects under large rotations (Table A1). As noted in Appendix C, while our current memory focuses on preserving expanded scene structures rather than simulating independent off-screen motion of newly synthesized objects (planned via coding-agent proxies), combining historical scene retrieval with current-time source evidence keeps revisits consistent with both the expanded scene and the evolving recorded event. To test revisit consistency in both regions on the same rollouts, we built 50 out-and-back paths (150-220 frames, 2 revisits, advancing event time) on 20 OpenVid-1M videos. In newly generated regions (Table R4), using initial-visit generated frames as the reference (GT) for the 1st and 2nd revisits, WiW achieves better consistency than No Retrieval and adapted WorldKV on the same backbone. At revisited source poses in source-observed regions (Table R5), using the corresponding source frames at the revisited event times as GT, WiW also outperforms streaming 4D InSpatio-World, confirming that WiW preserves expanded scene layout across revisits while tracking evolving source motion.

Table R4. Revisits in newly generated regions vs. initial-visit GT (₁, ₂: 1st and 2nd revisits).

MethodPSNR₁↑SSIM₁↑LPIPS₁↓PSNR₂↑SSIM₂↑LPIPS₂↓
No Retrieval14.350.5860.30914.570.5420.320
WorldKV (adapted)15.540.5950.29715.090.5560.290
WiW16.020.6030.23215.580.5650.252

Table R5. Revisits at source poses vs. current-time source frames (source-observed).

MethodPSNR↑SSIM↑LPIPS↓
InSpatio-World22.080.6960.102
WiW24.490.8410.049

2. Isolating CGAR's contribution. We agree "w/o CGAR" did not isolate routing and renamed it "w/o Source Branch". To test CGAR with identical source K/V and fixed settings, we compare ordinary attention, routing without logit bias, and full CGAR (Table R1 in Common Response). Ordinary source attention performs worse than removing the source branch: under object and camera motion, unguided attention reads unrelated source regions, adding wrong details. Routing avoids mismatches, but without logit bias, matched features get too little weight during joint normalization to guide generation; full CGAR resolves both. This especially benefits dynamic subjects by following their source appearance and pose at the corresponding time (Fig. A4(a)), as verified on 50 high-motion DAVIS/OpenVid-1M pairs (e.g., dancing, cycling, running; Table R2 in Common Response).

3. EWA versus simpler evidence amplification. We agree comparing with equally tuned alternatives is important. To test EWA against simpler amplification, we ran 6 controlled tests (Table R6): no amplification, constant logit bias, standard attention CFG, EWA without cosine gating or projection, and full EWA. All share identical evidence, active steps, sampling, and output-norm control, with each method's parameter ("Param") independently tuned under the same budget for best overall performance. Simple bias and attention CFG boost evidence weights or conditioned-vs.-native differences directly, strengthening guidance but also amplifying redundant or conflicting signals. EWA instead uses projection to isolate information complementary to the native response and cosine gating to suppress misaligned updates.

Table R6. Comparison of evidence amplification.

GuidanceParamVBench↑RotErr(°)↓TransErr↓PSNR↑SSIM↑LPIPS↓
w/o EWA083.88081.99440.056621.520.6690.156
Simple Attn Bias483.90311.91240.060122.070.6830.149
Native Attn CFG0.383.86072.07330.068921.380.6450.169
EWA (w/o cos gate max⁡(0,cos⁡(∗))\max(0,\cos(\ast)) )0.583.88531.80230.053322.580.7010.137
EWA (w/o proj Projo\mathrm{Proj}_{o})0.583.89011.83170.054722.120.6890.139
EWA (Full)0.583.90421.78230.053322.980.7270.133

In Table R6, component ablations verify each design, with full EWA achieving the best performance across all visual fidelity, camera control, and quality metrics. We hope these new results fully address your concerns.

Thank authors for the detailed rebuttal and additional experiments. The new results substantially address my concerns. I will raise my score.

AuthorsOpenReview ↗

Thank You for Your Positive Feedback and Encouragement

Thank you very much for your thoughtful and professional review, and for the time and effort you have devoted to evaluating our work. We sincerely appreciate your recognition of our rebuttal and additional experiments, as well as your encouraging reassessment of our work. Please feel free to let us know if you have any further questions or concerns.

Reviewer_mSh6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents World in World, a trainning-free framework to convert rendered video to new camera trajectories. The method reuses the model’s existing self-attention layers to incorporate clean K/V features from the source video, projected views, geometry renderings, and retrieved past outputs. Correspondence-guided attention routing (CGAR) directs queries toward matching source tokens, while evidence-wise attention CFG (EWA) regulates each evidence channel’s influence. The method is implemented on Lingbot-World-2.0, providing competitive results on DAVIS and OpenVid-1M. And some applications like bullet-time generation is quite interesting.

Critical strengths
  1. WiW extends a world model backbone to many new capabilities without finetuning, which is efficient and elegant. Without training, WiW still achieves impressive results.
  2. The experiments are well designed. The rerendering results on DAVIS and OpenVid-1M against baselines are competitive. And for ablation study, removing EWA reduces DAVIS PSNR from 22.98 to 21.52, supporting its contribution to evidence-guided generation.
  3. Qualitative demonstrations cover bullet-time generation, stabilization, editing, and scene revisits, illustrating the method's practical flexibility and extensibility.
Critical weaknesses
  1. The biggest concern is about difference between WiW and Warp-as-history. Warp-as-History already demonstrates training-free camera control by supplying temporally aligned, visibility-filtered warped observations through a frozen model’s native visual-history pathway. WiW incrementally adding source-video routing, geometry evidence, retrieved history, and evidence-wise guidance. The paper should more clearly establish the substantive advances provided by these additions.
  2. A direct comparison with Warp-as-History is also needed. By using the same backbone and matched rerendering inputs, WiW should be compared with Warp-as-History's zero-shot variant. I understand the “Only Target-View Warp” ablation is useful, but direct comparison is more preferable.
  3. Removing CGAR also removes the entire source-video attention branch. Retaining identical source K/V while comparing ordinary attention against correspondence-guided routing would better establish the routing mechanism’s value.
Questions for authors

Please address my concerns between WiW and Warp-as-History, and the CGAR ablation study. I would raise my score if these two biggest concerns are well solved.

3 comments
AuthorsOpenReview ↗

Response to Reviewer mSh6

We sincerely thank Reviewer mSh6 for the positive feedback and insightful comments on WaH and CGAR.

1. Difference from Warp-as-History (WaH). We agree that our warping branch shares the same basic design as WaH: injecting temporally aligned, visibility-masked warps through the native history pathway. WiW extends this idea to a broader challenge: exploring an evolving world captured in a source video from new viewpoints while preserving the original event's appearance and motion. Warped observations alone provide limited information, especially under large viewpoint changes, occlusions, and complex motion. Our focus is therefore on how to effectively combine complementary visual evidence, including unwarped source observations, geometric priors, and generated history, to guide the generation of continuously changing novel views. Simply injecting these cues through the native history pathway is insufficient, as different evidence sources can compete for attention and queries may attend to incorrect source regions. To address these challenges, we introduce EWA to regulate the influence of different evidence channels through attention-response guidance (Table 2, Table A1), and CGAR to route queries toward geometrically corresponding source features to preserve dynamic appearance (Tables R1–R2 in Common Response). In addition, rendered geometry provides priors for newly exposed surfaces under large viewpoint changes (Table A1), while historical K/V retrieval helps preserve previously generated scene structure across revisits. Together, these designs enable the frozen model to integrate complementary visual cues beyond aligned warps for flexible, consistent world exploration without training.

2. Direct comparison on the same backbone. We agree a direct comparison is valuable and apologize for not making our baseline clearer. "Only Target-View Warp" (Table 2) already adapts WaH's core zero-shot approach to LingBot-World 2.0, injecting target-view warps via the native history pathway with aligned RoPE and visibility filtering. For a more direct comparison, we also reproduced WaH's original zero-shot setting on the same backbone ("WaH Zero-Shot Port"): following its official settings: using the first source frame's mean RGB to fill invisible regions in the warped views and removing invisible tokens based on a confidence threshold. In our existing "Only Target-View Warp" implementation, we use boundary-pixel dilation to fill invisible holes and softly weight tokens according to their visible-pixel ratios. Testing different schedules showed guidance at denoising steps 0, 1, and 2 works best, matching our default. Using the same 200 DAVIS pairs, backbone, inputs, target cameras, and protocol (Table R3), both warp-only variants underperform WiW in fidelity and camera control.

Table R3. Quantitative comparison with zero-shot Warp-as-History on the same backbone.

MethodVBench Mean↑RotError(°)↓TransError↓PSNR↑SSIM↑LPIPS↓
Only Target-View Warp83.81622.27810.087920.920.6010.179
WaH Zero-Shot Port83.58342.27960.086520.510.5980.184
WiW (Ours)83.90421.78230.053322.980.7270.133

3. Isolating CGAR's contribution. Thank you for this suggestion. We renamed "w/o CGAR" in Table 2 to "w/o Source Branch". To test whether CGAR improves upon ordinary attention under identical source K/V (other settings fixed), we added two ablations (Table R1 in Common Response): (i) ordinary attention without routing or bias ("w/ Source + Ordinary Attention"), and (ii) routing without logit bias ("w/ CGAR (w/o Logit Bias)"). Ordinary source attention performs worse than removing the source branch: without correspondence guidance, object motion causes queries to attend to unrelated source regions, adding mismatched details. Routing without bias avoids mismatches, but matched features get too little weight relative to other evidence blocks during joint normalization. Full CGAR resolves both by directing attention to valid matches and strengthening their influence. As shown in Fig. A4(a), CGAR is more effective for highly dynamic subjects: when fast motion degrades depth-warped inputs, routing provides direct appearance and pose references from the unwarped source video at the corresponding time. Evaluating 50 high-motion DAVIS/OpenVid-1M pairs (e.g., dancing, cycling, running; Table R2 in Common Response) confirms larger gains from CGAR on these challenging cases, especially in preserving source appearance. We hope these new experiments and clarifications fully address your concerns.

AuthorsOpenReview ↗

Thank You for Your Valuable Comments and Suggestions

We are grateful for your thorough and insightful review, as well as the time and effort you have devoted to evaluating our work. We have carefully addressed your concerns in our rebuttal and incorporated your constructive suggestions into the revised manuscript, including an expanded discussion of the differences between WiW and Warp-as-History in the related work. We hope our responses address your concerns. If you have any further questions or suggestions, please let us know. We would be happy to provide additional clarification.

Reviewer_mSh6OpenReview ↗

Satisfying rebuttal

Thank you for the clarifications and additional experiments. The same-backbone comparison with WaH and the controlled CGAR ablations largely address my concerns, so I am raising my rating to weak accept.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank the Area Chair and all reviewers (mSh6, 2qif, and DkGE) for their time and constructive feedback. We are encouraged that the reviewers recognized the strengths of our work, including our unified and training-free visual-evidence interface, strong quantitative performance across benchmarks, and flexible applications such as bullet-time generation, stabilization, video editing, and scene revisits.

We have provided point-by-point responses to each reviewer's comments below and updated the manuscript accordingly, with all revisions highlighted in blue.

Since all three reviewers (mSh6, 2qif, and DkGE) raised a shared question regarding isolating the contribution of Correspondence-Guided Attention Routing (CGAR) from the unwarped source-video attention branch, we conducted two controlled ablations with identical source K/V features (holding all other settings fixed): Table R1 tests whether CGAR's geometric routing and logit bias improve upon ordinary source attention on the same 200 DAVIS video–trajectory pairs used in the paper ablation study, and Table R2 further evaluates CGAR's gains on 50 challenging high-motion video–trajectory pairs from DAVIS and OpenVid-1M (e.g., dancing, cycling, and running) where fast motion degrades depth-warped inputs.

Table R1. Quantitative ablation of CGAR and the source-video attention branch on DAVIS.

VariantVBench Mean↑RotError(°)↓TransError↓PSNR↑SSIM↑LPIPS↓
w/o Source Branch (original w/o CGAR)83.89181.78280.053922.880.7200.138
w/ Source + Ordinary Attention83.70111.89120.068321.930.6980.166
w/ CGAR (w/o Logit Bias)83.88931.78280.054222.900.7170.140
w/ CGAR (Full)83.90421.78230.053322.980.7270.133

Table R2. Quantitative ablation of CGAR on 50 high-motion video-trajectory pairs.

VariantVBench Mean↑RotError(°)↓TransError↓PSNR↑SSIM↑LPIPS↓
w/o Source Branch (original w/o CGAR)83.68032.52130.095721.1900.6790.169
w/ Source + Ordinary Attention83.51343.18250.108620.4470.6430.183
w/ CGAR (w/o Logit Bias)83.67912.53030.090921.2850.6840.165
w/ CGAR (Full)83.69172.49210.089721.4830.6970.158

Together, Tables R1 and R2 show that naively attending to unwarped source K/V via ordinary attention degrades performance due to motion-induced mismatches, whereas full CGAR (combining geometric routing with logit bias) consistently improves both visual fidelity and camera control, with even larger gains on high-motion subjects. Detailed explanations and analyses are provided in our individual responses to each reviewer below, and we thank the Area Chair and reviewers again for their valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.