acceptodds
Under review as a conference paper at ICLR 2027

Geometry as Address: Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation

Abstract

Long-horizon camera-controlled video generation requires recovering previously observed content from an ever-growing visual history. Existing approaches either search historical context implicitly or reconstruct it into persistent 3D memory, facing inefficient memory access or accumulated geometric errors. Our key insight is that geometry need not explain the scene—it only needs to determine *where visual memory should be read from, while attention decides what should be recovered*. Based on this insight, we introduce GEAR, a Geometry-Enabled Attention Routing framework that uses geometry as an explicit token-level address for visual memory. Rather than fusing historical observations into a persistent global 3D representation, GEAR retains them as frame latents and uses per-frame geometry only to establish token-level correspondences with target views, thereby avoiding persistent error accumulation from global fusion. Guided by these correspondences, Geometric Correspondence Attention (GCA) selectively injects geometrically matched historical features into noisy target patches during denoising. We further introduce an Invisible Octree to accumulate visibility evidence and reject geometrically plausible but occluded correspondences. Extensive experiments demonstrate that GEAR achieves state-of-the-art visual quality, precise camera control, and revisit consistency, enabling minute-long video generation along challenging trajectories. Additional videos are available at https://anonymousgear46.github.io/geometry-as-address/.

Primary area:
generative models
Keywords:
camera-controlled video generation, long-horizon video generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_RiBPOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GEAR, a Geometry-Enabled Attention Routing framework for long-horizon camera-controlled video generation. Experiments demonstrate state-of-the-art visual quality, precise camera control, and revisit consistency, enabling minute-long video generation along challenging user-specified camera trajectories.

Critical strengths
  1. I like GEAR’s idea of decoupling visual memory from geometric addressing. I believe this insight contributes somewhat to the community.

  2. Extensive experiments demonstrate that GEAR achieves state-of-the-art visual quality.

Critical weaknesses
  1. The paper claims that GEAR avoids persistent 3D memory. However, the model still constructs local meshes from per-frame depth estimates and camera poses, while continuously maintaining a global Invisible Octree and a global visible mesh. Their errors may still accumulate over time and affect subsequent generations. Could the authors clarify how GEAR differs from explicit 3D-memory approaches in this respect? Furthermore, I recommend authors provide some experiments to verify the accumulation errors between two variants. Among them, one variant is the GEAR using per-frame geometry for memory addressing. Another variant is that GEAR explicitly fuses historical observations into a persistent global 3D representation and uses its reprojection as the generation condition.

  2. In table.1, the long-horizon experiments only are conducted in 50 scenes from the WorldScore Static set, which are insufficient to evaluate each method's performance.

Questions for authors

Please see the critical weaknesses.

1 comment

We sincerely thank the reviewer for the positive and encouraging feedback. We appreciate the recognition of our core insight and experimental results. We provide detailed responses to each of the reviewer’s comments below.


Q1: Clarification on Persistent 3D Memory and Error Accumulation.

We thank the reviewer for this insightful observation. We clarify that GEAR does maintain persistent geometric structures, including an Invisible Octree and an auxiliary global visible mesh, but these structures serve only as visibility proxies rather than appearance-bearing 3D memories. They do not store visual features or provide rendered appearance conditions for generation.

(1) Geometric Representation and Error Propagation. Importantly, GEAR constructs history-to-target correspondences exclusively from independent per-frame local geometry, rather than from a globally fused mesh (Sec.3.2 and Page.Method). This design avoids consolidating cross-view geometric inconsistencies into persistent artifacts that would otherwise propagate through reprojected appearance conditions. We have added a supplementary visualization (supp/review-RiBP/local_correspondence.mp4 and our webpage) illustrating how the history-to-target patch correspondence cache is constructed from local historical geometry.

While inaccuracies in per-frame geometry may introduce erroneous correspondence candidates, errors in the global visibility proxies affect only correspondence filtering. Neither source of error directly corrupts the underlying visual memory, as visual content remains stored in independent frame latents. Moreover, GCA adaptively aggregates visual evidence from multiple historical observations, mitigating the impact of erroneous correspondences when valid alternatives remain available, as demonstrated by our correspondence-perturbation ablation (Fig. 11 and Table 2).

(2) Warped-conditioning Variant. We additionally trained a warped-conditioning variant of GEAR, following the global 3D fusion and reprojection pipeline of ViewCrafter [1], under matched training settings. As shown in the updated supplementary video (supp/review-RiBP/warp_vs_gca.mp4) and our updated webpage, cross-view geometric inconsistencies introduce persistent 3D floaters after global fusion. These artifacts are subsequently reprojected as noisy pixel-aligned conditions and degrade generation quality. In contrast, GEAR uses the same per-frame geometric inputs and independently retrieves historical features through correspondence attention, avoiding the direct propagation of fused geometric artifacts into appearance conditioning.

We also visualize GEAR's global Invisible Octree and visible mesh to clarify their roles as coarse visibility proxies for identifying unobserved regions, which are subsequently updated using newly generated frames, rather than as appearance-bearing memory representations.

Reference:
[1] ViewCrafter: Taming Video Diffusion Models for High-fidelity Novel View Synthesis


Q2: Extended Evaluation on WorldScore-Static.

We appreciate the reviewer's concern regarding evaluation scale. We have expanded the WorldScore-Static evaluation to 100 randomly sampled scenes with closed-loop camera trajectories, and updated the quantitative results in our paper and added more qualitative comparisons to our project page. GEAR consistently achieves the best content alignment, photometric consistency, style consistency, and revisit performance.

Reviewer_oN6pOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GEAR, a framework for long-horizon camera-controlled video generation from a single image. Built on Wan2.1-I2V-14B, GEAR keeps retrieved history frames as context latents, uses per-frame depth to build token-level correspondences between history and target views, and introduces Geometric Correspondence Attention (GCA) to inject the matched history features into the target tokens. An Invisible Octree is further used to filter occluded correspondences. Experiments on DL3DV and WorldScore-Static report improvements in visual quality, camera control, and revisit consistency.

Critical strengths
  • This paper addresses an important problem. Long-horizon camera-controlled generation with consistent revisits is still challenging for video world models.
  • The high-level idea is easy to understand, and GCA is lightweight.
  • paper overall well written with clear figures. The project page and visualizations are also well done.
Critical weaknesses
  • My main concern is about novelty. Method feels more like a combination of existing memory designs rather than a new idea. The modeling basically follows context-as-memory (temporal concat of history frames), and GCA is added as an extra correction branch on top. The coverage based retrieval over local per-frame memories is quite similar to anchorweave [1]. Also mosaicmem [2] is missing, which also lifts memory patches into 3D and reads them by attention, very close to the “geometry as address” idea. GCA itself may have some novelty, but overall the method reads more like a combination of these works.

  • I couldn’t find any explicit camera conditioning in the architecture. Looks neither Fig. 3 nor the method section/appendix shows how camera pose goes into the network, and GCA looks doesn’t take pose either. The correspondences are computed with cameras, but the injected features themselves don’t carry any pose info, so it seems hard for the model to learn how an injected patch relates to where it’s placed. So I am not sure how camera control actually works, especially for newly explored regions without correspondences. This might also be why the w/o GCA variant in Tab. 2 is worse than all baselines. A clarification would be helpful.

  • The ablation doesn’t compare GCA with other ways of injecting the same memory, e.g., warping the retrieved history to the target view and injecting it as control signal, like spmem [3] and anchorweave [1]. These are only compared as full systems with different backbones and data, so its hard to tell whether the gains really come from GCA

[1] AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories. ECCV 2026. [2] MosaicMem: Hybrid Spatial Memory for Controllable Video World Models. arXiv 2026. [3] Video World Models with Long-term Spatial Memory. NeurIPS 2025.

Questions for authors

See weaknesses. I am open to discussion + reconsider my score based on the author rebuttal.

4 comments

We thank the reviewer for the constructive feedback and recognition of our lightweight design and clear presentation.


Q1: Novelty and Distinction from Prior Designs.

We acknowledge that temporal concatenation and coverage-based history retrieval are established strategies. GEAR's novelty lies in using geometry to govern token-level access to visual memory, rather than combining existing memory designs. We distinguish GEAR from three existing paradigms:

  • Implicit frame-level memory (Context-as-Memory) relies on dense attention over retrieved historical frames to implicitly identify relevant evidence, introducing redundant interactions with unrelated context. GEAR instead restricts each target token to spatially corresponding historical tokens through geometric addressing.
  • Projection-based geometric conditioning (AnchorWeave [1], SPMem [3]) constructs appearance conditions by projecting geometric memories into target views. SPMem fuses historical observations into a global point cloud, causing cross-view geometric inconsistencies to accumulate and propagate through rendered conditions. AnchorWeave similarly avoids global fusion by maintaining local geometry, but still relies on appearance reprojection and a learned controller to reconcile inconsistencies. GEAR avoids both appearance fusion and warping, preserving historical visual features in their original latents while using geometry only for memory addressing.
  • Geometry-aligned latent conditioning (MosaicMem [2]) also retrieves memory patches using 3D geometry and constructs spatially aligned conditioning through latent warping, without explicitly constraining query-specific attention connectivity. GEAR instead converts geometric correspondences into sparse, query-specific attention connections, enabling adaptive aggregation of multiple historical observations without intermediate warped appearance conditions.

Our central novelty is using geometry to define which historical visual tokens each target query can attend to, enabling fine-grained and adaptive memory retrieval without intermediate appearance warping, rather than what appearance conditions to construct. Furthermore, we introduce an Invisible Octree that improves retrieval reliability by filtering geometrically projectable but occluded correspondences.

We will clarify these distinctions in the revised related work.


Q2: Camera Conditioning without Explicit Pose Embeddings.

We emphasize that camera information is conveyed through pose-dependent geometric correspondences which determine GCA's attention scope, rather than explicit pose embeddings.

(1) Observed regions. Prescribed camera poses determine history-to-target correspondences via local geometric projection, specifying which historical keys and values each target query can access. Although the historical features do not explicitly encode camera poses, GCA injects them directly into geometrically corresponding target tokens. This provides explicit spatial grounding for faithful reconstruction in previously observed regions.
(2) Unobserved regions. For newly explored regions without valid correspondences, GCA contributes no direct historical features. The pretrained backbone instead synthesizes plausible content through its learned generative prior and dense spatiotemporal attention to neighboring geometrically grounded regions.
(3) Experimental validation. The w/o GCA ablation demonstrates that camera control and memory consistency do not emerge from training alone. To further assess the necessity of explicit pose embeddings, we train an additional camera adapter alongside GCA following CameraCtrl. Explicit pose conditioning yields no additional benefit, validating the effectiveness of GCA's spatial guidance without additional pose-conditioning modules. Qualitative results are provided in supp/review-oN6p/camera_condition.mp4 and on our webpage.

VariantRevisit SSIM ↑ / LPIPS ↓ATE ↓
GEAR0.6489/0.20190.0436
+ Cam.0.6437/0.20240.0441

Q3: Ablation with Warped Conditions.

Under matched backbone and training settings, we replace GCA with channel-wise concatenated point-cloud renderings obtained through global fusion and target-view reprojection, following ViewCrafter.

VariantSSIM ↑LPIPS ↓ATE ↓
GEAR (GCA)0.36450.44590.0436
GEAR (Warp Cond.)0.30570.54920.2438

GCA substantially outperforms warped conditioning in visual fidelity and camera accuracy. Qualitative comparisons further reveal that noisy reprojections caused by cross-view geometric inconsistencies degrade subsequent generations, as shown in supp/review-oN6p/ablation_warped_condition.mp4 and on our webpage.

Thank you again for your time and feedback. We would appreciate hearing whether our clarifications and expanded comparisons address your concerns, and would be happy to discuss any remaining questions.

Reviewer_oN6pOpenReview ↗

Thanks for the rebuttal

Thank you for the response. The new experiments address my concerns on camera control and the comparison with warped conditions. On novelty side, I still feel the main claimed novelty using geometry to define sparse attention connections is reasonable but somewhat incremental, and the other components are not very novel, though they bring some system and engineering improvements (e.g. Invisible Octree). On the other hands this work is at least complete with good writing overall. My assessment is slightly above a borderline case but not quite at weak accept (mainly the novelty concern). I have raised my score to weak accept, since there is no borderline accept option, but my overall feeling is closer to borderline accept.

We sincerely thank the reviewer for the constructive feedback and are glad that our additional experiments have addressed the concerns regarding camera control and warped conditioning. Regarding novelty, we would like to further clarify the motivation behind our approach.

Precise camera control and persistent revisit memory remain challenging in long-horizon video generation. While explicit geometry offers an effective means of addressing these challenges, existing approaches predominantly exploit geometry through warping to construct appearance conditions, where geometric inconsistencies can accumulate through global fusion and propagate into subsequent generations, as demonstrated in our additional experiments.

Our work approaches this problem from the intrinsic attention mechanism of video generation models. Ideally, a video generation model should be able to retrieve geometrically corresponding historical tokens through dense attention to achieve precise camera control and consistent revisitation. In practice, however, learning such spatially precise retrieval amid numerous irrelevant token interactions remains challenging, as evidenced by the limitations of recent SOTA implicit methods, such as Lingbot-World and HY-WorldPlay, in our experiments.

This motivates us to ask: Given explicit geometry, why not directly constrain where attention retrieves visual evidence, rather than constructing additional appearance conditions or relying on dense attention to implicitly discover geometric correspondences? By restricting memory attention to geometrically corresponding historical tokens, GEAR preserves the backbone's inherent contextual modeling capabilities while explicitly directing memory retrieval toward the most spatially relevant historical evidence.

Therefore, our core contribution lies not in sparse attention itself, but in rethinking the role of geometry in video generation: from constructing explicit appearance conditions to directly governing token-level access to visual memory. We believe this perspective offers a meaningful methodological insight into long-horizon video generation beyond the individual system components.

We would be pleased to provide any further clarification if needed. We sincerely appreciate the reviewer's constructive feedback and support again!

Reviewer_Kt4xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a work for the long-horizon camera-controlled video generation. Rather than fusing history into a global 3D model or attending densely over all history tokens, GEAR keeps history as frame latents, uses per-frame estimated depth to build token-level correspondences, and lets each target token cross-attend only to its matched history tokens. The proposed approach is built on frozen Wan2.1-14B.

Critical strengths
  1. I like the idea of this paper and build long-term memory is a critical challenge for long video generation
  2. The paper is well organized
  3. The motivation is good.
Critical weaknesses

1.I can not see the demo, please include the video. I see a html file in the supplementary material, but the video can not play. Please include necessary examples and comparison

2.I also need the code to understand the detailed implementation. Or the author can include a section in the supplementary material to explain the implementation.

  1. The computation is so slow, make people worry about if the performance improvement is substantial and worth it.

  2. This work can only tackle static scene, but dynamic scene is most important in real-world.

  3. Everything depends on Depth Anything 3 run on generated frames

Questions for authors
  1. Can you include a viewable demo video and generated examples/comparisons of long video examples.

  2. Please include a code copy / implementation explanantion for reproducbility

  3. Explain about the trade-off between time and performance

  4. Any ideas about dynamic scene? will the performance degrade significantly? Any results?

4 comments

We thank the reviewer for the constructive feedback and the recognition of our motivation and presentation. We address each concern below.


Q1: Demo Accessibility.

We sincerely apologize for the inconvenience. The supplementary HTML file (supp/project_page.html) redirects to our anonymous project page. We have also updated an offline video (supp/review-Kt4x/demo_video.mp4) with long-horizon generations and baseline comparisons.


Q2: Code Implementation.

Appendix A.1 (Algorithm 1) already provides the complete streaming inference procedure, with detailed component descriptions in A.2–A.6. To further facilitate reproducibility, we have added annotated code implementations (supp/review-Kt4x/implement_code_detail), covering correspondence construction, GCA, Invisible Octree filtering, and streaming memory updates. We commit to open-sourcing the complete data preprocessing, training, and inference pipelines.


Q3: Efficiency–Performance Trade-off.

The dominant computational cost originates from the pretrained 14B backbone and its iterative denoising rather than our memory mechanism. GCA incurs only 2.1% additional per-forward latency and 1.9% additional parameters (Appendix Tab.4), while reducing ATE by 80.3% over the strongest baseline and improving SSIM, LPIPS, and FVD.

We further demonstrate GEAR's compatibility with diffusion acceleration by applying Distribution Matching Distillation (DMD) to obtain a 4-step bidirectional generator, while reataining the same geometry-addressed memory modules.

MethodStepsCFGDenoising Time / Chunk (s)ATE ↓Revisit SSIM ↑
Wan2.1-14B (backbone)25✓1682.4\\
+ GCA (GEAR)25✓1771.80.04360.6489
+ DMD4✗140.60.04530.6543

Note: Measured on a single 96 GB NVIDIA GPU.

Combining few-step denoising with CFG removal achieves a 12.6× speedup, while preserving camera accuracy and revisit consistency. Even without distillation, direct 4-step sampling still faithfully recovers observed regions, albeit with some loss of fine details. Qualitative comparisons are provided in supp/review-Kt4x/dmd_acceleration.mp4 and our webpage.

We emphasize that our primary contribution lies in accurate visual memory access rather than backbone acceleration. Our DMD results demonstrate establish GEAR's compatibility with diffusion acceleration, with further inference optimization left for future work.


Q4: Dynamic Scene Generation.

Although trained on static scenes, GEAR supports training-free dynamic scene generation by removing dynamic-object correspondences, allowing motion synthesis through recent temporal context and the backbone's generative prior while preserving background consistency.

We demonstrate two training-free applications in Appendix D:

  1. Dynamic object generation. Removing SAM-segmented object correspondences enables dynamic-prompt-guided vehicle motion while maintaining background consistency.
  2. Dynamic object control. In the cow case, prompt modification produces local motion, mask displacement controls object movement, and directly shifting correspondence addresses further improves motion coherence and controllability.

These results (supp/review-Kt4x/dynamic_scene.mp4 and our updated webpage) demonstrate preliminary training-free dynamic generation and motion control through correspondence manipulation. Further training on dynamic scenes could improve motion realism and generalization.


Q5: Dependency on Depth Anything 3.

GEAR uses DA3 solely for geometry-addressed visual memory access, rather than explicit appearance reconstruction or rendering. Historical visual information remains in frame latents, and GCA aggregates relevant features from multiple geometric correspondences, mitigating the impact of geometric inaccuracies.

Our robustness ablation (Fig. 11 and Tab. 2) shows comparable performance under 64–128 pixel correspondence perturbations, retaining only 2–3 valid matches among nine observations, highlighting GEAR's tolerance to imperfect geometry without requiring globally accurate 3D reconstruction.

We further evaluate alternative pose-conditioned geometry estimators: WorldMirror, MapAnything, and Pi3-X. As shown below, GEAR maintains comparable camera-control accuracy and revisit consistency across all alternatives.

VariantRevisit SSIM ↑ / LPIPS ↓ATE ↓
w/ DA30.6489/0.20190.0436
w/ WorldMirror0.6393/0.19200.0448
w/ MapAnything0.6314/0.21030.0457
w/ Pi3-X0.6378/0.20950.0441

These results demonstrate that GEAR requires neither DA3 specifically nor globally accurate 3D reconstruction, but only reliable geometric correspondences for visual memory access.

Dear Reviewer,

We sincerely appreciate your constructive feedback and positive recognition of our core idea and motivation, which are particularly encouraging to us.

Following your suggestions, we have made substantial efforts to address your concerns, including providing accessible demos and detailed implementations, demonstrating significant acceleration through distillation, exploring training-free dynamic scene generation, and validating robustness across alternative geometry estimators.

We sincerely hope that our responses and additional experiments have addressed your concerns and further clarified the value and potential of our approach. Your feedback is highly valuable to us, and we would greatly appreciate hearing your thoughts on our responses. We would also be honored to provide any further clarifications if needed.

Thank you again for your time and constructive feedback.

Thanks. Most of my concerns have been solved.

We sincerely thank the reviewer for the constructive feedback. We are glad that our responses have addressed most of your concerns and would also be pleased to provide any further clarifications if needed.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

Dear AC and Reviewers,

We are greatly encouraged by the reviewers’ recognition of our core insight and experimental results. We sincerely appreciate their constructive feedback and have conducted additional experiments, clarified our methodological contributions, and updated the manuscript and supplementary materials. We summarize the major updates below.

Main Paper & Appendix

  • We have expanded the WorldScore-Static evaluation to 100 randomly sampled scenes with closed-loop camera trajectories, and updated the quantitative metrics in Table 1. (Reviewer RiBP, Q2)
  • We have added a discussion on training-free dynamic scene generation and object motion control through geometric correspondence manipulation to the Appendix D. (Reviewer Kt4x, Q4)

Supplementary Materials

  • As suggested by Reviewer Kt4x (Q1), we provide an offline demonstration video (review-Kt4x/demo_video.mp4) containing representative long-horizon generation results and baseline comparisons to ensure accessibility.
  • As suggested by Reviewer Kt4x (Q2), we provide annotated code implementations (review-Kt4x/implement_code_detail) of our core components to improve reproducibility.
  • In response to Reviewer Kt4x (Q3), we provide additional comparisons with four-step DMD and direct few-step sampling (review-Kt4x/dmd_acceleration.mp4) to demonstrate GEAR's compatibility with diffusion acceleration.
  • In response to Reviewer Kt4x (Q4), we provide additional results on training-free dynamic scene generation. (review-Kt4x/dynamic_scene.mp4).
  • In response to Reviewer oN6p (Q2), we provide an additional ablation with explicit camera-pose conditioning (review-oN6p/camera_condition.mp4), demonstrating the effectiveness of correspondence-based camera control without explicit pose embeddings.
  • As suggested by Reviewer oN6p (Q3), we provide comparisons with warped conditioning (review-oN6p/ablation_warped_condition.mp4).
  • In response to Reviewer RiBP (Q1), we provide additional visualizations of correspondence construction with per-frame local geometry (review-RiBP/local_correspondence.mp4) and geometric error propagation with different variants (review-RiBP/warp_vs_gca.mp4), demonstrating the advantages of geometry-addressed attention over appearance-based conditioning.

We sincerely appreciate the reviewers’ time and constructive feedback, which have substantially improved the clarity, completeness, and experimental validation of our work. We also thank the AC for their efforts in coordinating the review process. We welcome any further questions or discussions.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.