pith. sign in

arxiv: 2606.02479 · v1 · pith:F4IL3AX5new · submitted 2026-06-01 · 💻 cs.CV

Retrieve What's Missing: Coverage-Maximizing Retrieval for Consistent Long Video Generation

Minseok Joo , Dogyun Park , Taehoon Lee , Kyujin Lee , Hyunwoo J. Kim This is my paper

Pith reviewed 2026-06-28 15:09 UTC · model grok-4.3

classification 💻 cs.CV
keywords long video generationgeometric consistencymemory retrievalcoverage maximizationdepth priorsautoregressive video modelssliding-window caching
0
0 comments X

The pith

COVRAG retrieves past frames by maximizing residual target-view coverage from depth priors to sustain geometric consistency in long autoregressive video generation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper targets the problem of geometric drift in long-horizon autoregressive video generation, where memory-augmented models must choose which historical frames to keep. It replaces coarse pose or field-of-view heuristics and expensive explicit 3D reconstruction with a lightweight target-view coverage map built from pretrained depth priors. Frame selection then proceeds by iteratively picking the memory that adds the largest new covered area in the target view. A sliding-window depth cache keeps the process efficient for extended rollouts. The resulting method is shown to raise consistency metrics on RealEstate10K and DL3DV10K while preserving low latency.

Core claim

COVRAG constructs a target-view coverage map from depth estimates produced by pretrained 3D priors; this map serves as lightweight 3D memory evidence that encodes pixel-wise visibility of past observations. Memory frames are then chosen by maximizing residual coverage gain, i.e., the additional target-view area explained by each new frame beyond what the current context and already-selected memories already cover. Sliding-window depth caching maintains efficiency across long sequences, avoiding the maintenance cost of full 3D reconstructions.

What carries the argument

The target-view coverage map (depth-derived pixel-wise visibility mask) together with residual coverage gain maximization for iterative memory-frame selection.

If this is right

  • Long-horizon geometric consistency improves on RealEstate10K and DL3DV10K relative to pose-based and reconstruction-based baselines.
  • Latency stays low because sliding-window depth caching avoids repeated full-scene reconstruction.
  • Memory selection becomes finer-grained than field-of-view overlap without incurring the storage cost of explicit 3D meshes.
  • The same retrieval logic scales to longer rollouts while the coverage map remains lightweight.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The coverage-maximization principle could transfer to other autoregressive domains such as 3D scene synthesis or audio where the goal is to retrieve context that fills currently unobserved structure.
  • If the coverage map proves reliable, downstream video pipelines might reduce reliance on explicit camera calibration or pose estimation.
  • Combining the coverage selector with learned retrieval scores rather than purely geometric gain could further improve results on diverse scene types.

Load-bearing premise

Pretrained 3D priors suffice to construct an accurate target-view coverage map that reasons about pixel-wise visibility without explicit 3D reconstruction or camera poses.

What would settle it

A controlled experiment on RealEstate10K or DL3DV10K in which replacing the coverage-map selector with a simple pose-overlap baseline yields equal or higher geometric-consistency scores and equal or lower latency.

Figures

Figures reproduced from arXiv: 2606.02479 by Dogyun Park, Hyunwoo J. Kim, Kyujin Lee, Minseok Joo, Taehoon Lee.

Figure 1
Figure 1. Figure 1: COVRAG improves consistency under occlusion and out-of-frame motion. COVRAG preserves reappearing content and achieves a better consistency–efficiency trade-off than FoV-based retrieval (WorldMem) and explicit 3D memory (VMem). Abstract Maintaining long-term geometric consistency remains challenging for long-horizon autoregressive video generation. Memory-augmented generative models address this by retriev… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of diffusion video generation with COVRAG. During autoregressive video generation, COVRAG augments the diffusion model with retrieved historical frames that provide complementary geometric evidence for the target view. It identifies regions already supported by the temporal context and retrieves additional memories (denoted by red and blue regions) that cover the remaining target-view regions, ena… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on DL3DV10K. We compare 100-frame autoregressive rollouts from WorldMem, VMem, COVRAG, and the ground-truth video on DL3DV10K trajectories [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of retrieval budget on RealEstate10K. We vary the number of retrieved memory frames n and compare COVRAG with WorldMem. COVRAG consistently achieves better consistency across budgets and saturates with only a few retrieved frames. view coverage computation, including VGGT evaluation, warping historical frames into the target view, and selecting frames by residual coverage gain, its per-step latency … view at source ↗
Figure 6
Figure 6. Figure 6: Example customized loop-closing camera trajectories for RealEstate10K evaluation. Starting from 41-frame RealEstate10K clips, we illustrate two example loop-closing variants obtained by reordering camera poses so that late frames revisit poses close to early frames. Here, Pi denotes the camera pose at frame i. The constructed loops make the final poses approximately match the initial poses, e.g., P0 ≈ P40 … view at source ↗
Figure 7
Figure 7. Figure 7: Examples of DL3DV camera trajectories. We show two representative examples of the original camera paths in the DL3DV10K, with colors indicating temporal frame indices. These examples illustrate how DL3DV10K trajectories can naturally span broad viewpoint changes and partial scene observations, providing a complementary benchmark to the controlled RealEstate10K loop-closing setting. frames and the final rev… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on loop-closing trajectories. We visualize two RealEstate10K sequences generated along loop-closing camera paths, where the camera revisits earlier views after a long temporal gap. COVRAG (top row) better preserves scene geometry and appearance across revisitation (highlighted in green), yielding consistent reconstruction of previously observed regions. In contrast, WorldMem and VMem… view at source ↗
Figure 9
Figure 9. Figure 9: Retrieval behavior and context coverage visualizations on RealEstate10K. For each RealEstate10K loop-closing example, the left panel shows generated frames along the trajectory, the middle panel shows the memory frames retrieved for the indicated target step, and the right panel shows the context coverage map mctx t before adding retrieved memories. The coverage map visualizes which target-view regions are… view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative rollouts on DL3DV10K. We visualize 100-frame autoregressive generations on diverse DL3DV10K scenes and camera trajectories. Frame indices mark temporal order. Across extended rollouts with substantial viewpoint changes, COVRAG maintains coherent scene layout and object structure over long horizons, illustrating the qualitative behavior of coverage￾based retrieval beyond the controll… view at source ↗
read the original abstract

Maintaining long-term geometric consistency remains challenging for long-horizon autoregressive video generation. Memory-augmented generative models address this by retrieving historical frames, but their effectiveness depends on two key design choices: what 3D-geometric evidence should represent past observations, and how memory frames should be selected from this evidence. Existing methods often rely on camera poses or field-of-view overlap, which are lightweight but too coarse to reason about pixel-wise visibility, or use explicit 3D reconstruction, which provides fine-grained evidence but is costly to maintain over long rollouts. We propose Coverage-Maximizing Retrieval-Augmented Generation (COVRAG), a depth-based memory retrieval framework that uses pretrained 3D priors to construct a target-view coverage map as lightweight 3D memory evidence. For frame selection, COVRAG maximizes residual coverage gain, iteratively retrieving frames that explain target-view regions not covered by the current context or previously selected memories. To improve scalability in long-video generation, we introduce sliding-window depth caching for efficient geometry estimation. Experiments on RealEstate10K and DL3DV10K show that COVRAG improves long-horizon geometric consistency while maintaining low latency compared to baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes Coverage-Maximizing Retrieval-Augmented Generation (COVRAG), a depth-based memory retrieval method for autoregressive long video generation. It constructs a target-view coverage map from pretrained 3D priors (monocular depth) as lightweight geometric evidence and selects memory frames by iteratively maximizing residual coverage gain. A sliding-window depth caching scheme is introduced for efficiency. The abstract claims that this yields improved long-horizon geometric consistency on RealEstate10K and DL3DV10K while keeping latency low relative to pose-based or explicit-reconstruction baselines.

Significance. If the coverage map delivers reliable pixel-wise visibility estimates, COVRAG would supply a practical middle ground between coarse pose/FoV heuristics and expensive 3D reconstruction, directly addressing a core bottleneck in memory-augmented video models. The sliding-window caching is a concrete engineering contribution that could transfer to other long-horizon generation pipelines.

major comments (2)
  1. [Method] Method section (coverage-map construction): the central claim that pretrained monocular depth priors suffice to produce an accurate target-view coverage map for pixel-wise visibility reasoning (without camera poses or explicit reconstruction) is load-bearing for the consistency improvement. No direct validation metric (visibility IoU, boundary error, or occlusion accuracy against ground truth) is supplied to show the map is sufficiently reliable for the residual-coverage-gain objective to outperform coarser alternatives.
  2. [Experiments] Experiments (RealEstate10K / DL3DV10K results): the abstract asserts quantitative improvements in long-horizon geometric consistency, yet supplies no numerical values, baseline comparisons, error bars, or ablation tables. Without these data the magnitude and attribution of gains to the coverage-maximization step cannot be assessed.
minor comments (1)
  1. [Abstract] Abstract: the sentence reporting experimental results omits all metrics, making the strength of the empirical claim difficult to gauge at first reading.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and describe the corresponding revisions.

read point-by-point responses

  1. Referee: [Method] Method section (coverage-map construction): the central claim that pretrained monocular depth priors suffice to produce an accurate target-view coverage map for pixel-wise visibility reasoning (without camera poses or explicit reconstruction) is load-bearing for the consistency improvement. No direct validation metric (visibility IoU, boundary error, or occlusion accuracy against ground truth) is supplied to show the map is sufficiently reliable for the residual-coverage-gain objective to outperform coarser alternatives.

    Authors: We agree that an explicit validation of the coverage map would strengthen the central claim. The current manuscript evaluates the map only indirectly via downstream geometric consistency. In the revision we will add a new subsection reporting visibility IoU, boundary error, and occlusion accuracy against ground-truth visibility derived from the datasets' known poses, thereby directly quantifying the reliability of the monocular-depth coverage map. revision: yes

  2. Referee: [Experiments] Experiments (RealEstate10K / DL3DV10K results): the abstract asserts quantitative improvements in long-horizon geometric consistency, yet supplies no numerical values, baseline comparisons, error bars, or ablation tables. Without these data the magnitude and attribution of gains to the coverage-maximization step cannot be assessed.

    Authors: The experiments section contains the requested quantitative results, including tables with consistency metrics, comparisons against pose-based and reconstruction baselines, error bars from multiple runs, and ablations isolating the coverage-maximization component. To address the abstract's lack of specificity, we will revise the abstract to include key numerical improvements and explicit references to the experimental tables. revision: partial

Circularity Check

0 steps flagged

No circularity: method uses external pretrained priors and reports empirical gains on held-out data

full rationale

The paper proposes COVRAG as a retrieval framework that constructs a coverage map from off-the-shelf monocular depth estimators and selects frames by maximizing residual coverage gain. No equations are presented that define the coverage map or selection objective in terms of the final consistency metric; the priors are imported from external pretrained models rather than fitted or derived within the paper. Experiments compare against baselines on RealEstate10K and DL3DV10K without any self-referential fitting of the claimed improvement. No self-citation chains or uniqueness theorems are invoked to justify core components. The derivation therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Central claim rests on the domain assumption that pretrained depth models yield sufficiently accurate coverage maps and that residual coverage gain is a good proxy for geometric consistency.

axioms (1)
  • domain assumption Pretrained 3D priors yield reliable per-pixel depth sufficient for visibility reasoning.
    Invoked when constructing the target-view coverage map from depth evidence.

pith-pipeline@v0.9.1-grok · 5757 in / 1020 out tokens · 25620 ms · 2026-06-28T15:09:32.368738+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 4 linked inside Pith

  1. [1]

    Vid- man: Exploiting implicit dynamics from video diffusion model for effective robot manipulation

    Youpeng Wen, Junfan Lin, Yi Zhu, Jianhua Han, Hang Xu, Shen Zhao, and Xiaodan Liang. Vid- man: Exploiting implicit dynamics from video diffusion model for effective robot manipulation. InNeural Information Processing Systems, NeurIPS, 2024

    2024

  2. [2]

    Diffusion models for robotic manipulation: A survey.Frontiers in Robotics and AI, 2025

    Rosa Wolf, Yitian Shi, Sheng Liu, and Rania Rayyes. Diffusion models for robotic manipulation: A survey.Frontiers in Robotics and AI, 2025

    2025

  3. [3]

    Unisim: A unified simulator for time-coarsened dynamics of biomolecules

    Ziyang Yu, Wenbing Huang, and Yang Liu. Unisim: A unified simulator for time-coarsened dynamics of biomolecules. InInternational Conference on Machine Learning, ICML, 2025

    2025

  4. [4]

    Irasim: A fine-grained world model for robot manipulation

    Fangqi Zhu, Hongtao Wu, Song Guo, Yuxiao Liu, Chilam Cheang, and Tao Kong. Irasim: A fine-grained world model for robot manipulation. InComputer Vision and Pattern Recognition, CVPR, 2025

    2025

  5. [5]

    Wonderjourney: Going from anywhere to everywhere

    Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. InComputer Vision and Pattern Recognition, CVPR, 2024

    2024

  6. [6]

    Cameractrl: Enabling camera control for video diffusion models

    Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for video diffusion models. InInternational Conference on Learning Representations, ICLR, 2025

    2025

  7. [7]

    Recammaster: Camera-controlled generative rendering from a single video

    Jianhong Bai, Menghan Xia, Xiao Fu, Xintao Wang, Lianrui Mu, Jinwen Cao, Zuozhu Liu, Haoji Hu, Xiang Bai, Pengfei Wan, et al. Recammaster: Camera-controlled generative rendering from a single video. InInternational Conference on Computer Vision, ICCV, 2025

    2025

  8. [8]

    CAT3D: Create anything in 3d with multi-view diffusion models

    Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. CAT3D: Create anything in 3d with multi-view diffusion models. InNeural Information Processing Systems, NeurIPS, 2024

    2024

  9. [9]

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  10. [10]

    Wonder- world: Interactive 3d scene generation from a single image

    Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonder- world: Interactive 3d scene generation from a single image. InComputer Vision and Pattern Recognition, CVPR. IEEE Computer Society, 2025

    2025

  11. [11]

    The matrix: Infinite-horizon world generation with real-time moving control

    Ruili Feng, Han Zhang, Zhantao Yang, Jie Xiao, Zhilei Shu, Zhiheng Liu, Andy Zheng, Yukun Huang, Yu Liu, and Hongyang Zhang. The matrix: Infinite-horizon world generation with real-time moving control. InNeural Information Processing Systems, NeurIPS, 2025

    2025

  12. [12]

    Gamefactory: Cre- ating new games with generative interactive videos

    Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Gamefactory: Cre- ating new games with generative interactive videos. InInternational Conference on Computer Vision, ICCV, 2025

    2025

  13. [13]

    Diffusion models are real-time game engines

    Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. InInternational Conference on Learning Representations, ICLR, 2025

    2025

  14. [14]

    Genie: Generative interactive environments

    Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. InInternational Conference on Machine Learning, ICML, 2024

    2024

  15. [15]

    WorldMem: Long-term consistent world simulation with memory

    Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, and Xingang Pan. WorldMem: Long-term consistent world simulation with memory. InNeural Information Processing Systems, NeurIPS, 2025

    2025

  16. [16]

    Context as memory: Scene-consistent interactive long video generation with memory retrieval

    Jiwen Yu, Jianhong Bai, Yiran Qin, Quande Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Context as memory: Scene-consistent interactive long video generation with memory retrieval. InProceedings of the SIGGRAPH Asia 2025 Conference Papers, 2025. 10

    2025

  17. [17]

    VMem: Consistent interactive video scene generation with surfel-indexed view memory

    Runjia Li, Philip Torr, Andrea Vedaldi, and Tomas Jakab. VMem: Consistent interactive video scene generation with surfel-indexed view memory. InInternational Conference on Computer Vision, ICCV, 2025

    2025

  18. [18]

    Learning world models for interactive video generation

    Taiye Chen, Xun Hu, Zihan Ding, and Chi Jin. Learning world models for interactive video generation. InNeural Information Processing Systems, NeurIPS, 2025

    2025

  19. [19]

    Streamingt2v: Consistent, dynamic, and extendable long video generation from text

    Roberto Henschel, Levon Khachatryan, Hayk Poghosyan, Daniil Hayrapetyan, Vahram Tade- vosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streamingt2v: Consistent, dynamic, and extendable long video generation from text. InComputer Vision and Pattern Recognition, CVPR, 2025

    2025

  20. [20]

    Memcam: Memory-augmented camera control for consistent video generation.arXiv preprint arXiv:2603.26193, 2026

    Xinhang Gao, Junlin Guan, Shuhan Luo, Wenzhuo Li, Guanghuan Tan, and Jiacheng Wang. Memcam: Memory-augmented camera control for consistent video generation.arXiv preprint arXiv:2603.26193, 2026

    arXiv 2026

  21. [21]

    VGGT: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. VGGT: Visual geometry grounded transformer. InComputer Vision and Pattern Recognition, CVPR, 2025

    2025

  22. [22]

    History-guided video diffusion

    Kiwhan Song, Boyuan Chen, Max Simchowitz, Yilun Du, Russ Tedrake, and Vincent Sitzmann. History-guided video diffusion. InInternational Conference on Machine Learning, ICML, 2025

    2025

  23. [23]

    Stereo magnifi- cation: Learning view synthesis using multiplane images.arXiv preprint arXiv:1805.09817, 2018

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnifi- cation: Learning view synthesis using multiplane images.arXiv preprint arXiv:1805.09817, 2018

    Pith/arXiv arXiv 2018

  24. [24]

    DL3DV-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. DL3DV-10k: A large-scale scene dataset for deep learning-based 3d vision. InComputer Vision and Pattern Recognition, CVPR, 2024

    2024

  25. [25]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InNeural Information Processing Systems, NeurIPS, 2020

    2020

  26. [26]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, ICLR, 2021

    2021

  27. [27]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. InInternational Conference on Learning Representations, ICLR, 2025

    2025

  29. [29]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Diffusion forcing: Next-token prediction meets full-sequence diffusion

    Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. InNeural Information Processing Systems, NeurIPS, 2024

    2024

  31. [31]

    From slow bidirectional to fast autoregressive video diffusion models

    Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. In Computer Vision and Pattern Recognition, CVPR, 2025

    2025

  32. [32]

    Self forcing: Bridging the train-test gap in autoregressive video diffusion

    Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train-test gap in autoregressive video diffusion. InNeural Information Processing Systems, NeurIPS, 2025. 11

    2025

  33. [33]

    Diffusion adversarial post-training for one-step video generation

    Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. InInternational Conference on Machine Learning, ICML, 2025

    2025

  34. [34]

    Mixture of contexts for long video generation

    Shengqu Cai, Ceyuan Yang, Lvmin Zhang, Yuwei Guo, Junfei Xiao, Ziyan Yang, Yinghao Xu, Zhenheng Yang, Alan Yuille, Leonidas Guibas, Maneesh Agrawala, Lu Jiang, and Gordon Wetzstein. Mixture of contexts for long video generation. InInternational Conference on Learning Representations, ICLR, 2026

    2026

  35. [35]

    Reconx: Reconstruct any scene from sparse views with video diffusion model

    Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024

    arXiv 2024

  36. [36]

    3d reconstruction with spatial memory

    Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. InInternational Conference on 3D Vision, 3DV, 2025

    2025

  37. [37]

    Gen3c: 3d-informed world- consistent video generation with precise camera control

    Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world- consistent video generation with precise camera control. InComputer Vision and Pattern Recognition, CVPR, 2025

    2025

  38. [38]

    World-consistent video diffusion with explicit 3d modeling

    Qihang Zhang, Shuangfei Zhai, Miguel Angel Bautista Martin, Kevin Miao, Alexander Toshev, Joshua Susskind, and Jiatao Gu. World-consistent video diffusion with explicit 3d modeling. In Computer Vision and Pattern Recognition, CVPR, 2025

    2025

  39. [39]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations, ICLR, 2023

    2023

  40. [40]

    The unrea- sonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unrea- sonable effectiveness of deep features as a perceptual metric. InComputer Vision and Pattern Recognition, CVPR, 2018

    2018

  41. [41]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 2004

    2004

  42. [42]

    MEt3R: Measuring multi-view consistency in generated images

    Mohammad Asim, Christopher Wewer, Thomas Wimmer, Bernt Schiele, and Jan Eric Lenssen. MEt3R: Measuring multi-view consistency in generated images. InComputer Vision and Pattern Recognition, CVPR, 2024

    2024

  43. [43]

    DUSt3R: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3d vision made easy. InComputer Vision and Pattern Recognition, CVPR, 2024

    2024

  44. [44]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InInternational Conference on Computer Vision, ICCV, 2021

    2021

  45. [45]

    Towards accurate generative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018

    Pith/arXiv arXiv 2018

  46. [46]

    GANs trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. InNeural Information Processing Systems, NeurIPS, 2017

    2017

  47. [47]

    Efros, and Angjoo Kanazawa

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InComputer Vision and Pattern Recogni- tion, CVPR, 2025

    2025

  48. [48]

    # 𝑃$%≈𝑃% 𝑃%≈𝑃$% 𝑃

    Jensen (Jinghao) Zhou, Hang Gao, Vikram V oleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models.arXiv preprint arXiv:2503.14489, 2025. 12 A Experimental Details A.1 Baseline Setup We provide additional details on the baseline configurat...

    arXiv 2025