Pith. sign in

REVIEW 3 major objections 4 minor 93 references

Training-free inference tricks turn frozen text-to-video models into multi-shot cinematic generators.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 13:51 UTC pith:KU5BPZJD

load-bearing objection A clean, useful training-free recipe for multi-shot reference-controlled long video generation, but the central premise is under-tested because the method bundles shot-wise conditioning with the continuity-breaking tricks. the 3 major comments →

arxiv 2607.26529 v1 pith:KU5BPZJD submitted 2026-07-29 cs.CV

CineWeaver: Training-Free Reference-Controllable Multi-Shot Long Video Generation for Cinematic Storytelling

Yuyang Huang , Yabo Chen , Wenrui Dai , Ziyang Zheng , Haibin Huang , Chi Zhang , Junni Zou , Hongkai Xiong
show 1 more author
Xuelong Li
This is my paper
classification cs.CV
keywords multi-shot video generationtraining-free inferencereference-controllable generationlong video generationdiffusion transformerstemporal continuityattention maskingcinematic video
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

CineWeaver claims that multi-shot, reference-controllable, long video generation does not require retraining or new architectures. The paper's central assertion is that pretrained text-to-video diffusion models already know how to compose distinct scenes; the only real obstacle is a structural bias toward temporal continuity built into the generation process. By manipulating rotary positional encoding, self-attention masks, per-shot cross-attention, and VAE decoding only at inference time, CineWeaver produces clear shot boundaries, per-shot reference adherence, and cross-segment global consistency using frozen Wan2.1 and Phantom backbones. If this claim holds, it would mean that sophisticated cinematic video editing can be achieved as an inference-time operation rather than a new learning problem.

Core claim

The paper claims that pretrained text-to-video diffusion models inherently possess the ability to compose visually distinct scenes, and that the apparent difficulty of multi-shot generation stems from four structural continuity biases: RoPE-based temporal proximity, global self-attention across adjacent frames, shared prompt conditioning, and temporal caching in VAE decoding. CineWeaver selectively relaxes these biases at inference time. Specifically, it inserts 'gap frames' to enlarge RoPE positional offsets between shots, uses masked self-attention with 'transition frames' to isolate early shot tokens, applies shot-wise cross-attention and FFN to avoid inter-shot prompt interference, and d

What carries the argument

The central mechanism is the interruption of temporal continuity in a frozen DiT-based video diffusion backbone. Concretely: (i) RoPE manipulation with gap frames—inserting dummy frames between shots so the relative positional offset between shots is enlarged, weakening pairwise attention; (ii) masked self-attention with transition frames—the first few frames of each shot attend only within their own shot, while remaining frames see all non-gap shot tokens; (iii) shot-wise cross-attention and FFN—each shot's tokens only attend to its own prompt tokens; and (iv) isolated VAE decoding—each shot is decoded with a reset cache to avoid visual leakage. Reference routing extends the mask so referen

Load-bearing premise

The load-bearing premise is that a frozen pretrained video diffusion model already contains sufficient priors for composing visually distinct scenes, so that interrupting temporal continuity at inference is all that is needed to produce valid multi-shot video.

What would settle it

Generate a multi-shot video on a style or scene combination well outside the training distribution of the backbone (e.g., a surreal Escher-like interior with no real-world counterpart) and measure whether the training-free gap-frame/transition-frame manipulation still produces perceptually valid shot boundaries and coherent content; if the model produces artifacts or fails to separate shots, the continuity-bias-only hypothesis is weakened.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If true, multi-shot and cinematic video generation becomes an inference-time capability of existing large video diffusion models, eliminating the need for specialized multi-shot training data and fine-tuning.
  • The same frozen backbone can switch between single-shot, multi-shot, reference-controllable, and long-form modes simply by changing the attention mask and positional encoding schedule.
  • Shot-level reference routing enables per-shot character, object, and background control without training a new conditioning head, which could lower the cost of customizing video generation for storytelling.
  • The anchor-memory design permits long videos (dozens of shots, minute-level durations) by decomposing generation into independently generated segments that share global appearance cues, and it parallelizes across GPUs.

Where Pith is reading between the lines

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

  • A testable extension would be to apply the same continuity-interruption recipe to other pretrained video diffusion models (e.g., open-source 7B or 30B backbones) to see whether the claimed sufficiency of the four inference-time manipulations generalizes beyond Wan2.1 and Phantom.
  • The framework implicitly suggests that 'temporal continuity' is a transferable dial: varying gap-frame counts, transition-frame counts, and mask sparsity might enable a continuous spectrum from hard cuts to soft dissolves, which the paper only partially explores.
  • If the central premise is correct, it would imply that much of the apparent need for custom multi-shot models is an artifact of training biases, and that future video foundation models could be made multi-shot-capable by design simply by not over-regularizing temporal smoothness.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes CineWeaver, an inference-only framework for reference-controllable multi-shot long video generation. It identifies a structural temporal-continuity bias in pretrained text-to-video diffusion models and argues that multi-shot composition can be achieved by manipulating this bias during inference rather than retraining. The method combines shot-aware RoPE with gap frames, masked self-attention with transition frames, shot-wise cross-attention and FFN, and isolated VAE decoding. It then adds a shot-routed reference-token mechanism for per-shot appearance control and an anchor-memory mechanism for cross-segment global consistency. Experiments compare CineWeaver with training-based multi-shot methods (CineTrans, EchoShot, HoloCine, MultiShotMaster) and with reference baselines constructed by concatenating single-shot Phantom and VACE outputs. Ablations and a user study support the component design.

Significance. If the central claim holds, CineWeaver is a useful contribution: it would show that frozen pretrained video diffusion models can produce multi-shot, reference-controlled long videos without retraining, and its attention-mask and anchor-token mechanisms could be reused by other inference-time systems. The paper is clearly written, the components are defined in detail, and the training-free setting is attractive for practical adoption. However, the empirical validation does not yet isolate the paper's main causal claim, and the reference-controllable baselines are weaker than the framing suggests. The authors do not release code, which would help substantiate the reproducibility of the method.

major comments (3)
  1. [§4.1.2, Table 1] The reference-controllable baselines are constructed by generating each shot independently with Phantom and VACE and concatenating the outputs. These are single-shot models with no multi-shot mechanism, no shot-routing, and no explicit handling of transition boundaries. The large gains in inter-shot consistency and narrative coherence in Table 1 therefore partly reflect the construction of the baseline, not necessarily superiority over a true multi-shot reference-controllable method. MultiShotMaster is cited as the only prior unified method, but its reference module was unavailable. Please compare against an available multi-shot reference method, or at least equalize shot-level conditioning for Phantom/VACE and clearly label the comparison as 'single-shot concatenation' rather than a multi-shot baseline.
  2. [§3.1, Table 2] The central claim that pretrained T2V models already possess multi-shot composition priors and that only temporal-continuity bias blocks multi-shot generation is not isolated by the ablations. Every row in Table 2 keeps shot-wise cross-attention, shot-wise FFN, and shot-wise VAE decoding. None of the ablations removes the shot-wise isolation itself; for example, there is no global-prompt/global-FFN control and no ablation without shot-wise FFN. Without such a control, the observed transitions could be caused by isolating prompts and features per shot rather than by releasing a pre-existing compositional prior through gap frames and masked attention. Add a 'global prompt / no shot-wise FFN' condition to Table 2, or weaken the causal claim in Section 3.1 accordingly.
  3. [§4.2, Tables 1–3] All quantitative comparisons are point estimates with no error bars, confidence intervals, or significance tests. Several differences supporting the conclusions are small (e.g., Text Align 0.259 vs 0.264 in Table 1; Narrative Coherence differences in Table 3 across NA). The long benchmark has only 15 prompt sets, so the reported margins may be within noise. Report per-sample means, standard errors or bootstrap intervals, and the number of generated videos per condition. The user study (Table 4) should also state the number of participants and pairs; the current preference percentages are not interpretable without that information.
minor comments (4)
  1. [§3.2.2] Typo: 'quey' should be 'query'.
  2. [Table 1] The first block has stray '%' symbols and inconsistent number formatting; the column alignment is also confusing as '0.089% %'.
  3. [§4.4.2] For NG/NS, the text states default NG=5 for Wan2.1 but later 'NG is set to 2 by default for reference-controllable generation'; please make the default values per setting explicit and consistent across Sections 4.1.1 and 4.4.2.
  4. [Reproducibility] The project page is mentioned but no code or model weights are linked. Releasing the inference implementation would materially strengthen the claim of a training-free framework.

Circularity Check

0 steps flagged

No circular derivation; inference-time design tested against frozen external backbones.

full rationale

The paper's load-bearing claim—that frozen T2V diffusion models can perform multi-shot composition once temporal-continuity biases are interrupted—is an empirical hypothesis, not a quantity derived from assumptions. Section 3.2 defines explicit masks and RoPE offsets (Eqs. 1-2), and Sections 3.3-3.4 extend them; no parameter is fitted to the headline metrics and no predicted value is algebraically equal to an input. The methods are evaluated against external baselines (CineTrans, EchoShot, HoloCine, MultiShotMaster, Phantom, VACE) on frozen Wan2.1-14B/Phantom-14B backbones, and the only backbone reuse is transparent: CineWeaver routes Phantom's clean-to-noisy reference tokens per shot, while Phantom is also independently concatenated as a baseline. That creates a possible comparison asymmetry, but not circularity. Self-citations in the related-work section are contextual and not load-bearing for the central derivation. The main scientific risk is external validity—the decisive premise that pretrained models already possess composition priors is not isolated from shot-wise FFN/VAE isolation in the ablations—but that is a threat to validity, not a self-referential reduction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper does not introduce a new natural-kind entity; Anchor Memory is an algorithmic component, not a postulated physical object. The central claim relies on four domain assumptions about pretrained-model priors and attention behavior, plus three tuned hyperparameters (NG, NS, NA). These are the main uncharged inputs the reader must grant.

free parameters (3)
  • NG (number of gap frames) = 2 for Phantom-based reference-controllable generation; 5 for Wan2.1 multi-shot T2V
    Chosen from qualitative ablation in Fig. 6; text says NG=2 already gives clear transitions and larger values only add compute.
  • NS (number of transition frames) = 2
    Chosen from qualitative ablation in Fig. 7; NS=0 weakens transitions, larger values add compute.
  • NA (number of anchor frames) = 5
    Chosen from Table 3 and Fig. 8 ablation; constrained to 4n+1 by Wan VAE temporal compression; NA=5 balances transition accuracy, consistency, and compute.
axioms (4)
  • ad hoc to paper Pretrained video diffusion models already possess sufficient scene priors to compose visually distinct scenes; only the structural temporal-continuity bias blocks multi-shot generation.
    Stated in the Introduction and Section 3.1 as the key insight; no independent mechanism-level proof is given.
  • domain assumption RoPE attention scores between temporally adjacent frames are higher due to smaller relative offsets, so adding excluded gap frames weakens inter-shot coupling without harmful side effects.
    Uses Eq. 1 and the monotonic dependence of attention on |Δd|; assumes discarding gap tokens does not disturb latent distributions.
  • domain assumption Masking transition-frame tokens to their own shot and restricting reference tokens to their own shot prevents cross-shot leakage without destroying intra-shot coherence.
    The masks in Eqs. 2 and 5 enforce isolation by construction; the assumption that this preserves generation quality is only supported by ablations, not proven.
  • domain assumption Anchor tokens from the first generated segment propagate global style to later segments, and restricting anchor self-attention to themselves prevents their representation from being altered by the new context.
    Section 3.4.2 and Eq. 7; the paper provides empirical trade-offs (Table 3) but no theoretical justification for why anchors transfer style without structural interference.

pith-pipeline@v1.3.0-daily-deepseek · 17019 in / 10395 out tokens · 97108 ms · 2026-08-01T13:51:44.006887+00:00 · methodology

0 comments
read the original abstract

Cinematic video generation is challenging for text-to-video diffusion models due to concurrent requirements on multi-shot generation, fine-grained controllability over characters and scenes, and long-form generation across extended temporal horizons. Existing methods rely on customization and retraining to separately address specific requirements, and cannot simultaneously fulfill all the requirements with a unified framework. In this paper, we shed light on the training-free paradigm with the key insight that the difficulty of multi-shot generation arises from a structural bias toward temporal continuity in pretrained video diffusion models, and consequently, propose a unified framework named CineWeaver to achieve reference-controllable multi-shot long-video generation without retraining. We manipulate positional encoding and attention patterns to break temporal continuity during inference to enable clear shot transitions using pretrained video diffusion models. Furthermore, we extend the proposed framework with a shot-routed reference conditioning mechanism for per-shot fine-grained controllability, and develop an anchor memory mechanism to allow long-form generation with consistent global appearance cues. To our best knowledge, CineWeaver is the first unified framework to simultaneously enable \textbf{long-form}, \textbf{reference-controllable}, and \textbf{multi-shot} video generation in a training-free fashion. Experimental results demonstrate that CineWeaver produces high-quality cinematic videos of long durations with consistent identities, stable global appearance, and clear shot transitions. The project page is available at: https://cineweaver.github.io.

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

93 extracted references · 19 linked inside Pith

  1. [1]

    FirstName Alpher , title =

  2. [2]

    Journal of Foo , volume = 13, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =

  3. [3]

    Journal of Foo , volume = 14, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =

  4. [4]

    FirstName Alpher and FirstName Gamow , title =

  5. [5]

    Computer Vision -- ECCV 2022 , year =

    2022

  6. [6]

    2025 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Phantom: Subject-consistent video generation via cross-modal alignment , author=. 2025 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    2025

  7. [7]

    Wu, Xiaoxue and Gao, Bingjie and Qiao, Yu and Wang, Yaohui and Chen, Xinyuan , booktitle=

  8. [8]

    Wang, Jiahao and Sheng, Hualian and Cai, Sijia and Zhang, Weizhan and Yan, Caixia and Feng, Yachuang and Deng, Bing and Ye, Jieping , booktitle=

  9. [9]

    Meng, Yihao and Ouyang, Hao and Yu, Yue and Wang, Qiuyu and Wang, Wen and Cheng, Ka Leong and Wang, Hanlin and Li, Yixuan and Chen, Cheng and Zeng, Yanhong and Zhu, Xing and Shen, Yujun and Qu, Huamin , booktitle=

  10. [10]

    Wang, Qinghe and Shi, Xiaoyu and Li, Baolu and Bian, Weikang and Liu, Quande and Lu, Huchuan and Wang, Xintao and Wan, Pengfei and Gai, Kun and Jia, Xu , booktitle=

  11. [11]

    Jiang, Zeyinzi and Han, Zhen and Mao, Chaojie and Zhang, Jingfeng and Pan, Yulin and Liu, Yu , booktitle=

  12. [12]

    2025 , month = nov, note =

    A new era of intelligence with Gemini 3 , author =. 2025 , month = nov, note =

    2025

  13. [13]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  14. [14]

    Wang, Yi and He, Yinan and Li, Yizhuo and Li, Kunchang and Yu, Jiashuo and Ma, Xin and Li, Xinhao and Chen, Guo and Chen, Xinyuan and Wang, Yaohui and Luo, Ping and Liu, Ziwei and Wang, Yali and Wang, Limin and Yu Qiao , booktitle =

  15. [15]

    Khanam, Rahima and Hussain, Muhammad , journal=

  16. [16]

    2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Segment anything , author=. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    2023

  17. [17]

    Transactions on Machine Learning Research , year=

    Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research , year=

  18. [18]

    Soucek, Tom. Proc. 32nd ACM Int. Conf. Multimedia , pages=

  19. [19]

    arXiv preprint arXiv:2503.20314 , year=

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

    Pith/arXiv arXiv

  20. [21]

    Fei, Zhengcong and Li, Debang and Qiu, Di and Wang, Jiahua and Dou, Yikun and Wang, Rui and Xu, Jingtao and Fan, Mingyuan and Chen, Guibin and Li, Yang and Zhou, Yahui , journal=

  21. [23]

    Yang, Zhuoyi and Teng, Jiayan and Zheng, Wendi and Ding, Ming and Huang, Shiyu and Xu, Jiazheng and Yang, Yuanming and Hong, Wenyi and Zhang, Xiaohan and Feng, Guanyu and Yin, Da and Zhang, Yuxuan and Wang, Weihan and Cheng, Yean and Bin, Xu and Gu, Xiaotao and Dong, Yuxiao and Tang, Jie , booktitle=

  22. [24]

    2024 , publisher=

    Su, Jianlin and Ahmed, Murtadha and Lu, Yu and Pan, Shengfeng and Bo, Wen and Liu, Yunfeng , journal=. 2024 , publisher=

    2024

  23. [25]

    2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    From slow bidirectional to fast autoregressive video diffusion models , author=. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    2025

  24. [26]

    Advances in Neural Information Processing Systems 38 , pages =

    Self forcing: Bridging the train-test gap in autoregressive video diffusion , author=. Advances in Neural Information Processing Systems 38 , pages =

  25. [27]

    The Fourteenth International Conference on Learning Representations , year=

    Rolling forcing: Autoregressive long video diffusion in real time , author=. The Fourteenth International Conference on Learning Representations , year=

  26. [28]

    SIGGRAPH Asia 2024 Conference Papers , pages=

    Lumiere: A space-time diffusion model for video generation , author=. SIGGRAPH Asia 2024 Conference Papers , pages=

    2024

  27. [29]

    Latte: Latent diffusion

    Ma, Xin and Wang, Yaohui and Chen, Xinyuan and Jia, Gengyun and Liu, Ziwei and Li, Yuan-Fang and Chen, Cunjian and Qiao, Yu , journal=. Latte: Latent diffusion

  28. [31]

    Henschel, Roberto and Khachatryan, Levon and Poghosyan, Hayk and Hayrapetyan, Daniil and Tadevosyan, Vahram and Wang, Zhangyang and Navasardyan, Shant and Shi, Humphrey , booktitle=

  29. [32]

    2026 , pages=

    Cui, Justin and Wu, Jie and Li, Ming and Yang, Tao and Li, Xiaojie and Wang, Rui and Bai, Andrew and Ban, Yuanhao and Hsieh, Cho-Jui , booktitle=. 2026 , pages=

    2026

  30. [33]

    Advances in Neural Information Processing Systems 38 , pages =

    Frame context packing and drift prevention in next-frame-prediction video diffusion models , author=. Advances in Neural Information Processing Systems 38 , pages =

  31. [34]

    She, D and Liu, Mushui and Pang, Jingxuan and Wang, Jin and Yang, Zhen and He, Wanggui and Zhang, Guanghao and Wang, Yi and Huang, Qihan and Tang, Haobin and Yu, Yunlong and Fu, Siming , journal=

  32. [35]

    Mai, Ziyang and Tai, Yu-Wing , journal=

  33. [36]

    Proceedings of the 24th ACM International Conference on Multimedia , pages=

    Magic mirror: A virtual fashion consultant , author=. Proceedings of the 24th ACM International Conference on Multimedia , pages=

  34. [37]

    arXiv preprint arXiv:2510.18573 , year=

    Kaleido: Open-Sourced Multi-Subject Reference Video Generation Model , author=. arXiv preprint arXiv:2510.18573 , year=

    arXiv

  35. [38]

    2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Scaling Zero-Shot Reference-to-Video Generation , author=. 2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    2026

  36. [39]

    Lu, Yu and Liang, Yuanzhi and Zhu, Linchao and Yang, Yi , booktitle=

  37. [40]

    2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Longdiff: Training-free long video generation in one go , author=. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    2025

  38. [41]

    Yang, Shuai and Huang, Wei and Chu, Ruihang and Xiao, Yicheng and Zhao, Yuyang and Wang, Xianbang and Li, Muyang and Xie, Enze and Chen, Yingcong and Lu, Yao and Han, Song and Chen, Yukang , booktitle=

  39. [42]

    Advances in Neural Information Processing Systems 37 , pages=

    Diffusion forcing: Next-token prediction meets full-sequence diffusion , author=. Advances in Neural Information Processing Systems 37 , pages=

  40. [43]

    Malt diffusion: Memory-augmented latent

    Yu, Sihyun and Hahn, Meera and Kondratyuk, Dan and Shin, Jinwoo and Gupta, Agrim and Lezama, Jos. Malt diffusion: Memory-augmented latent. CVPR 2025 Workshop on AI for Content Creation , year=

    2025

  41. [44]

    arXiv preprint arXiv:2510.01784 , year=

    Pack and force your memory: Long-form and consistent video generation , author=. arXiv preprint arXiv:2510.01784 , year=

    arXiv

  42. [45]

    Huang, Yuyang and Chen, Yabo and Ding, Li and Zhang, Xiaopeng and Dai, Wenrui and Zou, Junni and Xiong, Hongkai and Tian, Qi , booktitle=

  43. [46]

    Towards Holistic Modeling for Video Frame Interpolation with Auto-regressive Diffusion

    Peng, Xinyu and Li, Han and Huang, Yuyang and Zheng, Ziyang and Wang, Yaoming and Chen, Xin and Dai, Wenrui and Li, Chenglin and Zou, Junni and Xiong, Hongkai , booktitle=. Towards Holistic Modeling for Video Frame Interpolation with Auto-regressive Diffusion. 2026 , pages=

    2026

  44. [47]

    Chen, Yabo and Liang, Yuanzhi and Wang, Jiepeng and Chen, Tingxi and Cheng, Junfei and Gu, Zixiao and Huang, Yuyang and Jiang, Zicheng and Li, Wei and Li, Tian and Li, Weichen and Li, Zuoxin and Liu, Guangce and Liu, Jialun and Liu, Junqi and Wang, Haoyuan and Weng, Qizhen and Wu, Xuan'er and Xiang, Xunzhi and Yang, Xiaoyan and Zhang, Xin and Zhang, Shiwe...

  45. [48]

    arXiv preprint arXiv:2508.03334 , year=

    Macro-from-micro planning for high-quality and parallelized autoregressive long video generation , author=. arXiv preprint arXiv:2508.03334 , year=

    arXiv

  46. [49]

    2026 , publisher=

    Wang, Zhao and Li, Aoxue and Zhu, Lingting and Guo, Yong and Dou, Qi and Li, Zhenguo , journal=. 2026 , publisher=

    2026

  47. [50]

    Zhao, Haoyu and Gu, Jiaxi and Wang, Shicong and Lu, Tianyi and Zhang, Xing and Wu, Zuxuan and Xu, Hang and Jiang, Yu-Gang , journal=

  48. [51]

    A Benchmark for Controllable Text -Image-to-Video Generation , year=

    Hu, Yaosi and Luo, Chong and Chen, Zhenzhong , journal=. A Benchmark for Controllable Text -Image-to-Video Generation , year=

  49. [52]

    IEEE Transactions on Multimedia , volume=

    Tuning-Free High-Resolution Video Diffusion With Spatial-Temporal Latent Grouping , author=. IEEE Transactions on Multimedia , volume=

  50. [53]

    Chen, Hong and Wang, Xin and Zeng, Guanning and Zhang, Yipeng and Zhou, Yuwei and Han, Feilin and Wu, Yaofei and Zhu, Wenwu , journal=

  51. [54]

    2025 , publisher=

    Shao, Jiawei and Li, Xuelong , journal=. 2025 , publisher=

    2025

  52. [55]

    2026 , publisher=

    An, Hongjun and Hu, Wenhan and Huang, Sida and Huang, Siqi and Li, Ruanjun and Liang, Yuanzhi and Shao, Jiawei and Song, Yiliang and Wang, Zihan and Yuan, Cheng and Zhang, Chi and Zhang, Hongyuan and Zhuang, Wenhao and Li, Xuelong , journal=. 2026 , publisher=

    2026

  53. [56]

    arXiv preprint arXiv:2606.27964 , year=

    Directing the World: Fast Autoregressive Video Generation with Compositional Human-Camera Control , author=. arXiv preprint arXiv:2606.27964 , year=

    Pith/arXiv arXiv

  54. [57]

    arXiv preprint arXiv:2605.25500 , year=

    Full-4D: Generating Full-Scope 4D Scenes from a Single-View Video , author=. arXiv preprint arXiv:2605.25500 , year=

    Pith/arXiv arXiv

  55. [58]

    arXiv preprint arXiv:2605.20290 , year=

    TelePhysics: Physics-Grounded Multi-Object Scene Generation from a Single Image with Real-Time Interaction , author=. arXiv preprint arXiv:2605.20290 , year=

    Pith/arXiv arXiv

  56. [59]

    arXiv preprint arXiv:2604.03723 , year=

    SymphoMotion: Joint Control of Camera Motion and Object Dynamics for Coherent Video Generation , author=. arXiv preprint arXiv:2604.03723 , year=

    Pith/arXiv arXiv

  57. [60]

    arXiv preprint arXiv:2412.09597 , year=

    LiftImage3D: Lifting any single image to 3D Gaussians with video generation priors , author=. arXiv preprint arXiv:2412.09597 , year=

    Pith/arXiv arXiv

  58. [61]

    European Conference on Computer Vision , pages=

    Cascade-zero123: One image to highly consistent 3d with self-prompted nearby views , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  59. [62]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    3DGabSplat: 3D Gabor Splatting for Frequency-adaptive Radiance Field Rendering , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  60. [63]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Phantom: Subject-consistent video generation via cross-modal alignment , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  61. [64]

    arXiv preprint arXiv:2508.11484 , year=

    Cinetrans: Learning to generate videos with cinematic transitions via masked diffusion models , author=. arXiv preprint arXiv:2508.11484 , year=

    arXiv

  62. [65]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    EchoShot: Multi-Shot Portrait Video Generation , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  63. [66]

    arXiv preprint arXiv:2510.20822 , year=

    Holocine: Holistic generation of cinematic multi-shot long video narratives , author=. arXiv preprint arXiv:2510.20822 , year=

    arXiv

  64. [67]

    arXiv preprint arXiv:2512.03041 , year=

    Multishotmaster: A controllable multi-shot video generation framework , author=. arXiv preprint arXiv:2512.03041 , year=

    arXiv

  65. [68]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Vace: All-in-one video creation and editing , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  66. [69]

    Nano Banana , author =. n.d. , note =

  67. [70]

    arXiv preprint arXiv:2307.06942 , year=

    Internvid: A large-scale video-text dataset for multimodal understanding and generation , author=. arXiv preprint arXiv:2307.06942 , year=

    Pith/arXiv arXiv

  68. [71]

    arXiv preprint arXiv:2410.17725 , year=

    Yolov11: An overview of the key architectural enhancements , author=. arXiv preprint arXiv:2410.17725 , year=

    Pith/arXiv arXiv

  69. [72]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Segment anything , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  70. [73]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    Pith/arXiv arXiv

  71. [74]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Transnet v2: An effective deep network architecture for fast shot transition detection , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  72. [75]

    arXiv preprint arXiv:2511.18870 , year=

    Hunyuanvideo 1.5 technical report , author=. arXiv preprint arXiv:2511.18870 , year=

    Pith/arXiv arXiv

  73. [76]

    arXiv preprint arXiv:2504.02436 , year=

    Skyreels-a2: Compose anything in video diffusion transformers , author=. arXiv preprint arXiv:2504.02436 , year=

    Pith/arXiv arXiv

  74. [77]

    arXiv preprint arXiv:2510.22200 , year=

    Longcat-video technical report , author=. arXiv preprint arXiv:2510.22200 , year=

    arXiv

  75. [78]

    arXiv preprint arXiv:2408.06072 , year=

    Cogvideox: Text-to-video diffusion models with an expert transformer , author =. arXiv preprint arXiv:2408.06072 , year=

    Pith/arXiv arXiv

  76. [79]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  77. [80]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    From slow bidirectional to fast autoregressive video diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  78. [81]

    arXiv preprint arXiv:2506.08009 , year=

    Self forcing: Bridging the train-test gap in autoregressive video diffusion , author=. arXiv preprint arXiv:2506.08009 , year=

    Pith/arXiv arXiv

  79. [82]

    arXiv preprint arXiv:2509.25161 , year=

    Rolling forcing: Autoregressive long video diffusion in real time , author=. arXiv preprint arXiv:2509.25161 , year=

    Pith/arXiv arXiv

  80. [83]

    arXiv preprint arXiv:2401.03048 , year=

    Latte: Latent diffusion transformer for video generation , author=. arXiv preprint arXiv:2401.03048 , year=

    Pith/arXiv arXiv

Showing first 80 references.