Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

FreeLong++: Training-Free Long Video Generation via Multi-band SpectralFusion

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A training-free, multi-branch attention plugin can push short-video diffusion models to 4x-8x native length by fusing the frequency bands of differently sized attention windows.

desk verdict Practical training-free long-video method with solid empirical results, but the Nyquist-based frequency-band justification is mathematically off and needs a direct spectral test or a reframing. read the letter →

arxiv 2507.00162 v1 pith:B234Z4JQ submitted 2025-06-30 cs.CV

Yu Lu , Yi Yang This is my paper
classification cs.CV
keywords videogenerationdiffusiontransformerstraining-freeadaptationlongfrequency-domainfusiontemporalattentionspectralanalysisnoiseinitialization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Short-video diffusion models degrade when asked to generate longer clips, and this paper argues the cause is specifically a loss of high temporal frequencies: low-frequency content stays stable while fine detail and fast motion become distorted as the sequence grows beyond native length. The paper's proposal is FreeLong++, a training-free attention plugin that splits temporal attention into several branches with different window sizes, converts each branch's output to the frequency domain, keeps only the band of frequencies that branch is supposed to represent, and sums the filtered spectra before transforming back. The claim is that this Multi-band SpectralFusion lets pretrained models such as Wan2.1 and LTX-Video generate clips four to eight times their native length with temporal consistency and visual fidelity close to native clips, with no training and no added parameters. If the claim is right, long-video generation becomes an inference-time retrofit rather than a retraining problem, and it also extends to multi-prompt storytelling and long pose- or depth-guided control.

What carries the argument

The load-bearing mechanism is Multi-band SpectralFusion (MSF): temporal attention is masked into L parallel branches with window sizes $\alpha_l T_\alpha$ (e.g., $\alpha = 1,2,4$ for 4x, $\alpha = 1,2,4,8$ for 8x), each branch's latent is transformed with a 3D fast Fourier transform, multiplied by a scale-specific band-pass mask $P_l$, summed across branches, and transformed back. A second component, SpecMix, initializes the diffusion noise as a frequency-dependent blend between a shuffled consistency baseline and independent Gaussian noise, with the mixing angle growing from the sequence center outward. The band-pass masks are the part that carries the paper's frequency story: the upper edge of each branch's band is set by $\tfrac{1}{2\alpha_l}\pi$, with adjacent branches covering successive bands from low to high frequencies, so the most local branch supplies the highest frequencies and the most global branch supplies the lowest.

What would settle it

Compute the actual Fourier spectrum of a single attention branch's output under FreeLong++ while the input is a fixed long video latent, and check whether the measured passband matches the claimed band for that window size. If a branch with $\alpha_l = 4$ shows significant energy above $\tfrac{1}{8}\pi$, or if the filter bank works equally well after swapping which branch's output is assigned to which band, the frequency-alignment story is not what is doing the work.

Watch

Extended reading notes

Core claim

The central discovery claimed is that the failure mode of long video generation is visible in the frequency domain before it is visible in pixels: the signal-to-noise ratio of low-frequency components stays roughly constant as length grows (about 1.0 at native length to 0.97 at 8x for the models tested), while high-frequency SNR drops substantially (to about 0.6 at 8x). The paper further claims that this high-frequency distortion can be corrected by a multi-branch attention scheme in which each branch's temporal window length $\alpha_l T_\alpha$ determines a frequency band, and the bands are fused by 3D FFT, band-pass filtering, summation, and inverse FFT. The stated result is that this restores fine spatial-temporal detail while keeping global consistency, producing better scores than direct sampling, sliding-window attention, FreeNoise, and an autoregressive fine-tuned baseline, and supporting 4x and 8x extensions.

Load-bearing premise

The load-bearing premise is that a temporal attention window of length $\alpha_l T_\alpha$ actually determines the branch's maximum representable frequency, so a fixed frequency band can be assigned to each window size; if window length only changes frequency resolution rather than the passband, the multi-band fusion is not derived and the method's justification reduces to an arbitrary hand-designed filter bank.

Editorial extensions

If this is right

  • Any diffusion transformer that uses standard temporal attention can adopt FreeLong++ by replacing the attention module, so future short-video models should inherit the same length extension without retraining.
  • Long-form video generation from a single prompt becomes feasible at the same model size, since the extra cost is only in attention windows, not in parameters or fine-tuning.
  • Multi-prompt storytelling can be done by assigning prompts to segments and relying on the fused low-frequency backbone for smooth transitions, as demonstrated by the paper's scene-transition results.
  • Pose- or depth-guided video generation over hundreds of frames becomes more reliable, because the global branch preserves layout while local branches keep motion detail.
  • The frequency analysis implies that other training-free long-video methods should be judged by how well they protect high-frequency components, not only by boundary consistency.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not draw this conclusion, but its own numbers suggest the mid-frequency branches carry much of the gain: at 4x length only about 3% of frequency bands remain reliable with a two-branch design, so a test that removes only the $\alpha=2$ branch would show whether mid-range motion, not pure high-frequency detail, is the bottleneck.
  • Because the filter cutoffs are fixed constants ($D_0 = 0.25$ and band edges at $\tfrac{1}{8}\pi$ and $\tfrac{1}{4}\pi$), the method may be sensitive to the temporal statistics of the base model; an extension would derive cutoffs from each model's measured SNR curve instead of setting them by hand.
  • SpecMix and the spectral fusion are separate mechanisms, and the paper does not report a full $2\times2$ ablation of the two on the same backbone; a reader testing FreeLong++ with SpecMix removed and with a single fused branch would learn which component restores which metric.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes FreeLong++, a training-free framework for extending short-video diffusion models (Wan2.1, LTX-Video) to 4–8× longer clips. FreeLong++ replaces the standard temporal attention with multiple masked attention branches at different window sizes, transforms each branch output into the spatiotemporal frequency domain, and fuses them with band-pass masks P_l whose bands are supposedly determined by the window size via the Nyquist criterion. A spectral noise initialization (SpecMix) is also introduced. Experiments on VBench-Long report consistent improvements over direct sampling, sliding window, FreeNoise, and CausVid on consistency and fidelity metrics, plus demonstrations for multi-prompt and control-guided generation.

Significance. If the empirical results hold, the contribution is practically valuable: a training-free, parameter-free plug-in that improves long-video generation across two modern backbones, with ablations isolating the main components and a sparse-attention variant that reduces inference cost. The paper also makes a useful empirical observation about high-frequency degradation in longer videos. However, the central theoretical justification for the frequency-band assignment is mathematically unsupported, the SNR metric underlying the motivating analysis is undefined, and the quantitative claims lack error bars. The method may still work as a hand-designed filter bank, but the paper currently overstates the derivation of its core mechanism. These gaps are load-bearing and require substantive revision.

major comments (4)
  1. [§4.2.4 (Multi-band Spectral Fusion)] The statement "The temporal window α_l T_α for branch l determines its maximum frequency 1/(2α_l)π based on the Nyquist criterion" is not correct for the discrete-time latents actually processed. In a sequence of length T, the DFT frequencies always span [0, π] regardless of the attention window size; the window size changes the temporal support and frequency resolution of the branch output, not the Nyquist limit. Consequently the passbands [0, 1/8π], [1/8π, 1/4π], [1/4π, 1.0π] for α_l = 4, 2, 1 are not derived from the window sizes. Since this assignment is what the multi-band fusion equation (Z' = Σ_l P_l ⊙ F_3D(Z_l)) treats as "frequency-aligned," the central theoretical justification is unsupported. The method may still work as a heuristic filter bank, but the paper should either provide an empirical spectral analysis showing that each branch's energy actually concentrates in its assigned band (e.g., spectral plots of branch outputs before fusion) or explicitly downgrade the claim to a heuristic design with ablations over band splits.
  2. [§4.1.1 and Figures 2, 4 (Frequency analysis)] The Signal-to-Noise Ratio (SNR) used to motivate the high-frequency distortion is never defined. It is unclear what the "signal" and "noise" are (e.g., comparing generated long videos against what reference?), how the frequency bands are extracted, and what "relative SNR above 0.9" means. Without this definition, the central observation that high-frequency components degrade with length cannot be reproduced or quantitatively checked. Please provide a precise formula, the reference signal, and ideally release the analysis code.
  3. [Tables 1–3 (Quantitative comparison)] All quantitative results are reported as single numbers with no error bars, multiple seeds, or statistical significance tests. Video generation is stochastic, and the reported between-method differences (e.g., imaging quality 62.42 vs 64.93 in Table 1) may be within run-to-run noise. Report mean±std over at least three seeds for the main comparison and ablations, and state the number of prompts per metric (the text says 100 prompts for VBench-Long, which is fine).
  4. [§4.2.4 (Filter construction)] The band-pass masks P_l are not specified beyond "band-pass filter." For reproducibility, the paper must state whether P_l are hard masks or smooth (e.g., Gaussian) filters, how the cutoff frequencies are normalized (relative to the latent sampling rate, and with what roll-off), and the exact bands for the 4-branch 8× configuration (α_l = 1, 2, 4, 8). Currently the implementation details in §4.2.6 only mention window sizes and sparse attention, not the spectral masks.
minor comments (7)
  1. [Abstract] The phrase "extends FreeLong dual-branch design" should be "extends FreeLong's dual-branch design."
  2. [§4.2.6] The text contains a typo: "varing" should be "varying."
  3. [§4.2.4 (Rationale paragraph)] The rationale paragraph refers to "low-frequency global features (Z1)," but Z1 was defined earlier as the most local branch (l = 1); this should be Z_L or another clearly global label.
  4. [Figures 2 and 4] The figures lack axis labels and a full legend; the captions should define "relative SNR" and state the reference video length used to compute the ratios.
  5. [§4.1.4] The Gaussian Low Pass Filter cutoff D0 = 0.25 is not precisely defined; specify whether D0 is the standard deviation, the radius where the response falls to a given value, and how it is normalized (e.g., relative to the Nyquist frequency).
  6. [§4.2.5 (SpecMix)] In the SpecMix description, the text says "For each time index t" after transforming to the spectral domain; clarify whether t indexes temporal frequency bins or frame positions, since the subsequent low/high-frequency discussion depends on that distinction.
  7. [References] References [56] and [57] are reprints of Nyquist and Shannon; give the original publication years or use a consistent citation style for reprints.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FreeLong++ is an observation-motivated heuristic whose improvements are measured on independent VBench-Long metrics; the disputed Nyquist band assignment in §4.2.4 is a mathematical correctness issue, not a self-referential derivation.

full rationale

FreeLong++ does not fit any parameter to the metrics it later reports as predictions, and it does not invoke a load-bearing self-citation. The design chain starts from an SNR-based observation of high-frequency degradation in the base models (Figs. 2 and 4), and the proposed remedy—windowed attention branches combined by hand-designed spectral masks—is a heuristic response to that observation, not an equation whose output is its own input. The evaluation uses VBench-Long consistency and fidelity metrics that are external to the design-time SNR analysis, so the reported gains are independent evidence. The only mathematically load-bearing assertion, in §4.2.4, is 'The temporal window α_l T_α for branch l determines its maximum frequency 1/(2α_l)π based on the Nyquist criterion'; even if this application of Nyquist is unsupported for discrete-time latent features (window length affects resolution, not the passband upper bound), that is a correctness risk rather than a circular step, because the band masks are stipulated design choices and are not derived from the evaluation targets. The self-citations in the related-work list ([6], [7]) are contextual and carry no argumentative weight. Hence no step in the paper reduces by construction to its inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method relies on two hand-chosen filter settings (D0, α_l sets) and on two unproven measurement and band assumptions. No new physical entities are introduced.

free parameters (3)
  • Gaussian LPF cutoff D0 = 0.25
    Hand-set normalized spatiotemporal stop frequency for the low-pass filter in FreeLong (Sec. 4.1.4). Not fitted to data, but chosen by the authors and kept fixed across models.
  • Branch scale set α_l = [1,2,4] for 4x; [1,2,4,8] for 8x
    Hand-chosen window size multipliers for the multi-scale attention branches (Sec. 4.2.6). No selection criterion is given.
  • Number of branches L = 3 for 4x, 4 for 8x
    Set by the scale set α_l; increases with target length. Not data-derived.
assumptions (3)
  • domain assumption The SNR measure computed on generated videos correctly identifies high-frequency distortion as the primary degradation cause.
    Introduced in Sec. 4.1.1 without a formal definition of SNR; the entire motivation depends on this measurement being meaningful.
  • ad hoc to paper The temporal window length α_l T_α determines the maximum frequency of that branch through the Nyquist criterion, permitting the fixed band split in Sec. 4.2.4.
    This is not a standard signal processing result: the window size controls frequency resolution, not the highest resolvable frequency. The band assignment is therefore an unproven premise.
  • domain assumption The pretrained video diffusion models (Wan2.1, LTX-Video) can be treated as fixed black boxes whose intermediate features are compatible with the spectral fusion operation.
    The method replaces attention outputs with filtered, summed features without fine-tuning; the paper assumes the model's subsequent layers tolerate this distribution shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FreeLong++: Training-Free Long Video Generation via Multi-band SpectralFusion." pith.science (2026). https://pith.science/paper/B234Z4JQ

@misc{pith2026250700162,
  author       = {Pith},
  title        = {Pith review of: FreeLong++: Training-Free Long Video Generation via Multi-band SpectralFusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B234Z4JQ}},
  note         = {Machine review of arXiv:2507.00162}
}
read the original abstract

Recent advances in video generation models have enabled high-quality short video generation from text prompts. However, extending these models to longer videos remains a significant challenge, primarily due to degraded temporal consistency and visual fidelity. Our preliminary observations show that naively applying short-video generation models to longer sequences leads to noticeable quality degradation. Further analysis identifies a systematic trend where high-frequency components become increasingly distorted as video length grows, an issue we term high-frequency distortion. To address this, we propose FreeLong, a training-free framework designed to balance the frequency distribution of long video features during the denoising process. FreeLong achieves this by blending global low-frequency features, which capture holistic semantics across the full video, with local high-frequency features extracted from short temporal windows to preserve fine details. Building on this, FreeLong++ extends FreeLong dual-branch design into a multi-branch architecture with multiple attention branches, each operating at a distinct temporal scale. By arranging multiple window sizes from global to local, FreeLong++ enables multi-band frequency fusion from low to high frequencies, ensuring both semantic continuity and fine-grained motion dynamics across longer video sequences. Without any additional training, FreeLong++ can be plugged into existing video generation models (e.g. Wan2.1 and LTX-Video) to produce longer videos with substantially improved temporal consistency and visual fidelity. We demonstrate that our approach outperforms previous methods on longer video generation tasks (e.g. 4x and 8x of native length). It also supports coherent multi-prompt video generation with smooth scene transitions and enables controllable video generation using long depth or pose sequences.

Figures

Figures reproduced from arXiv: 2507.00162 by the authors.

Figure 1
Figure 1. Results of Short and Longer Videos. The first row of each case shows short videos generated using short video [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Ratio of short video SNR on high (0.25π￾1.0π)/low (0.0π-0.25π) frequency to longer videos. Our findings reveal that when direct extend short video diffusion model to generate longer videos, the SNR of high-frequency components in the space-time frequency domain degrades significantly as video length increases. ing low-frequency global components with high-frequency local components. Building on the FreeLong, we furt… view at source ↗
Figure 4
Figure 4. Fine-grained frequency analysis on longer video generation. (a) As video length increases, both the range and severity of frequency distortion grow substantially. (b) We define available frequency bands as those with a relative SNR above 0.9. As shown, the number of available bands drops significantly when the video length increases from 2× to 4×, indicating that a fixed two-branch structure in FreeLong is insuffici… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Overview of FreeLong. FreeLong facilitates consistent and high-fidelity video generation using SpectralBlend Attention. SpectralBlend effectively blends low-frequency global video features with high-frequency local video features through a two-step process: local-globa…
Figure 6
Figure 6. Figure 6: Overview of FreeLong++. The FreeLong++ framework extends FreeLong by introducing Multi-band SpectralFusion Attention. Multi-scale temporal branches with varying window sizes capture motion dynamics at different frequency bands. Each branch is processed in the frequency…
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Results of Multi-Prompt Video Generation. Our method ensures coherent visual continuity and motion consistency across different video segments. dynamics. 5.5 Multi-Prompt Video Generation Our method easily extends to multi-prompt video genera￾tion by assigning distinct…
Figure 9
Figure 9. Figure 9 [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Spectral Self-Anchoring fuses low-frequency anchor attention with high-frequency local attention to stop autoregressive video collapse, enabling 24× length extrapolation without retraining.

  2. Surprise Forcing: What to Remember, When to Skip in Long Video Generation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A training-free 'surprise' controller decides which old frames to keep in memory and which chunks need fewer denoising steps, improving long-video consistency at real-time speed.

Reference graph

Works this paper leans on

70 extracted references · 24 canonical work pages · cited by 2 Pith papers

  1. [1]

    Wan: Open and advanced large-scale video generative models,

    WanTeam, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, J. Wang, J. Zhang, J. Zhou, J. Wang, J. Chen, K. Zhu, K. Zhao, K. Yan, L. Huang, M. Feng, N. Zhang, P . Li, P . Wu, R. Chu, R. Feng, S. Zhang, S. Sun, T. Fang, T. Wang, T. Gui, T. Weng, T. Shen, W. Lin, W. Wang, W. Wang, W. Zhou, W. Wang, W. Shen, W. Yu, X. Shi,...

    arXiv 2025

  2. [2]

    Ltx-video: Realtime video latent diffusion,

    Y. HaCohen, N. Chiprut, B. Brazowski, D. Shalem, D. Moshe, E. Richardson, E. Levin, G. Shiran, N. Zabari, O. Gordon, P . Panet, S. Weissbuch, V . Kulikov, Y. Bitterman, Z. Melumian, and O. Bibi, “Ltx-video: Realtime video latent diffusion,”arXiv preprint arXiv:2501.00103, 2024

    arXiv 2024

  3. [3]

    Videocrafter2: Overcoming data limitations for high- quality video diffusion models,

    H. Chen, Y. Zhang, X. Cun, M. Xia, X. Wang, C. Weng, and Y. Shan, “Videocrafter2: Overcoming data limitations for high- quality video diffusion models,” 2024

    2024

  4. [4]

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

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y. Yang, W. Hong, X. Zhang, G. Feng et al. , “Cogvideox: Text-to-video diffusion models with an expert transformer,” arXiv preprint arXiv:2408.06072, 2024

    arXiv 2024

  5. [5]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning,

    Y. Guo, C. Yang, A. Rao, Y. Wang, Y. Qiao, D. Lin, and B. Dai, “Animatediff: Animate your personalized text-to-image diffusion models without specific tuning,” arXiv preprint arXiv:2307.04725 , 2023

    arXiv 2023

  6. [6]

    FlowZero: Zero-Shot Text-to-Video Synthesis with LLM-Driven Dynamic Scene Syntax

    Y. Lu, L. Zhu, H. Fan, and Y. Yang, “Flowzero: Zero-shot text- to-video synthesis with llm-driven dynamic scene syntax,” arXiv preprint arXiv:2311.15813, 2023

    work page Pith review arXiv 2023

  7. [7]

    Eva: Zero-shot accu- rate attributes and multi-object video editing,

    X. Yang, L. Zhu, H. Fan, and Y. Yang, “Eva: Zero-shot accu- rate attributes and multi-object video editing,” arXiv preprint arXiv:2403.16111, 2024

    arXiv 2024

  8. [8]

    Hunyuanvideo: A systematic framework for large video generative models,

    Z. Z. e. Weijie Kong, Qi Tian, “Hunyuanvideo: A systematic framework for large video generative models,” 2024. [Online]. Available: https://arxiv.org/abs/2412.03603

    arXiv 2024

Show all 70 references
  1. [9]

    Step-video-t2v technical report: The practice, challenges, and future of video foundation model,

    G. Ma, H. Huang, K. Yan, L. Chen, N. Duan, S. Yin, C. Wan, R. Ming, X. Song, X. Chen et al., “Step-video-t2v technical report: The practice, challenges, and future of video foundation model,” arXiv preprint arXiv:2502.10248, 2025

    2025 arXiv

  2. [10]

    Cosmos world foundation model platform for physical ai,

    N. Agarwal, A. Ali, M. Bala, Y. Balaji, E. Barker, T. Cai, P . Chattopadhyay, Y. Chen, Y. Cui, Y. Ding et al. , “Cosmos world foundation model platform for physical ai,” arXiv preprint arXiv:2501.03575, 2025

    2025 arXiv

  3. [11]

    Movie gen: A cast of media foundation models,

    A. Polyak, A. Zohar, A. Brown, A. Tjandra, A. Sinha, A. Lee, A. Vyas, B. Shi, C.-Y. Ma, C.-Y. Chuanget al., “Movie gen: A cast of media foundation models,” arXiv preprint arXiv:2410.13720, 2024

    2024 arXiv

  4. [12]

    Pyramidal flow matching for efficient video generative modeling,

    Y. Jin, Z. Sun, N. Li, K. Xu, H. Jiang, N. Zhuang, Q. Huang, Y. Song, Y. Mu, and Z. Lin, “Pyramidal flow matching for efficient video generative modeling,” arXiv preprint arXiv:2410.05954, 2024

    2024

  5. [13]

    Open-sora plan: Open-source large video generation model,

    B. Lin, Y. Ge, X. Cheng, Z. Li, B. Zhu, S. Wang, X. He, Y. Ye, S. Yuan, L. Chen et al., “Open-sora plan: Open-source large video generation model,” arXiv preprint arXiv:2412.00131, 2024

    2024 arXiv

  6. [14]

    Stable video IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 13 diffusion: Scaling latent video diffusion models to large datasets,

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y. Levi, Z. English, V . Voleti, A. Lettset al., “Stable video IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 13 diffusion: Scaling latent video diffusion models to large datasets,” ar...

    2023 arXiv

  7. [15]

    Internvid: A large-scale video-text dataset for multimodal understanding and generation,

    Y. Wang, Y. He, Y. Li, K. Li, J. Yu, X. Ma, X. Li, G. Chen, X. Chen, Y. Wanget al., “Internvid: A large-scale video-text dataset for multimodal understanding and generation,” in The Twelfth International Conference on Learning Representations, 2023

    2023

  8. [16]

    Frozen in time: A joint video and image encoder for end-to-end retrieval,

    M. Bain, A. Nagrani, G. Varol, and A. Zisserman, “Frozen in time: A joint video and image encoder for end-to-end retrieval,” in IEEE International Conference on Computer Vision, 2021

    2021

  9. [17]

    Panda-70m: Captioning 70m videos with multiple cross-modality teachers,

    T.-S. Chen, A. Siarohin, W. Menapace, E. Deyneka, H.-w. Chao, B. E. Jeon, Y. Fang, H.-Y. Lee, J. Ren, M.-H. Yanget al., “Panda-70m: Captioning 70m videos with multiple cross-modality teachers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2024

  10. [18]

    Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content,

    Q. Wang, Y. Shi, J. Ou, R. Chen, K. Lin, J. Wang, B. Jiang, H. Yang, M. Zheng, X. Tao et al. , “Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content,” in Proceedings of the Computer Vision and Pattern Recogni- tion Conf...

    2025

  11. [19]

    Vidgen-1m: A large- scale dataset for text-to-video generation,

    Z. Tan, X. Yang, L. Qin, and H. Li, “Vidgen-1m: A large- scale dataset for text-to-video generation,” arXiv preprint arXiv:2408.02629, 2024

    2024 arXiv

  12. [20]

    Hoigen-1m: A large-scale dataset for human-object interaction video generation,

    K. Liu, Q. Liu, X. Liu, J. Li, Y. Zhang, J. Luo, X. He, and W. Liu, “Hoigen-1m: A large-scale dataset for human-object interaction video generation,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 24 001–24 010

    2025

  13. [21]

    Egovid-5m: A large-scale video- action dataset for egocentric video generation,

    X. Wang, K. Zhao, F. Liu, J. Wang, G. Zhao, X. Bao, Z. Zhu, Y. Zhang, and X. Wang, “Egovid-5m: A large-scale video- action dataset for egocentric video generation,” arXiv preprint arXiv:2411.08380, 2024

    2024 arXiv

  14. [22]

    Stream- ingt2v: Consistent, dynamic, and extendable long video genera- tion from text,

    R. Henschel, L. Khachatryan, D. Hayrapetyan, H. Poghosyan, V . Tadevosyan, Z. Wang, S. Navasardyan, and H. Shi, “Stream- ingt2v: Consistent, dynamic, and extendable long video genera- tion from text,” arXiv preprint arXiv:2403.14773, 2024

    2024 arXiv

  15. [23]

    Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models,

    F. Bao, C. Xiang, G. Yue, G. He, H. Zhu, K. Zheng, M. Zhao, S. Liu, Y. Wang, and J. Zhu, “Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models,” arXiv preprint arXiv:2405.04233, 2024

    2024 arXiv

  16. [24]

    Videotetris: Towards compositional text-to- video generation,

    Y. Tian, L. Yang, H. Yang, Y. Gao, Y. Deng, J. Chen, X. Wang, Z. Yu, X. Tao, P . Wanet al., “Videotetris: Towards compositional text-to- video generation,” arXiv preprint arXiv:2406.04277, 2024

    2024 arXiv

  17. [25]

    Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation,

    W. Wang, H. Yang, Z. Tuo, H. He, J. Zhu, J. Fu, and J. Liu, “Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation,” arXiv preprint arXiv:2305.10874, 2023

    2023 arXiv

  18. [26]

    Vlogger: Make your dream a vlog,

    S. Zhuang, K. Li, X. Chen, Y. Wang, Z. Liu, Y. Qiao, and Y. Wang, “Vlogger: Make your dream a vlog,” arXiv preprint arXiv:2401.09414, 2024

    2024 arXiv

  19. [27]

    Packing input frame contexts in next- frame prediction models for video generation,

    L. Zhang and M. Agrawala, “Packing input frame contexts in next- frame prediction models for video generation,” Arxiv, 2025

    2025

  20. [28]

    Long-context autoregres- sive video modeling with next-frame prediction,

    Y. Gu, W. Mao, and M. Z. Shou, “Long-context autoregres- sive video modeling with next-frame prediction,” arXiv preprint arXiv:2503.19325, 2025

    2025 arXiv

  21. [29]

    Long context tuning for video generation,

    Y. Guo, C. Yang, Z. Yang, Z. Ma, Z. Lin, Z. Yang, D. Lin, and L. Jiang, “Long context tuning for video generation,”arXiv preprint arXiv:2503.10589, 2025

    2025 arXiv

  22. [30]

    One-minute video generation with test-time training,

    K. Dalal, D. Koceja, J. Xu, Y. Zhao, S. Han, K. C. Cheung, J. Kautz, Y. Choi, Y. Sun, and X. Wang, “One-minute video generation with test-time training,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 702–17 711

    2025

  23. [31]

    Vide- orag: Retrieval-augmented generation with extreme long-context videos,

    X. Ren, L. Xu, L. Xia, S. Wang, D. Yin, and C. Huang, “Vide- orag: Retrieval-augmented generation with extreme long-context videos,” arXiv preprint arXiv:2502.01549, 2025

    2025 arXiv

  24. [32]

    Videoauteur: Towards long narrative video generation,

    J. Xiao, F. Cheng, L. Qi, L. Gui, J. Cen, Z. Ma, A. Yuille, and L. Jiang, “Videoauteur: Towards long narrative video generation,” arXiv preprint arXiv:2501.06173, 2025

    2025 arXiv

  25. [33]

    Owl-1: Omni world model for consistent long video generation,

    Y. Huang, W. Zheng, Y. Gao, X. Tao, P . Wan, D. Zhang, J. Zhou, and J. Lu, “Owl-1: Omni world model for consistent long video generation,” arXiv preprint arXiv:2412.09600, 2024

    2024 arXiv

  26. [34]

    Freenoise: Tuning-free longer video diffusion via noise reschedul- ing,

    H. Qiu, M. Xia, Y. Zhang, Y. He, X. Wang, Y. Shan, and Z. Liu, “Freenoise: Tuning-free longer video diffusion via noise reschedul- ing,” arXiv preprint arXiv:2310.15169, 2023

    2023 arXiv

  27. [35]

    Gen- l-video: Multi-text to long video generation via temporal co- denoising,

    F.-Y. Wang, W. Chen, G. Song, H.-J. Ye, Y. Liu, and H. Li, “Gen- l-video: Multi-text to long video generation via temporal co- denoising,” arXiv preprint arXiv:2305.18264, 2023

    2023 arXiv

  28. [36]

    Fifo-diffusion: Generating infinite videos from text without training,

    J. Kim, J. Kang, J. Choi, and B. Han, “Fifo-diffusion: Generating infinite videos from text without training,” in NeurIPS, 2024

    2024

  29. [37]

    Vstar: Generative temporal nursing for longer dynamic video synthesis,

    Y. Li, W. Beluch, M. Keuper, D. Zhang, and A. Khoreva, “Vstar: Generative temporal nursing for longer dynamic video synthesis,” arXiv preprint arXiv:2403.13501, 2024

    2024 arXiv

  30. [38]

    Riflex: A free lunch for length extrapolation in video diffusion transformers,

    M. Zhao, G. He, Y. Chen, H. Zhu, C. Li, and J. Zhu, “Riflex: A free lunch for length extrapolation in video diffusion transformers,” arXiv preprint arXiv:2502.15894, 2025

    2025 arXiv

  31. [39]

    Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation,

    M. Cai, X. Cun, X. Li, W. Liu, Z. Zhang, Y. Zhang, Y. Shan, and X. Yue, “Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation,” in Proceedings of the Computer Vision and Pattern Recog- nition Conference,...

    2025

  32. [40]

    Freepca: Integrating consistency information across long-short frames in training-free long video generation via principal component analysis,

    J. Tan, H. Yu, J. Huang, J. Xiao, and F. Zhao, “Freepca: Integrating consistency information across long-short frames in training-free long video generation via principal component analysis,” in Pro- ceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp...

    2025

  33. [41]

    Longdiff: Training-free long video generation in one go,

    Z. Li, H. Rahmani, Q. Ke, and J. Liu, “Longdiff: Training-free long video generation in one go,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 789–17 798

    2025

  34. [42]

    Scalingnoise: Scaling inference-time search for generating infinite videos,

    H. Yang, F. Tang, M. Hu, Q. Yin, Y. Li, Y. Liu, Z. Peng, P . Gao, J. He, Z. Ge et al., “Scalingnoise: Scaling inference-time search for generating infinite videos,” arXiv preprint arXiv:2503.16400, 2025

    2025 arXiv

  35. [43]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  36. [44]

    Lavie: High-quality video gen- eration with cascaded latent diffusion models,

    Y. Wang, X. Chen, X. Ma, S. Zhou, Z. Huang, Y. Wang, C. Yang, Y. He, J. Yu, P . Yang et al. , “Lavie: High-quality video gen- eration with cascaded latent diffusion models,” arXiv preprint arXiv:2309.15103, 2023

    2023 arXiv

  37. [45]

    Video generation models as world simulators,

    T. Brooks, B. Peebles, C. Holmes, W. DePue, Y. Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhman, C. Ng, R. Wang, and A. Ramesh, “Video generation models as world simulators,” 2024, accessed: 2024-05-09. [Online]. Available: https://openai. com/research/video-generatio...

    2024

  38. [46]

    Scalable diffusion models with trans- formers,

    W. Peebles and S. Xie, “Scalable diffusion models with trans- formers,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4195–4205

    2023

  39. [47]

    Mochi 1,

    G. Team, “Mochi 1,” https://github.com/genmoai/models, 2024

    2024

  40. [48]

    From slow bidirectional to fast autoregres- sive video diffusion models,

    T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shecht- man, and X. Huang, “From slow bidirectional to fast autoregres- sive video diffusion models,” in CVPR, 2025

    2025

  41. [49]

    Magi-1: Autoregressive video generation at scale,

    S. ai, H. Teng, H. Jia, L. Sun, L. Li, M. Li, M. Tang, S. Han, T. Zhang, W. Q. Zhang, W. Luo, X. Kang, Y. Sun, Y. Cao, Y. Huang, Y. Lin, Y. Fang, Z. Tao, Z. Zhang, Z. Wang, Z. Liu, D. Shi, G. Su, H. Sun, H. Pan, J. Wang, J. Sheng, M. Cui, M. Hu, M. Yan, S. Yin, S. Zhang, T. Li...

    2025 arXiv

  42. [50]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  43. [51]

    Flow matching for generative modeling,

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  44. [52]

    Spargeattn: Accurate sparse attention accelerating any model inference,

    J. Zhang, C. Xiang, H. Huang, J. Wei, H. Xi, J. Zhu, and J. Chen, “Spargeattn: Accurate sparse attention accelerating any model inference,” arXiv preprint arXiv:2502.18137, 2025

    2025

  45. [53]

    Flashvideo: Flowing fidelity to detail for efficient high-resolution video generation,

    S. Zhang, W. Li, S. Chen, C. Ge, P . Sun, Y. Zhang, Y. Jiang, Z. Yuan, B. Peng, and P . Luo, “Flashvideo: Flowing fidelity to detail for efficient high-resolution video generation,”arXiv preprint arXiv:2502.05179, 2025

    2025

  46. [54]

    Fast video generation with sliding tile attention,

    P . Zhang, Y. Chen, R. Su, H. Ding, I. Stoica, Z. Liu, and H. Zhang, “Fast video generation with sliding tile attention,” arXiv preprint arXiv:2502.04507, 2025

    2025 arXiv

  47. [55]

    Sparse videogen: Accelerating video dif- fusion transformers with spatial-temporal sparsity,

    H. Xi, S. Yang, Y. Zhao, C. Xu, M. Li, X. Li, Y. Lin, H. Cai, J. Zhang, D. Li et al. , “Sparse videogen: Accelerating video dif- fusion transformers with spatial-temporal sparsity,” arXiv preprint arXiv:2502.01776, 2025

    2025 arXiv

  48. [56]

    Certain topics in telegraph transmission theory,

    H. Nyquist, “Certain topics in telegraph transmission theory,” Transactions of the American Institute of Electrical Engineers , vol. 47, no. 2, pp. 617–644, 2009

    2009

  49. [57]

    Communication in the presence of noise,

    C. E. Shannon, “Communication in the presence of noise,” Proceed- ings of the IRE, vol. 37, no. 1, pp. 10–21, 2006

    2006

  50. [58]

    Freeinit: Bridg- ing initialization gap in video diffusion models,

    T. Wu, C. Si, Y. Jiang, Z. Huang, and Z. Liu, “Freeinit: Bridg- ing initialization gap in video diffusion models,” arXiv preprint arXiv:2312.07537, 2023. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 14

    2023 arXiv

  51. [59]

    Free- bloom: Zero-shot text-to-video generator with llm director and ldm animator,

    H. Huang, Y. Feng, C. Shi, L. Xu, J. Yu, and S. Yang, “Free- bloom: Zero-shot text-to-video generator with llm director and ldm animator,” Advances in Neural Information Processing Systems , vol. 36, pp. 26 135–26 158, 2023

    2023

  52. [60]

    FlashAttention-2: Faster attention with better parallelism and work partitioning,

    T. Dao, “FlashAttention-2: Faster attention with better parallelism and work partitioning,” in International Conference on Learning Representations (ICLR), 2024

    2024

  53. [61]

    Vbench: Comprehensive benchmark suite for video generative models,

    Z. Huang, Y. He, J. Yu, F. Zhang, C. Si, Y. Jiang, Y. Zhang, T. Wu, Q. Jin, N. Chanpaisit et al., “Vbench: Comprehensive benchmark suite for video generative models,”arXiv preprint arXiv:2311.17982, 2023

    2023 arXiv

  54. [62]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P . Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF international confer- ence on computer vision, 2021, pp. 9650–9660

    2021

  55. [63]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clark et al. , “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

    2021

  56. [64]

    Amt: All-pairs multi-field transforms for efficient frame inter- polation,

    Z. Li, Z.-L. Zhu, L.-H. Han, Q. Hou, C.-L. Guo, and M.-M. Cheng, “Amt: All-pairs multi-field transforms for efficient frame inter- polation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9801–9810

    2023

  57. [65]

    Musiq: Multi- scale image quality transformer,

    J. Ke, Q. Wang, Y. Wang, P . Milanfar, and F. Yang, “Musiq: Multi- scale image quality transformer,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 5148–5157

    2021

  58. [66]

    Perceptual quality assessment of smartphone photography,

    Y. Fang, H. Zhu, Y. Zeng, K. Ma, and Z. Wang, “Perceptual quality assessment of smartphone photography,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 3677–3686

    2020

  59. [67]

    aesthetic-predictor,

    LAION-AI, “aesthetic-predictor,” 2024, accessed: 2025- 06-04. [Online]. Available: https://github.com/LAION-AI/ aesthetic-predictor

    2024

  60. [68]

    Vace: All- in-one video creation and editing,

    Z. Jiang, Z. Han, C. Mao, J. Zhang, Y. Pan, and Y. Liu, “Vace: All- in-one video creation and editing,”arXiv preprint arXiv:2503.07598, 2025

    2025 arXiv

  61. [69]

    Kling, “Kling,” https://kling.kuaishou.com/en, 2025, accessed: 2025-06-06, 11, 13

    2025

  62. [70]

    Pika.art,

    Pika.art, “Pika.art,” https://pika.art, 2025, accessed: 2025-06-06

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.