Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Diff-VF claims that a frozen short-video diffusion model can generate long, coherent, and dynamic videos with no training, fine-tuning, or architecture changes, by replacing only the noise, sampling, and fusion procedure.

desk verdict Sensible combination of known tricks, but the headline consistency-diversity balance is not supported by the HunyuanVideo numbers, and the evaluation lacks error bars. read the letter →

arxiv 2608.05976 v1 pith:VU2LPHB2 submitted 2026-08-06 cs.CV

Haoning Yang , Xinyuan Chen , Yaohui Wang , Guo Lu This is my paper · ORCID
classification cs.CV
keywords diffusionmodelsvideogenerationlongtraining-freemethodsnoiseinitializationtemporalcoherencemotiondiversityenhancement
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

Diff-VF claims that a frozen short-video diffusion model can be turned into a long-video generator without any training, fine-tuning, or architectural change—only the inference procedure is modified. The paper argues that long-video quality degrades for two opposing reasons that are usually traded off: over-strong temporal coupling yields static or repetitive videos, while weak coupling yields flicker and drift. Diff-VF couples three latent-space strategies—hybrid noise initialization, weighted window sampling, and temporal extended sampling—to keep global semantics anchored while allowing local motion to evolve. On VBench-Long it reports a better consistency-diversity balance than FreeNoise, FreeLong, and RIFLEx, across both a U-Net backbone with 2D+1D attention and a DiT backbone with coupled 3D attention.

What carries the argument

The load-bearing device is a latent-space local-global denoising loop. Hybrid Noise Initialization sets the noise of clip $n$ as $\boldsymbol{Z}^n_T = \sqrt{1-w}\,\boldsymbol{Z}^0_T + \sqrt{w}\,\boldsymbol{\epsilon}_n$ with $\boldsymbol{\epsilon}_n \sim \mathcal{N}(0, \boldsymbol{I})$, so every clip shares the first clip's structure but receives fresh stochastic content. Weighted Window Sampling denoises overlapping temporal windows and aggregates them with the triangular weight $m_{i,j} = (U+1)/2 - |C_j - i|$, which ramps each window's influence down toward its edges. Temporal Extended Sampling re-indexes the long latent into interleaved clips $\boldsymbol{Z}^n_{\text{extended}} = \boldsymbol{Z}_t[n + k N]$ so that the backbone's short-range temporal modeling is applied to frames $N$ steps apart, enlarging its effective temporal receptive field. The local and global paths are fused by $c(t) = \alpha\,(\tfrac12(1+\cos(\pi (T-t)/T)))^{c_s}$, letting global structure dominate early denoising steps and local detail dominate later ones; all operations sit in latent space, so no backbone module is modified.

What would settle it

Generate a long video whose text prompt contains an abrupt transition (e.g., 'a person walks through a door into a completely different room' or an explicit scene cut) and measure subject consistency and temporal flickering across the transition; a clear failure—frozen or flickering frames at the cut—would confirm the stationarity limit of HNI. A second check is sweeping $w$ from 0 to 1: at $w=0$ the video should become repetitive, at $w=1$ it should lose long-range coherence, and the reported favorable balance should peak at an intermediate value.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the consistency-diversity trade-off in training-free long video generation is governed by three independent control knobs, all operating on noisy latent sequences outside the backbone: a noise-mixing weight that sets how much of the first clip's noise is reused, a position-dependent window fusion that removes boundaries without uniform averaging, and a timestep-varying soft fusion between a local windowed denoising path and a global temporally-dilated denoising path. With these, a short-video model generates videos of 3–4 times its training length that preserve subject consistency and motion smoothness while keeping a dynamic degree close to the unconstrained base model—a combination the paper claims none of FreeNoise, FreeLong, or RIFLEx achieves. The method is presented as model-agnostic because it never touches attention modules, positional encodings, or model weights.

Load-bearing premise

Every later clip's noise is anchored to the first clip's noise, so the video content must be roughly temporally stationary: if the scene suddenly changes, cuts, or drifts non-stationarily, the shared noise anchor forces the generation toward either repetition or incoherence—a limitation the paper states explicitly.

Editorial extensions

If this is right

  • Long videos up to 4× the base model's length can be generated with subject consistency and motion smoothness comparable to or better than FreeNoise, FreeLong, and RIFLEx, while retaining a higher dynamic degree.
  • The same three-component recipe transfers across a U-Net backbone with 2D+1D attention and a DiT backbone with coupled 3D attention, with only hyperparameters such as $w$, $\alpha$, and $c_s$ re-tuned.
  • Adding Skip Residual Guidance makes the framework usable for long-video enhancement: the original video's latents are noised to match each timestep and softly blended in, improving fidelity without extra modules.
  • Hyperparameters $w$, $\alpha$, and $c_s$ act as explicit trade-off dials, so practitioners can favor coherence or diversity without changing the base model.

Reading between the lines

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

  • An implication the authors leave implicit: the mixing weight $w$ parameterizes a continuum of clip-to-clip noise correlation, so Diff-VF could be adapted to shot-based videos by choosing a small $w$ inside a shot and a larger $w$ at a cut—a simple testable extension to handle the stationarity limitation.
  • Temporal Extended Sampling is the temporal analogue of spatial receptive-field dilation used in high-resolution image generation; combining both dilations would likely yield a single training-free recipe for extending video models in both space and time.
  • A direct stress test of the stationarity assumption: prompts with abrupt scene transitions or quick camera cuts should break the method (freezing or flickering at the transition), which the paper itself flags as a limitation.
  • Since all operations are on latents, the framework is a candidate retrofit for any future short-video diffusion backbone, including large DiTs, without waiting for long-video training data.
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

3 major / 5 minor

Summary. The paper proposes Diff-VF, a training-free, plug-and-play framework that converts short-video diffusion backbones into long-video generators by operating only on latent noise sequences. It combines three strategies: Hybrid Noise Initialization (HNI), which mixes the first clip's noise with fresh Gaussian noise to anchor global semantics while allowing diversity; Weighted Window Sampling (WWS), which denoises overlapping windows and fuses them with position-dependent weights; and Temporal Extended Sampling (TES), which re-indexes the latent sequence into temporally dilated clips to establish long-range dependencies. A fourth component, Skip Residual Guidance, is proposed for long-video enhancement. The method is evaluated on LaVie and HunyuanVideo using VBench-Long, with comparisons against FreeNoise, FreeLong, and RIFLEx, and the paper claims a favorable balance between temporal coherence and motion diversity.

Significance. If substantiated, the framework is practically useful because it is model-agnostic, requires no training or fine-tuning, and does not modify internal attention modules, making it applicable to both U-Net and DiT backbones. The paper provides explicit equations for each component, component-wise ablations, and experiments on two architectures, including a 3D-attention DiT. However, the central trade-off claim currently rests on small-margin benchmark numbers without uncertainty estimates or a pre-specified balance criterion, and the HunyuanVideo results do not uniformly support the claim. The contribution is incremental but potentially publishable if the evidence is made robust.

major comments (3)
  1. [Section 4.2, Tables 1 and 2] The paper's headline claim is that Diff-VF achieves 'a more favorable balance between temporal coherence and motion diversity,' but no balance metric is defined or measured. In Table 2 (HunyuanVideo), Diff-VF obtains the lowest Subject Consistency (0.9594), Motion Smoothness (0.9920), and Temporal Flickering (0.9864) among all compared methods and is below FreeNoise on Imaging Quality (0.6292 vs 0.6646), with its only advantage being Dynamic Degree (0.4413). In Table 1 (LaVie), the difference between Ours (0.9164) and FreeNoise (0.9106) on Subject Consistency is only 0.006. Without a pre-specified criterion that combines these metrics, the claim that this is a 'more favorable balance' is not operational or testable. I recommend defining a balance score (e.g., a weighted combination or a Pareto-style criterion), reporting it for all methods, and explicitly discussing the HunyuanVideo results in light of that score.
  2. [Sections 4.1 and 4.4.2] All hyperparameters (w=0.1, alpha=0.4, c_s=6 on LaVie; w=0.2, alpha=0.5, c_s=6 on HunyuanVideo) are selected from ablations run on the same VBench-Long prompt sets used for the final comparisons, with no held-out prompt set and no repeated-seed statistics. Table 4 shows that changing alpha from 0.2 to 0.4 moves Subject Consistency by 0.0017 and Dynamic Degree by 0.0215; such differences are within the range of sampling noise for video generation. The paper should either report means and standard deviations over multiple seeds or use a held-out prompt set for hyperparameter selection and final evaluation, so that the reader can judge whether the reported margins are meaningful and not artifacts of in-sample tuning.
  3. [Section 5 (Limitations)] The limitation paragraph states that HNI 'assumes temporal stationarity of video content, so that our method may face challenges when generating videos with abrupt motions or scene cuts.' This is a structural premise, not a peripheral caveat, because Eq. (1) anchors every subsequent clip's initial noise to the first clip. If the target video contains a scene change, the shared noise anchor cannot represent the new scene, and the method will either repeat the old content or produce incoherent transitions. Given that the abstract and Section 1 claim a general training-free long-video generation framework, the paper should either add experiments with multi-prompt or scene-change content to characterize the failure mode, or explicitly restrict the scope of the central claim to temporally stationary videos.
minor comments (5)
  1. [Section 3.4] Skip Residual Guidance is evaluated only qualitatively (Fig. 6); no quantitative VBench numbers are provided for the long-video enhancement extension, so the claimed fidelity-realism balance is not measured.
  2. [Section 4.1] The paper should clarify whether the AccVideo acceleration is applied identically to all compared methods on HunyuanVideo, and how the accelerated sampler interacts with the comparison; otherwise the reader cannot assess whether the reported differences are affected by the acceleration strategy.
  3. [Section 3.3] For HunyuanVideo, TES is performed only for t < 900, but no ablation or sensitivity analysis of this cutoff is reported; since Eq. (9) already schedules the fusion weight, this additional manual restriction deserves explanation.
  4. [Equation (2)] The notation in Eq. (2) is unclear: the shuffled clip index n' is not defined, and the relationship between n and n' should be stated explicitly.
  5. [General presentation] There are minor typographical issues, including 'with4×extended' in Section 1 and 'ms-vid2vid-xl' in Section 4.1, which does not match the cited reference [49] (I2VGen-XL).

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: Diff-VF is an empirical latent-space heuristic evaluated on an external benchmark, and the in-sample hyperparameter selection weakens evidence without making the claim self-referential.

full rationale

Diff-VF's three components are defined by explicit latent-space operations, not by the outcomes they are claimed to produce. HNI constructs later noise clips as a mixture of the first clip's noise and fresh Gaussian noise (Eq. 1); WWS is a position-dependent weighted average of overlapping denoised windows (Eq. 5); TES is a temporal re-indexing and inverse-scatter operation (Eqs. 6-7) with a timestep-varying fusion (Eqs. 8-9). None of these equations takes consistency, diversity, or VBench scores as an input, and no fitted parameter is renamed as a prediction. The central claim is an empirical trade-off statement supported by direct measurements on the external VBench-Long benchmark (Tables 1 and 2), not by a derivation that presupposes the conclusion. The self-references in the paper, notably LaVie [41] as a base model, AccVideo [48] as an accelerator, and Seine [9] as related work, are tools or context rather than load-bearing evidence for Diff-VF's effectiveness; the comparison and ablation numbers come from running the method. The main legitimate concern is that the hyperparameters w=0.1, alpha=0.4, and cs=6 are selected from ablations on the same VBench-Long prompt sets used for the headline comparison (Sections 4.2 and 4.4.2), which makes the reported superiority partly in-sample. That is an evidence-quality and overfitting concern, not circularity: the reported metrics are independent external measurements, and the paper does not claim to predict them from the tuning procedure. Section 5's stated limitation that hybrid noise initialization assumes temporal stationarity of video content is an honest scope caveat and not a self-justifying premise. No circular step is identifiable.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

Diff-VF is a parameterized heuristic framework: w, alpha, c_s, c_r, U, s, N, and a TES activation threshold are chosen by hand or by tuning on VBench-Long. The core postulates are that noise mixing controls semantic correlation, that weighted overlap averaging produces smooth transitions, and that a frozen model denoising temporally dilated clips builds long-range dependencies. No new physical or mathematical entities are introduced.

free parameters (6)
  • w (noise mixing weight) = 0.1 for LaVie, 0.2 for HunyuanVideo, 0.3 for enhancement
    Controls the balance between global consistency and local diversity in Hybrid Noise Initialization; tuned on VBench-Long in Section 4.4.2.
  • alpha (TES fusion scale) = 0.4 for LaVie, 0.5 for HunyuanVideo, 1 for enhancement
    Scales the timestep-varying weight c(t) in Eq. 9; tuned on VBench-Long in Section 4.4.2.
  • c_s (schedule exponent) = 6 for generation, 5 for enhancement
    Shapes the cosine schedule of c(t); tuned on VBench-Long in Section 4.4.2.
  • c_r (skip residual schedule exponent) = not specified in text
    Appears in Eq. 10 for Skip Residual Guidance but no value is reported in Section 4.1; treated as a tuned but undisclosed parameter.
  • U, s, N (window size, stride, number of clips) = LaVie: U=16, s=8, N=4; HunyuanVideo: U=32, s=16, N=3; enhancement: U=32, s=16, N=2
    Chosen per base model and task in Section 4.1; these structural choices affect the trade-off and are not derived from first principles.
  • TES activation threshold = t < 900 for HunyuanVideo
    TES is disabled at early timesteps for HunyuanVideo to avoid semantic discrepancies, as stated in Section 4.1; this is an ad hoc choice.
assumptions (5)
  • domain assumption The first clip's noise determines the global semantics of the entire long video.
    Used in HNI, Eq. 1, and acknowledged as temporal stationarity in Section 5.
  • domain assumption Linear interpolation of latent noises with weight w yields semantically interpolated video content.
    Core to HNI; no proof is provided that mixing latents in noise space controls semantic correlation.
  • domain assumption A frozen short-video diffusion model denoising temporally dilated clips establishes dependencies among distant frames.
    The premise of TES in Section 3.3; the paper argues by analogy to receptive field dilation but provides no direct evidence.
  • domain assumption VBench-Long metrics are a faithful proxy for the claimed consistency-diversity trade-off.
    The whole evaluation rests on these metrics, used in Section 4.1, without calibration against human judgment in this paper.
  • domain assumption Weighted overlap averaging of latents (Eq. 5) produces temporally smooth video transitions.
    The central mechanism of WWS; plausible but not formally justified beyond the illustration in Fig. 3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model." pith.science (2026). https://pith.science/paper/VU2LPHB2

@misc{pith2026260805976,
  author       = {Pith},
  title        = {Pith review of: Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VU2LPHB2}},
  note         = {Machine review of arXiv:2608.05976}
}
read the original abstract

Recently, diffusion models have made great progress in video generation. However, most existing video diffusion models are trained with short videos, and degrade when extrapolated to long videos, struggling to maintain long-range temporal coherence while retaining diverse motions. To generate consistent, high-quality and dynamic long videos, we propose Diff-VF, a training-free, plug-and-play and model-agnostic framework that converts existing short-video diffusion backbones into long-video generators without modifying or fine-tuning the base model. Diff-VF couples three complementary strategies: Hybrid Noise Initialization (HNI) to constrain global semantics, Weighted Window Sampling (WWS) to remove inter-window discontinuities, and Temporal Extended Sampling (TES) to establish long-range dependencies with a timestep-varying fusion. We further extend Diff-VF to long-video enhancement via Skip Residual Guidance that balances fidelity and realism through timestep-dependent guidance. VBench-Long evaluation results show that Diff-VF achieves a more favorable balance between temporal coherence and motion diversity than base models and recent training-free long video generation baselines, including FreeNoise, FreeLong, and RIFLEx, while maintaining competitive frame-wise quality. Experiments on two base models demonstrate the applicability to video diffusion models with different spatial-temporal modeling strategies. Extensive ablations validate the contribution of each component and hyperparameters.

Figures

Figures reproduced from arXiv: 2608.05976 by the authors.

Figure 1
Figure 1. Illustration of long videos generated by our method based on HunyuanVideo. Both of them are [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of Diff-VF. We first build the initial noise through hybrid noise initialization, which replicates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Detailed illustration of simple fusion and weighted fusion. (a) If we simply calculate the average of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison based on LaVie. Videos generated by the base model LaVie [ [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison based on HunyuanVideo. Complete videos corresponding to the visual results [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of our long video enhancement method. Compared to original videos, the enhanced [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of ablation studies. Prompt: A dog running happily. (a) A video generated without [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: More qualitative comparison based on LaVie. Complete videos corresponding to the visual results in [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison based on HunyuanVideo (global view). Complete videos corresponding to the [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison based on HunyuanVideo (local view). Complete videos corresponding to the [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Illustration of long videos generated by our method based on CogVideoX1.5-5B. Complete videos [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Illustration of long videos generated by our method based on VideoCrafter2. Complete videos [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 38 canonical work pages

  1. [1]

    Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. 2021. Frozen in time: A joint video and image encoder for end-to-end retrieval. InProceedings of the IEEE/CVF international conference on computer vision. 1728–1738

    2021

  2. [2]

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. 2023. MultiDiffusion: fusing diffusion paths for controlled image generation. InProceedings of the 40th International Conference on Machine Learning. 1737–1752

    work page 2023

  3. [3]

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis

    work page

  4. [4]

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. 2024. Video generation models as world simulators. (2024). https://openai.com/research/video-generation-models-as-world-simulators

    2024

  5. [5]

    Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. 2022. Investigating tradeoffs in real-world video super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5962–5971

    work page 2022

  6. [6]

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. 2023. Videocrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512(2023)

    arXiv 2023

  7. [7]

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. 2024. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7310–7320

    work page 2024

  8. [8]

    Jingyuan Chen, Fuchen Long, Jie An, Zhaofan Qiu, Ting Yao, Jiebo Luo, and Tao Mei. 2025. Ouroboros-diffusion: Exploring consistent content generation in tuning-free long video diffusion. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 2079–2087

    work page 2025

Show all 53 references
  1. [9]

    Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. 2023. Seine: Short-to-long video diffusion model for generative transition and prediction. InThe Twelfth International Conference on Learning Representations

    2023

  2. [10]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34 (2021), 8780–8794

    2021

  3. [11]

    Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. 2024. Demofusion: Democratising high-resolution image generation with no $$$. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6159–6168

    2024

  4. [12]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks.Commun. ACM63, 11 (2020), 139–144

    2020

  5. [13]

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. 2024. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In International Conference on Learning Representations

    2024

  6. [14]

    Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu

  7. [15]

    Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. 2023. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In Diff-VF: Training-free High-quality Long Video Generati...

    2023

  8. [16]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  9. [17]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video diffusion models.Advances in Neural Information Processing Systems35 (2022), 8633–8646

    2022

  10. [18]

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. 2023. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. InInternational Conference on Learning Representations

    2023

  11. [19]

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision an...

    2024

  12. [20]

    Jianwen Jiang, Weihong Zeng, Zerong Zheng, Jiaqi Yang, Chao Liang, Wang Liao, Han Liang, Yuan Zhang, and Mingyuan Gao. 2025. Omnihuman-1.5: Instilling an active mind in avatars via cognitive simulation.arXiv preprint arXiv:2508.19209(2025)

    2025 arXiv

  13. [21]

    Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. 2024. Fifo-diffusion: Generating infinite videos from text without training.Advances in Neural Information Processing Systems37 (2024), 89834–89868

    2024

  14. [22]

    Yulhwa Kim, Dongwon Jo, Hyesung Jeon, Taesu Kim, Daehyun Ahn, Hyungjun Kim, and jae-joon kim. 2023. Leveraging Early-Stage Robustness in Diffusion Models for Efficient and High-Quality Image Synthesis. InAdvances in Neural Information Processing Systems, Vol. 36. 1229–1244

    2023

  15. [23]

    Diederik P Kingma, Max Welling, et al. 2013. Auto-encoding variational bayes

    2013

  16. [24]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. 2024. HunyuanVideo: A Systematic Framework For Large Video Generative Models.CoRR(2024)

    2024

  17. [25]

    Wenhao Li, Yichao Cao, Xiu Su, Xi Lin, Shan You, Mingkai Zheng, Yi Chen, and Chang Xu. 2024. Training-free Long Video Generation with Chain of Diffusion Model Experts.CoRR(2024)

    2024

  18. [26]

    Zhuoling Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. 2025. LongDiff: Training-Free Long Video Generation in One Go. InProceedings of the Computer Vision and Pattern Recognition Conference. 17789–17798

    2025

  19. [27]

    Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. 2024. Freelong: Training-free long video generation with spectralblend temporal attention.Advances in Neural Information Processing Systems37 (2024), 131434–131455

    2024

  20. [28]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4195–4205

    2023

  21. [29]

    Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. 2024. FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling. InThe Twelfth International Conference on Learning Representations

    2024

  22. [30]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  23. [31]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning. PMLR, 2256–2265

    2015

  24. [32]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising diffusion implicit models. InInternational Conference on Learning Representations

    2021

  25. [33]

    Xue Song, Jingjing Chen, Bin Zhu, and Yu-Gang Jiang. 2024. Text-Driven Video Prediction.ACM Trans. Multimedia Comput. Commun. Appl.20, 9, Article 296 (Sept. 2024), 15 pages. doi:10.1145/3675171

    2024 doi

  26. [34]

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

    2021

  27. [35]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing568 (2024), 127063

    2024

  28. [36]

    Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, et al. 2025. Wan: Open and Advanced Large-Scale Video Generative Models.CoRR(2025)

    2025

  29. [37]

    Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. 2023. Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising.CoRR(2023)

    2023

  30. [38]

    Hanzhang Wang, Haoran Wang, Zhongrui Yu, Mingming Sun, Junjun Jiang, Xianming Liu, and Deming Zhai. 2025. FAST: Flexibly Controllable Arbitrary Style Transfer via Latent Diffusion Models.ACM Trans. Multimedia Comput. Commun. Appl.21, 9, Article 268 (Sept. 2025), 20 pages. doi:...

    2025 doi

  31. [39]

    Xingrui Wang, Xin Li, and Zhibo Chen. 2024. CoNo: Consistency Noise Injection for Tuning-free Long Video Diffusion. CoRR(2024)

    2024

  32. [40]

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. 2021. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. InProceedings of the IEEE/CVF international conference on computer vision. 1905–1914. 18 Yang et al

    2021

  33. [41]

    Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. 2025. Lavie: High-quality video generation with cascaded latent diffusion models.International Journal of Computer Vision133, 5 (2025), 3059–3078

    2025

  34. [42]

    Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. 2023. Internvid: A large-scale video-text dataset for multimodal understanding and generation.arXiv preprint arXiv:2307.06942(2023)

    2023 arXiv

  35. [43]

    Chan, and Zhen Cui

    Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B. Chan, and Zhen Cui. 2024. Edit Temporal- Consistent Videos with Image Diffusion Model.ACM Trans. Multimedia Comput. Commun. Appl.20, 12, Article 368 (Nov. 2024), 16 pages. doi:10.1145/3691344

    2024 doi

  36. [44]

    Yuan Wang, Bin Zhu, Yanbin Hao, Chong-Wah Ngo, Yi Tan, and Xiang Wang. 2025. CookingDiffusion: Cooking Procedural Image Generation with Stable Diffusion.ACM Trans. Multimedia Comput. Commun. Appl.(Oct. 2025). doi:10.1145/3771995

    2025 doi

  37. [45]

    Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. 2024. Freeinit: Bridging initialization gap in video diffusion models. InEuropean Conference on Computer Vision. Springer, 378–394

    2024

  38. [46]

    Shaoshu Yang, Zhe Kong, Feng Gao, Meng Cheng, Xiangyu Liu, Yong Zhang, Zhuoliang Kang, Wenhan Luo, Xunliang Cai, Ran He, et al. 2025. Infinitetalk: Audio-driven video generation for sparse-frame video dubbing.arXiv preprint arXiv:2508.14033(2025)

    2025 arXiv

  39. [47]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. 2024. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. CoRR(2024)

    2024

  40. [48]

    Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 2025. Accvideo: Accelerating video diffusion model with synthetic dataset.arXiv preprint arXiv:2503.19462(2025)

    2025 arXiv

  41. [49]

    Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. 2023. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models.arXiv preprint arXiv:2311.04145(2023)

    2023 arXiv

  42. [50]

    Min Zhao, Guande He, Yixiao Chen, Hongzhou Zhu, Chongxuan Li, and Jun Zhu. 2025. RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers. InForty-second International Conference on Machine Learning

    2025

  43. [51]

    Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. 2024. Upscale-A-Video: Temporal- Consistent Diffusion Model for Real-World Video Super-Resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2535–2545. Diff-V...

    2024

  44. [2023]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Align your latents: High-resolution video synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22563–22575

  45. [2024]

    VEnhancer: Generative Space-Time Enhancement for Video Generation.CoRR(2024)

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.