Pith. sign in

REVIEW 3 major objections 4 minor 60 references

SPEED is a one-step pixel diffusion model for video frame interpolation that claims to beat latent-diffusion baselines on perceptual quality while cutting inference time and memory use substantially.

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 22:50 UTC pith:VRAM5GG7

load-bearing objection Strong, well-ablated system paper: the one-step boundary extrapolation is a real missing analysis rather than a fatal flaw, but missing code/error bars and an undisclosed self-comparison need referee attention. the 3 major comments →

arxiv 2607.15585 v1 pith:VRAM5GG7 submitted 2026-07-17 cs.MM

SPEED: One-Step Pixel Diffusion for High-quality Video Frame Interpolation

Zihao Zhang , Haoyu Zhao , Siqian Yang , Yidi Wu , Yudong Jiang , Zuxuan Wu This is my paper
classification cs.MM
keywords video frame interpolationpixel-space diffusionone-step inferencedrift-aware timestep samplingnoise-update-only attentionprogressive patch scalinglarge-motion interpolation4K video generation
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.

The paper tries to establish that video frame interpolation can be done by a diffusion model operating directly on pixels in a single forward pass, without sacrificing the quality that multi-step latent diffusion models provide. The proposed system, SPEED, predicts the clean intermediate frame from pure Gaussian noise at the maximum noise timestep, conditioned on the two boundary frames. It argues that latent diffusion's compressed representation is a hard ceiling on fine detail, and that multi-step sampling is what makes diffusion-based interpolation slow and memory-hungry. On standard benchmarks it reports lower LPIPS than previous diffusion models while running in tens of milliseconds and using less memory, and on 4K inputs it reports large perceptual gains. The sympathetic reading is that the combination of a progressive patch-size architecture, an asymmetric attention that keeps condition frames clean, and a timestep-sampling curriculum is what makes one-step pixel diffusion feasible.

Core claim

On its own terms, SPEED establishes a recipe for one-step pixel-space diffusion in video frame interpolation. Instead of encoding the input frames into a VAE latent space and iteratively denoising, SPEED feeds the two clean boundary frames and a pure Gaussian noise tensor, labeled as timestep t=1, through a transformer that directly outputs the interpolated frame. Three designs carry the claim: a progressive multi-stage architecture whose token patch size shrinks from 64 to 32 to 16, so the network first models global motion and then refines texture; a Noise-Update-Only attention where queries come only from the noisy target tokens while keys and values span all three frames, giving a global

What carries the argument

Progressive multi-stage pixel-space transformer: patch size decreases 64 to 32 to 16 across three stages to separately model motion, structure, and texture. Noise-Update-Only (NUO) attention: an asymmetric attention where the query comes only from the noisy intermediate tokens and keys and values come from all three frames, reducing complexity from O(9N^2) to O(3N^2) while preventing noise leakage into the clean condition frames. Drift-aware Timestep Sampling (DTS): a curriculum that maps uniformly sampled timestep t to t' = s(p)t / (1 + (s(p)-1)t) with s(p) = 1/(1-p), pushing training mass toward t=1 as progress p approaches 1, so the model learns one-step projection from pure noise. x0 pre

Load-bearing premise

The load-bearing assumption is that the network, trained on timesteps strictly below 1, extrapolates smoothly to the t=1 boundary used at inference, so that a single forward pass from pure Gaussian noise lands on the same image manifold that multi-step denoising would have reached.

What would settle it

The decisive check is to run the trained model at t=0.999, t=0.9999, and t=1.0 with the same Gaussian noise draw and compare output LPIPS; a jump larger than about 0.02 between the last trained timestep and t=1.0, or high variance across noise draws at t=1.0, would show the boundary extrapolation is not smooth. A second check is to retrain with the DTS progress capped at p=0.99 so that timesteps never reach t=1 and then test one-step inference; an LPIPS jump would show the curriculum's final stage, not the architecture, is carrying the result.

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

If this is right

  • If one-step pixel diffusion works as claimed, diffusion-based video frame interpolation no longer needs a VAE decoder, removing the reconstruction upper bound that blurs micro-textures in latent methods.
  • One-step inference makes near-real-time interpolation practical: the paper reports 36.6 ms per frame on DAVIS and 959.4 ms on 4K, with memory below 7.3GB at 4K.
  • Additional sampling steps become unnecessary: the paper's ablation shows 1-step and 10-step outputs are nearly identical when DTS is active.
  • Recursive multi-frame interpolation remains stable, suggesting the one-step projection does not accumulate errors across synthesized frames.
  • The method reports better perceptual quality than optical-flow-based specialists at 4K, indicating pixel-space diffusion can handle large non-linear motion without explicit flow estimation.

Where Pith is reading between the lines

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

  • The DTS curriculum is not tied to video frame interpolation: any conditional diffusion task with strong input conditioning—such as super-resolution, depth estimation, or video prediction—could use the same drift toward t=1 to obtain one-step inference without adversarial distillation, though the paper does not test this.
  • The x0-prediction advantage appears specific to pixel space: the paper's own latent-space experiment shows only marginal gains from x0 prediction inside a VAE latent, suggesting future one-step diffusion efforts should focus on raw-pixel models rather than latent ones.
  • The 4K results imply that resolution limits in diffusion models may come more from latent reconstruction than from transformer capacity; if so, progressive patch scaling could be a general route to higher-resolution generation.
  • One-step pixel-space interpolation is a natural testbed for temporal consistency in generative video, since recursive synthesis either compounds artifacts or does not—and the paper reports the latter.

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 SPEED, a pixel-space diffusion framework for video frame interpolation that performs one-step inference from Gaussian noise. The method combines a progressive multi-stage transformer with patch sizes 64/32/16, a Noise-Update-Only (NUO) attention mechanism that restricts attention queries to the noisy target while using global keys/values from all three frames, and a Drift-aware Timestep Sampling (DTS) curriculum plus direct x0-prediction with LPIPS/style losses to enable one-step generation. Experiments on DAVIS, SNU-FILM, and XTest4K report lower LPIPS/FloLPIPS than latent-diffusion and flow-based baselines, with substantially lower latency and memory usage.

Significance. If the reported results hold, SPEED is a valuable contribution: it demonstrates that pixel-space single-step diffusion is practical for VFI and avoids VAE reconstruction loss. The architecture is well motivated, and the ablations are internally consistent: Tables 5 and 6 clearly show that the x0 objective and DTS are both important, and the NUO attention gives a global receptive field at lower cost than full attention. The speed/memory gains are substantial and relevant for deployment. However, the central one-step claim is not fully established because the forward noising process is never defined and the DTS curriculum's treatment of the t=1 boundary is under-specified. The paper also does not provide code, error bars, or a clear separation between model selection and final test evaluation, which tempers confidence in the exact headline numbers.

major comments (3)
  1. [§3.3, Eq. (8)–(10), Alg. 1] DTS never samples exactly t'=1 during training. Eq. (8) defines s(p)=1/(1-p) on p∈[0,1), so for any finite training progress p the warped timestep t'=s(p)t/(1+(s(p)-1)t) is strictly less than 1 for every t<1; the only preimage of t'=1 is the measure-zero event t=1. Algorithm 1 nevertheless evaluates f_θ(ε,I0,I1,t=1). The assertion that the curriculum 'perfectly transitions' to instantaneous one-step projection therefore relies on an unstated smoothness/extrapolation assumption across the boundary. Table 6 does not isolate this concern: both the DTS and no-DTS models are evaluated at t=1, and the no-DTS model is strongly t-sensitive. Please report the realized p schedule, the final p, the maximum t' actually sampled in the last training epochs, and a sensitivity study around t=1 (e.g., t=0.9, 0.99, 0.999, 1.0). If the boundary extrapolation fails, the headline LPIPS gains could shrink or
  2. [§3.3, Eq. (11), Alg. 1] The forward noising process is never defined. The training loss in Eq. (11) uses a noisy intermediate frame I_t, but the paper does not state how I_t is generated from the clean target x0 and ε, nor which noise schedule (e.g., α_t) is used. The statement that t=1 corresponds to pure Gaussian noise is an assumption, not a consequence of any stated schedule. Without this definition, the training objective and the one-step inference protocol are not reproducible, and the claim that the model is trained to project from pure noise cannot be verified. Please specify the forward process explicitly, e.g., I_t = sqrt(α_t)x0 + sqrt(1−α_t)ε, and give the exact α_t schedule.
  3. [§4.4, Tables 3 and 5; §5.1] The model configuration and loss weights appear to be selected on the same benchmarks used for the headline comparisons. The 2-6-4 architecture in Table 3 is chosen using SNU-FILM (Extreme), and the loss weights w1=1.0, wl=1.0, ws=20.0 in §5.1 are set empirically; the one-step objectives in Table 5 and DTS analysis in Table 6 are evaluated on DAVIS-256. These same datasets appear in the main comparison tables (Tables 1–2). This creates a risk of overfitting to the test split, and no error bars or multiple-seed experiments are reported, making it unclear whether the improvements are statistically significant. Please use a held-out validation split for configuration selection, or at minimum state the selection protocol, and provide variance estimates for the main results.
minor comments (4)
  1. [§3.1 / §3.3] The notation is inconsistent: Eq. (1) uses I_t for the noised intermediate frame, while Algorithm 1 passes ε directly to f_θ. Please clarify that at t=1 the noised input is I_t = ε under the (to-be-specified) forward process.
  2. [§4.4 vs §5.1] Section 4.4 says all ablations are trained for 200k steps, while §5.1 says the model is optimized over 100 epochs. Please reconcile these numbers or clarify the relationship between steps and epochs.
  3. [Table 3] The row order for multi-stage configurations is not intuitive: '2-4-6' appears after '3-6-3' but is discussed as the fine-stage-heavy variant. Reordering the rows (e.g., by increasing depth of the p=16 stage) would make the table easier to read.
  4. [Reproducibility] The project page is given, but no code or model weights are provided. Given the paper's strong efficiency claims, releasing code would substantially improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: all key claims rest on empirical evaluations against external baselines.

full rationale

SPEED's derivation chain is self-contained in the sense that its central claims do not reduce to their inputs by construction. The one-step inference protocol (Algorithm 1, t=1) is not trained exactly at t=1; the DTS curriculum samples t'<1, so the claim depends on an extrapolation that is empirically validated in Table 6 rather than being a tautology. The performance gains (LPIPS reductions, speedups) are measured against external baselines, including EDEN, which is co-authored by some of the present authors but is a fixed existing method, not an input to SPEED's computations. The training objective (Eq. 11) uses standard pixel, LPIPS, and style losses; the LPIPS loss does not predetermine the reported LPIPS values on unseen test sets. Architecture choices and loss weights are selected through ablations and empirical tuning, not by requiring the final reported metrics to be matched. No equation in the paper is shown to be equivalent to another by construction, and no fitted parameter is disguised as a prediction. The only notable self-reference is the comparison and extension to EDEN, but that is not load-bearing for the validity of SPEED's own design. Therefore, no circularity is found.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim is empirical; it rests on a handful of hand-set hyperparameters and domain assumptions about pixel-space training and DTS transfer. No new physical or latent entities are introduced; NUO attention and DTS are mechanisms, not entities.

free parameters (2)
  • Loss weights w1, wl, ws = w1=1.0, wl=1.0, ws=20.0
    Empirically set in Section 5.1; they control the trade-off between reconstruction, perceptual, and style losses, and therefore affect the reported quality numbers.
  • Stage block configuration (2-6-4) = 2 blocks p=64, 6 blocks p=32, 4 blocks p=16
    Selected from Table 3 ablation as the best trade-off on SNU-FILM Extreme; the central efficiency and quality numbers depend on this configuration.
axioms (4)
  • domain assumption Pixel-space diffusion transformers can model high-resolution video frames without a VAE bottleneck and without prohibitive training instability.
    Used to justify operating entirely in pixel space (Section 2.2); supported by cited PixelDiT/DiP/DeCo, but still a domain assumption about scalability to 4K VFI.
  • ad hoc to paper A timestep-sampling curriculum that concentrates on t->1 during later training lets the network learn an instantaneous projection from pure noise at t=1, even though t=1 itself is never sampled exactly.
    Introduced in Section 3.3, Eq. (8)-(10); validated only empirically on DAVIS-256 (Table 6); no theoretical guarantee.
  • domain assumption LPIPS and FloLPIPS are sufficient metrics to rank interpolation quality, and PSNR/SSIM are not needed in the main comparison.
    Metric protocol in Section 4.1; PSNR/SSIM appear only in the DAVIS-256 ablation (Table 6).
  • domain assumption Training on LAVIB with random crops up to 1296px transfers to 2160x4096 resolution at inference.
    Training details in Section 5.1 and evaluation on XTest4K in Table 2; no analysis of the resolution gap.

pith-pipeline@v1.3.0-alltime-deepseek · 16003 in / 14611 out tokens · 165785 ms · 2026-08-01T22:50:58.261787+00:00 · methodology

0 comments
read the original abstract

Despite the success of diffusion models in Video Frame Interpolation (VFI), existing methods still suffer from two critical limitations. First, latent diffusion inevitably loses fine-grained details when reconstructing images from latent representations back to the pixel space. Second, multi-step sampling incurs prohibitive memory consumption and inference latency. To address these issues, we propose SPEED, a one-step pixel diffusion framework for high-quality VFI. Specifically, SPEED employs a progressive multi-stage architecture with dynamic patch scaling to effectively learn multi-scale motion, structural, and appearance representations. Furthermore, we propose a novel Noise-Update-Only Attention mechanism to prevent semantic degradation of the clean condition frames while reducing the computational overhead by nearly 50%. Besides, we introduce a Drift-aware Timestep Sampling strategy coupled with a tailored training objective to directly predict images in the pixel space, enabling one-step inference without compromising the quality of the generated frames. Extensive experiments show that SPEED achieves state-of-the-art performance. On SNU-FILM, SPEED reduces LPIPS by 8.8% while delivering 63.3% faster inference and 10.6% lower memory usage. On challenging 4K benchmarks, it further surpasses prior methods by up to 51.5% in LPIPS.

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

60 extracted references · 13 linked inside Pith

  1. [1]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXivpreprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Learningtosynthesizemotionblur

    TimBrooksandJonathanTBarron. Learningtosynthesizemotionblur. In ProceedingsoftheIEEE/CVFConference on ComputerVisionand PatternRecognition, pages 6840–6848, 2019

    2019

  3. [3]

    Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The TwelfthInternational Conferenceon Learning Representations, 2024

    2024

  4. [4]

    Videoinr: Learning video implicit neural representation for continuous space-time super-resolution

    Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vidit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. In Proceedings ofthe IEEE/CVF ConferenceonComputerVisionand PatternRecognition, pages 2047–2057, 2022

    2047

  5. [5]

    Dip: Taming diffusion models in pixel space.arXiv preprintarXiv:2511.18822, 2025

    Zhennan Chen, Junwei Zhu, Xu Chen, Jiangning Zhang, Xiaobin Hu, Hanzhen Zhao, Chengjie Wang, Jian Yang, and Ying Tai. Dip: Taming diffusion models in pixel space.arXiv preprintarXiv:2511.18822, 2025

    arXiv 2025

  6. [6]

    Channel attention is all you need for video frame interpolation

    Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Kyoung Mu Lee. Channel attention is all you need for video frame interpolation. InProceedings of the AAAI Conferenceon Artificial Intelligence, volume 34, pages 10663–10671, 2020

    2020

  7. [7]

    Flolpips: A bespoke video quality metric for frame interpolation

    Duolikun Danier, Fan Zhang, and David Bull. Flolpips: A bespoke video quality metric for frame interpolation. In 2022 PictureCodingSymposium(PCS), pages 283–287. IEEE, 2022

    2022

  8. [8]

    St-mfnet: Aspatio-temporalmulti-flownetworkforframeinterpolation

    DuolikunDanier,FanZhang,andDavidBull. St-mfnet: Aspatio-temporalmulti-flownetworkforframeinterpolation. In Proceedings ofthe IEEE/CVF Conferenceon ComputerVisionand PatternRecognition, pages 3521–3531, 2022

    2022

  9. [9]

    Ldmvfi: Video frame interpolation with latent diffusion models

    Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings ofthe AAAI Conferenceon Artificial Intelligence, volume 38, pages 1472–1480, 2024

    2024

  10. [10]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-firstInternational Conferenceon MachineLearning, 2024

    2024

  11. [11]

    Deepstereo: Learning to predict new views from the world’s imagery

    John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world’s imagery. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5515–5524, 2016

    2016

  12. [12]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. InFirstconferenceon languagemodeling, 2024

    2024

  13. [13]

    Denoisingdiffusionprobabilisticmodels

    JonathanHo,AjayJain,andPieterAbbeel. Denoisingdiffusionprobabilisticmodels. Advancesinneuralinformation processing systems, 33:6840–6851, 2020

    2020

  14. [14]

    Real-timeintermediateflowestimation for video frame interpolation

    ZheweiHuang,TianyuanZhang,WenHeng,BoxinShi,andShuchangZhou. Real-timeintermediateflowestimation for video frame interpolation. InEuropean ConferenceonComputerVision, pages 624–642, 2022

    2022

  15. [15]

    Scale-adaptivefeatureaggregation for efficient space-time video super-resolution

    ZheweiHuang,AilinHuang,XiaotaoHu,ChenHu,JunXu,andShuchangZhou. Scale-adaptivefeatureaggregation for efficient space-time video super-resolution. InWinterConferenceon ApplicationsofComputerVision, 2024

    2024

  16. [16]

    High-resolution frame interpolation with patch-based cascaded diffusion

    JunhwaHur,CharlesHerrmann,SaurabhSaxena,JanneKontkanen,Wei-ShengLai,YichangShih,MichaelRubinstein, David J Fleet, and Deqing Sun. High-resolution frame interpolation with patch-based cascaded diffusion. In Proceedings ofthe AAAI Conferenceon Artificial Intelligence, volume 39, pages 3868–3876, 2025

    2025

  17. [17]

    Video interpolation with diffusion models

    Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. InProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition,pages7341–7351, 2024

    2024

  18. [18]

    Lc-mamba: Local and continuous mamba with shifted windows for frame interpolation

    Min Wu Jeong and Chae Eun Rhee. Lc-mamba: Local and continuous mamba with shifted windows for frame interpolation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17671–17681, 2025. 13

    2025

  19. [19]

    Neighbor correspondence matching for flow-based video frame synthesis

    Zhaoyang Jia, Yan Lu, and Houqiang Li. Neighbor correspondence matching for flow-based video frame synthesis. In Proceedings ofthe 30th ACMInternational Conferenceon Multimedia, pages 5389–5397, 2022

    2022

  20. [20]

    Super slomo: High quality estimation of multiple intermediate frames for video interpolation

    Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. InProceedings of the IEEE conferenceon computervision and pattern recognition, pages 9000–9008, 2018

    2018

  21. [21]

    Auto-encoding variational bayes.arXiv preprintarXiv:1312.6114, 2013

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprintarXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  22. [22]

    Ifrnet: Intermediate feature refine network for efficient frame interpolation

    Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. InProceedings of the IEEE/CVF Conferenceon ComputerVisionand PatternRecognition, pages 1969–1978, 2022

    1969

  23. [23]

    Back to basics: Let denoising generative models denoise

    Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise. arXiv preprint arXiv:2511.13720, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Amt: All-pairs multi-field transforms for efficient frame interpolation

    Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. InProceedingsofthe IEEE/CVFConferenceonComputer VisionandPatternRecognition, pages 9801–9810, 2023

    2023

  25. [25]

    Sparseglobalmatchingforvideoframeinterpolationwith large motion

    ChunxuLiu,GuozhenZhang,RuiZhao,andLiminWang. Sparseglobalmatchingforvideoframeinterpolationwith large motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19125–19134, 2024

    2024

  26. [26]

    Sora: A review on background, technology, limitations, and opportunities of large vision models

    Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprintarXiv:2402.17177, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Video frame interpolation with transformer

    Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings ofthe IEEE/CVF ConferenceonComputerVisionand PatternRecognition, pages 3532–3542, 2022

    2022

  28. [28]

    Tlb-vfi: Temporal-awarelatentbrownianbridgediffusionforvideoframeinterpolation

    ZonglinLyuandChenChen. Tlb-vfi: Temporal-awarelatentbrownianbridgediffusionforvideoframeinterpolation. In Proceedings ofthe IEEE/CVF International ConferenceonComputerVision, pages 16260–16269, 2025

    2025

  29. [29]

    Frame interpolation with consecutive brownian bridge diffusion

    Zonglin Lyu, Ming Li, Jianbo Jiao, and Chen Chen. Frame interpolation with consecutive brownian bridge diffusion. In Proceedings ofthe 32nd ACMInternational Conferenceon Multimedia, pages 3449–3458, 2024

    2024

  30. [30]

    Latte: Latent diffusion transformer for video generation.arXivpreprintarXiv:2401.03048, 2024

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.arXivpreprintarXiv:2401.03048, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Deco: Frequency-decoupled pixel diffusion for end-to-end image generation.arXivpreprintarXiv:2511.19365, 2025

    Zehong Ma, Longhui Wei, Shuai Wang, Shiliang Zhang, and Qi Tian. Deco: Frequency-decoupled pixel diffusion for end-to-end image generation.arXivpreprintarXiv:2511.19365, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Softmax splatting for video frame interpolation

    Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. InProceedings of the IEEE/CVF Conferenceon ComputerVisionand PatternRecognition, pages 5437–5446, 2020

    2020

  33. [33]

    Building the gist of a scene: The role of global image features in recognition

    Aude Oliva and Antonio Torralba. Building the gist of a scene: The role of global image features in recognition. Progress inbrainresearch, 155:23–36, 2006

    2006

  34. [34]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conferenceon ComputerVision, pages 4195–4205, 2023

    2023

  35. [35]

    The 2017 davis challenge on video object segmentation.arXivpreprintarXiv:1704.00675, 2017

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation.arXivpreprintarXiv:1704.00675, 2017

    Pith/arXiv arXiv 2017

  36. [36]

    Film: Frame interpolation for large motion

    Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. InEuropeanConferenceon ComputerVision, pages 250–266, 2022

    2022

  37. [37]

    High-resolution image synthesiswithlatentdiffusionmodels

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesiswithlatentdiffusionmodels. In ProceedingsoftheIEEE/CVFconferenceoncomputervisionandpattern recognition, pages 10684–10695, 2022

    2022

  38. [38]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European ConferenceonComputerVision, pages 87–103, 2024. 14

    2024

  39. [39]

    Bim-vfi: Bidirectional motion field-guided frame interpolation for video with non-uniform motions

    Wonyong Seo, Jihyong Oh, and Munchurl Kim. Bim-vfi: Bidirectional motion field-guided frame interpolation for video with non-uniform motions. InProceedings of the ComputerVisionand PatternRecognitionConference, pages 7244–7253, 2025

    2025

  40. [40]

    Xvfi: extreme video frame interpolation

    Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. InProceedings of the IEEE/CVF international conferenceon computervision, pages 14489–14498, 2021

    2021

  41. [41]

    Lavib: A large-scale video interpolation benchmark

    Alexandros Stergiou. Lavib: A large-scale video interpolation benchmark. Advances in Neural Information ProcessingSystems, 37:29091–29105, December 2024

    2024

  42. [42]

    Biological motion perception.Thesenses: Acomprehensivereference, 2:231–238, 2008

    Nikolaus F Troje and A Basbaum. Biological motion perception.Thesenses: Acomprehensivereference, 2:231–238, 2008

    2008

  43. [43]

    Motionfollower: Editingvideomotionvialightweightscore-guideddiffusion

    Shuyuan Tu, Qi Dai, Zihao Zhang, Sicheng Xie, Zhi-Qi Cheng, Chong Luo, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motionfollower: Editingvideomotionvialightweightscore-guideddiffusion. arXivpreprintarXiv:2405.20325, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Functional neuroanatomy of biological motion perception in humans.Proceedings of the NationalAcademyof Sciences, 98(20):11656–11661, 2001

    Lucia M Vaina, Jeffrey Solomon, Sanjida Chowdhury, Pawan Sinha, and John W Belliveau. Functional neuroanatomy of biological motion perception in humans.Proceedings of the NationalAcademyof Sciences, 98(20):11656–11661, 2001

    2001

  45. [45]

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

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

    Pith/arXiv arXiv 2025

  46. [46]

    Pixnerd: Pixel neural field diffusion.arXiv preprint arXiv:2507.23268, 2025

    Shuai Wang, Ziteng Gao, Chenhui Zhu, Weilin Huang, and Limin Wang. Pixnerd: Pixel neural field diffusion.arXiv preprint arXiv:2507.23268, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Video compression through image interpolation

    Chao-Yuan Wu, Nayan Singhal, and Philipp Krahenbuhl. Video compression through image interpolation. In Proceedings ofthe European conferenceoncomputervision (ECCV), pages 416–431, 2018

    2018

  48. [48]

    Qwen-image technical report.arXiv preprintarXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprintarXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Improved distribution matching distillation for fast image synthesis.Advancesin neuralinformation processing systems, 37:47455–47487, 2024

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis.Advancesin neuralinformation processing systems, 37:47455–47487, 2024

    2024

  50. [50]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. InProceedings of the IEEE/CVF conference on computervision andpattern recognition, pages 6613–6623, 2024

    2024

  51. [51]

    Pixeldit: Pixel diffusion transformers for image generation.arXivpreprint arXiv:2511.20645, 2025

    Yongsheng Yu, Wei Xiong, Weili Nie, Yichen Sheng, Shiqiu Liu, and Jiebo Luo. Pixeldit: Pixel diffusion transformers for image generation.arXivpreprint arXiv:2511.20645, 2025

    Pith/arXiv arXiv 2025

  52. [52]

    Extracting motion and appearance via inter-frame attention for efficient video frame interpolation

    Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. InProceedings ofthe IEEE/CVF Conferenceon ComputerVisionand PatternRecognition, pages 5682–5692, 2023

    2023

  53. [53]

    Vfimamba: Video frame interpolationwithstatespacemodels

    Guozhen Zhang, Chunxu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Vfimamba: Video frame interpolationwithstatespacemodels. AdvancesinNeuralInformationProcessingSystems,37:107225–107248,2024

    2024

  54. [54]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  55. [55]

    Eden: Enhanced diffusion for high-quality large-motion video frame interpolation

    Zihao Zhang, Haoran Chen, Haoyu Zhao, Guansong Lu, Yanwei Fu, Hang Xu, and Zuxuan Wu. Eden: Enhanced diffusion for high-quality large-motion video frame interpolation. InProceedings of the Computer Vision and PatternRecognitionConference, pages 2105–2115, 2025

    2025

  56. [56]

    Cameranoise: Enabling faithful camera control in video diffusion through geometry-flow-guided noise warping

    HaoyuZhao,JiaxiGu,HaoranChen,QingpingZheng,YeyingJin,HongyiYang,JunqiCheng,YuangZhang,Zenghui Lu, Huan Yu, et al. Cameranoise: Enabling faithful camera control in video diffusion through geometry-flow-guided noise warping. InForty-thirdInternational Conferenceon MachineLearning. 15

  57. [57]

    Magdiff: Multi-alignment diffusion for high-fidelity video generation and editing

    Haoyu Zhao, Tianyi Lu, Jiaxi Gu, Xing Zhang, Qingping Zheng, Zuxuan Wu, Hang Xu, and Yu-Gang Jiang. Magdiff: Multi-alignment diffusion for high-fidelity video generation and editing. InEuropean Conference on Computer Vision, pages 205–221. Springer, 2025

    2025

  58. [58]

    Dynamictrl: Rethinking the basic structure and the role of text for high-quality human image animation.arXiv preprint arXiv:2503.21246, 2025

    Haoyu Zhao, Zhongang Qi, Cong Wang, Qingping Zheng, Guansong Lu, Fei Chen, Hang Xu, and Zuxuan Wu. Dynamictrl: Rethinking the basic structure and the role of text for high-quality human image animation.arXiv preprint arXiv:2503.21246, 2025

    Pith/arXiv arXiv 2025

  59. [59]

    Lstd: Long short-term temporal diffusion for video generation.IEEETransactionson Multimedia, 2026

    Haoyu Zhao, Jiaxi Gu, Shicong Wang, Tianyi Lu, Xing Zhang, Zuxuan Wu, Hang Xu, and Yu-Gang Jiang. Lstd: Long short-term temporal diffusion for video generation.IEEETransactionson Multimedia, 2026

    2026

  60. [60]

    Blur interpolation transformer for real- worldmotionfromblur

    Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real- worldmotionfromblur. In ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition, pages 5713–5723, 2023. 16 5.1 Implementation Details SPEED employs a base hidden state dimension ofd= 768. To facilitate our macroscopic-to-microsco...

    2023