Pith. sign in

REVIEW 5 major objections 7 minor 60 references

The paper proposes AnchorSync, a two-stage diffusion framework that edits sparse anchor frames jointly and then interpolates the rest with optical-flow and edge guidance, to keep long videos globally consistent across thousands of frames.

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 ↗

By jointly editing sparse anchor frames and interpolating with flow and edge guidance, AnchorSync produces temporally consistent edits on videos longer than previous diffusion methods could handle.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid engineering paper with a plausible two-stage pipeline, but the headline long-term metrics are computed on the method's own anchor grid, so the temporal-consistency claim is weaker than the tables suggest. the 5 major comments →

arxiv 2508.14609 v1 pith:FHJWZSEF submitted 2025-08-20 cs.CV

AnchorSync: Global Consistency Optimization for Long Video Editing

Zichi Liu , Yinggui Wang , Tao Wei , Chao Ma This is my paper
classification cs.CV
keywords long video editingdiffusion modelstemporal consistencyanchor framesbidirectional attentionmultimodal guidanceoptical flowvideo interpolation
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 reading

This paper tackles long-video editing, where current diffusion methods drift or flicker past a few hundred frames. It claims that decoupling the job into sparse anchor-frame editing plus guided interpolation makes minute-long edits stable. A trainable bidirectional-attention block lets each anchor frame influence its neighbors while Plug-and-Play inversion preserves structure. Interpolation then uses a video diffusion model steered by Canny edges and optical flow so the in-between frames follow the original motion. If right, it offers a practical path to editing videos thousands of frames long with a single prompt.

Core claim

AnchorSync's central claim is that long-range video editing can be made globally consistent by editing a sparse set of anchor frames jointly, not one by one, and by interpolating the rest with a video diffusion model conditioned on the original video's edges and optical flow. The anchor stage uses pairwise diffusion with a bidirectional attention layer, so each anchor attends to its neighbors, and progressive fusion across overlapping pairs propagates coherence through the whole sequence. The interpolation stage runs Stable Video Diffusion both forward and backward and blends the two trajectories, while a multimodal ControlNet injects Canny edges and optical flow. The paper reports that this

What carries the argument

The load-bearing pieces are: (1) Bidirectional Attention, a parallel attention branch that lets one anchor frame's denoising stream attend to another's keys and values; (2) progressive pairwise diffusion, which jointly denoises adjacent anchor pairs and averages shared latents across pairs; (3) Plug-and-Play feature injection from an inversion stream to preserve structure; and (4) a multimodal ControlNet on Stable Video Diffusion, fusing Canny edges (static structure) and optical flow (motion) to interpolate frames between anchors.

Load-bearing premise

The load-bearing premise is that Canny edges and optical flow extracted from the original, unedited video remain valid guidance for the edited frames; if the edit changes shape or geometry, the interpolation is guided by stale structure and motion.

What would settle it

Take a video where the edit changes an object's category or shape (e.g., a camel into a horse) and measure frame-to-frame warp error or CLIP similarity between edited frames. The paper predicts a drop in temporal consistency; if the output stays perfectly stable, the claimed reliance on original edges and flow is not actually binding. Alternatively, run the interpolation stage on a fast-moving scene and compare the multimodal version against variants using only edges or only flow; the paper predicts the multimodal version is clearly better.

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

If this is right

  • Long videos of several minutes can be edited in one pass with stable identity and motion because edits are computed jointly on sparse anchors rather than frame-by-frame.
  • The two-stage decomposition reduces memory pressure: only anchor frames go through joint diffusion, while interpolation uses a video diffusion model with light conditioning.
  • Bidirectional attention plus progressive pairwise fusion propagates consistency across all anchors, avoiding the cumulative error of sequential segment editing.
  • Multimodal control with edges and flow preserves both structure and motion during interpolation, avoiding artifacts that appear when using flow-only warping like Ebsynth.
  • The reported numbers suggest the method sets a new baseline for long-term CLIP similarity, frame continuity, warp error, and user preference in long-video editing.

Where Pith is reading between the lines

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

  • The same anchor-plus-interpolation decomposition could be applied to video inpainting, style transfer, or multi-shot narrative generation, where long-range coherence is the bottleneck.
  • Because the flow is taken from the original video, edits that change motion direction or articulation (e.g., 'make the car go backward') would be actively resisted; an extension would predict a warped flow field for the edited content.
  • The fixed 24-frame anchor spacing is a practical choice; adaptive anchor selection based on scene cuts or motion magnitude could reduce interpolation load and improve robustness.
  • Progressive pairwise fusion can be viewed as an implicit pairwise consistency prior; one could test whether it is equivalent to minimizing a simple energy over anchor latents, which would give a cleaner explanation for why it works.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 7 minor

Summary. AnchorSync proposes a two-stage diffusion-based framework for long video editing. In Stage 1, a sparse set of anchor frames (every K=24 frames) is jointly edited via pairwise diffusion with a bidirectional attention module, PnP feature injection, and multi-conditional guidance. In Stage 2, intermediate frames are synthesized by a Stable Video Diffusion model augmented with a multimodal ControlNet (Canny edges + optical flow) and bidirectional temporal frame fusion. Experiments on short/medium/long videos compare against Rerender, Gen-L-Video, Anyv2v*, and StreamV2V, reporting quantitative metrics, user studies, and ablations. The paper claims superior long-term consistency, temporal stability, and visual quality, with source code promised.

Significance. If substantiated, the two-stage decoupling of anchor-frame editing and interpolation is a practical and scalable approach to minute-long video editing, combining global consistency with local smoothness. The paper ships several useful components: a trainable bidirectional attention for cross-frame consistency, a PnP-based anchor editing scheme, and a multimodal control strategy for interpolation using both flow and edges. The evaluation is extensive in terms of dataset length coverage and ablation scope. However, the central claim of 'global consistency' rests heavily on custom metrics computed on the anchor grid and on a user study without significance testing, so the current evidence is not yet convincing.

major comments (5)
  1. [Section 4, 'Evaluation Metrics'; Table 2] The long-term metrics I-I CLIP Sim★ and I-I CLIP Sim† are defined on frames that are exactly 24 frames apart or on every 24th frame versus the first. Since Stage 1 edits anchor frames every 24 frames and explicitly optimizes their consistency, these metrics measure the method's own anchor grid rather than the 23 interpolated frames between anchors, where temporal drift would actually appear. This makes the headline margins (e.g., 97.84 vs. 96.74) potentially self-fulfilling. Please report the same metrics on all frames or on uniformly sampled non-anchor frames, and validate the custom metrics against human judgments or standard full-video temporal-consistency measures (e.g., LPIPS-based warp error computed on the edited video).
  2. [Tables 2, 3, 4; Section 4] No error bars, confidence intervals, or significance tests are reported. Many quantitative differences are small (e.g., I-I CLIP Sim★ 97.84 vs. 96.74; Canny Error 10.36 vs. 10.81; I-I CLIP Sim. 99.64 vs. 99.36), and the user-study scores are given as single numbers without measures of variance or inter-rater agreement. Per-video confidence intervals and paired significance tests are needed to support the claimed superiority. Additionally, hyperparameters (anchor interval K, injection ratios, guidance scales) appear to have been tuned on the same dataset used for final evaluation; state whether the reported numbers are on a held-out split or the tuning set.
  3. [Table 4] Table 4 is titled 'Comparison of anchor frame selection strategies,' but the row 'w/o Interpolation' is not an anchor-selection strategy; it appears to evaluate the interpolation module. This conflates two separate ablation axes. If 'w/o Interpolation' means evaluating only the anchor frames, then the long-term metric on anchors is expected to be favorable by construction, and the row does not disentangle the contributions of anchor spacing vs. interpolation. Clarify the protocol for this row and consider separating the anchor-selection ablation from the interpolation ablation.
  4. [Section 3.3 and Section 5] The interpolation stage conditions on Canny edge maps and optical flow extracted from the original, unedited video. As Section 5 concedes, when the edit changes object shape or category (e.g., camel to horse), these cues no longer match the edited content, and temporal consistency breaks. This means the paper's central claim of 'high-quality long video editing' is effectively scoped to structure-preserving edits, but neither the abstract nor the introduction states this scope. Either qualify the claims explicitly, or provide quantitative evidence on edits with moderate geometric changes to show the method degrades gracefully.
  5. [Section 3.2, Eq. (5)] The progressive pairwise fusion procedure is underspecified. For a shared anchor frame f1 in pairs (f0,f1) and (f1,f2), how are the two pairwise denoising trajectories initialized and synchronized at each timestep? After averaging the latent for f1 via Eq. (5), how is the averaged latent fed back into both pair processes for the next denoising step? Provide pseudocode or a precise algorithmic description, including how the noise latents for overlapping pairs are aligned, so that the implementation is reproducible and the cross-pair fusion component can be properly evaluated.
minor comments (7)
  1. [Section 1 and Related Work] Reference [16] (Prompt-to-Prompt) is cited in the introduction and related work as a video editing method, but it is an image editing method. Please correct the citation or replace it with the intended video-editing reference.
  2. [Section 4.1] The sentence 'Anyv2v* underperforms on all metrics' is contradicted by Table 2, where Anyv2v* achieves the second-best Warp Error (5.19) and better than several baselines on that metric. Rephrase to 'underperforms on most metrics.'
  3. [Table 1 and Section 4, User Study] Table 1 lists 20+20+20=60 videos, but the user study section mentions 'a pool of 284' edited videos. Clarify the relationship between these counts; presumably multiple prompts/edits per video were used, but this is not stated.
  4. [Section 3.2, Eq. (5)] The notation e_t(f0,f1)[1] and e_t(f1,f2)[0] is ambiguous. Define the bracket indexing explicitly (e.g., output for the second frame of the pair).
  5. [Section 4, Implementation Details] The injection ratios are described as '0.44 and 0.65, respectively' but Section 4.2 and Fig. 8 refer to 'conv = 0.65, attn = 0.44'. Make the order consistent throughout the paper.
  6. [Figure 6] Caption typo: 'Analysing' should be 'Analysis'.
  7. [Section 4, Evaluation Metrics] The symbols ★ and † in I-I CLIP Sim★ and I-I CLIP Sim† are not defined in the text; they only appear in the table header. Explain the notation at first use.

Circularity Check

2 steps flagged

Long-term consistency metric is defined on the method's own 24-frame anchor grid, and key hyperparameters are tuned on the same evaluation videos, making part of the claimed superiority an evaluation artifact.

specific steps
  1. self definitional [Section 4 (Evaluation Metrics) and Section 3.2 (Anchor Editing via Pairwise Diffusion)]
    "(i) I-I CLIP Sim★ measures the average CLIP similarity between frames that are 24 frames apart, capturing stability over time. ... Anchor frames are selected every 24 frames to balance editing efficiency and interpolation quality, following the recommended temporal spacing in SVD."

    Anchors are sampled every K=24 frames (f_i = v_{i·K}), so 'frames that are 24 frames apart' are exactly adjacent anchor pairs. Stage 1 is engineered to make adjacent anchor pairs consistent: Bidirectional Attention (Eq. 1-2) and cross-pair fusion (Eq. 5) explicitly couple each anchor with its neighbor. The headline long-term metric therefore measures the method's own optimization target, not the 23 interpolated frames in each segment where temporal drift would appear. Table 4 confirms this: the 'w/o Interpolation' row still scores 96.81 on I-I CLIP Sim★, showing the metric is nearly insensitive to interpolation quality. The claimed superiority in long-term/global consistency is thus partly self-fulfilling.

  2. fitted input called prediction [Section 4 (Ablation Studies: Impact of Multi-conditional Guidance Scale; Impact of Injection Ratios; Anchor Frame Selection) and Table 2]
    "In practice, we find that setting s_T = 6.0 and s_J = 0.8 strikes a good balance between edit fidelity and structural consistency. ... Our default setting (conv = 0.65, attn = 0.44) provides a good trade-off ... We conduct an ablation study on anchor frame selection strategies, as shown in Tab. 4."

    The default hyperparameters (s_T=6.0, s_J=0.8; conv=0.65, attn=0.44; anchor interval 24) are selected via ablations on the evaluation videos (Fig. 7, Fig. 8, Tab. 4). The same videos and the same metrics are then reported in Table 2 as evidence that AnchorSync surpasses baselines. Thus the quantitative comparison is in-sample: the configuration was chosen to maximize the reported metrics on the exact data used for the final comparison. The 'prediction' of superiority is statistically forced by the selection process, not an out-of-sample result.

full rationale

The core algorithm is not derived from its outputs; the two-stage pipeline is a straightforward engineering contribution and the method is not built on a self-citation chain or uniqueness theorem. However, the quantitative evidence for the central claim of global consistency is partially circular. First, the custom long-term metrics I-I CLIP Sim★/† operate on frames spaced exactly at the anchor interval K=24, and Stage 1's bidirectional attention and cross-pair fusion explicitly optimize consistency between those frames; thus the metric measures the method's own optimization target rather than the interpolated frames where drift would appear. Table 4's 'w/o Interpolation' row still scores 96.81 on the long-term metric, confirming the metric is insensitive to the interpolation stage. Second, the default hyperparameters (guidance scales, injection ratios, anchor interval) are selected via ablations on the same evaluation dataset and then reported in the final comparison table, so the reported margins are in-sample rather than out-of-sample predictions. The paper does include a user study and additional frame-level metrics (Frame Continuity, Warp Error, Canny Error) that provide some independent support, so the circularity is partial, not total. The self-citation to VidToMe [27] for warp/canny protocols is not load-bearing and does not raise the score further.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The results are empirical; the central claim depends on a set of hand-tuned hyperparameters and borrowed components (PnP, SVD, ControlNet) whose assumptions are not re-derived. No new physical or conceptual entities are introduced.

free parameters (3)
  • Anchor frame interval K = 24 frames
    Chosen as best in ablation (Tab. 4); balances interpolation quality and number of anchors.
  • PnP injection ratios (conv, attn) = conv 0.65, attn 0.44
    Tuned via ablation (Fig. 8) for best edit fidelity vs structure retention.
  • Guidance scales (s_T, s_J) = 6.0, 0.8
    Selected in ablation (Fig. 7); higher s_T increases edit strength, higher s_J improves coherence but can hurt fidelity.
axioms (4)
  • domain assumption Plug-and-Play (PnP) feature injection preserves source spatial structure while enabling semantic edits.
    Stage 1 relies on PnP injection (Tumanyan et al. [47]) to keep edited anchors aligned with the source layout; this is an empirical property, not derived.
  • domain assumption Stable Video Diffusion can generate temporally coherent intermediate frames given a start frame and control signals.
    Stage 2 uses SVD [3] as the backbone; the method inherits its capabilities and limits.
  • domain assumption Canny edge maps and optical flow of the original video are valid structural/motion guides for the edited content.
    The multimodal ControlNet conditions on these cues; the paper's own Limitations section (Section 5) shows this breaks for drastic structural changes.
  • domain assumption Uniform anchor sampling every 24 frames captures the semantic variation of the video without missing important changes.
    The framework assumes sparse anchors are representative; ablation in Tab. 4 tunes the interval but does not prove it generalizes.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AnchorSync: Global Consistency Optimization for Long Video Editing." pith.science (2026). https://pith.science/paper/FHJWZSEF

@misc{pith2026250814609,
  author       = {Pith},
  title        = {Pith review of: AnchorSync: Global Consistency Optimization for Long Video Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FHJWZSEF}},
  note         = {Machine review of arXiv:2508.14609}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Editing long videos remains a challenging task due to the need for maintaining both global consistency and temporal coherence across thousands of frames. Existing methods often suffer from structural drift or temporal artifacts, particularly in minute-long sequences. We introduce AnchorSync, a novel diffusion-based framework that enables high-quality, long-term video editing by decoupling the task into sparse anchor frame editing and smooth intermediate frame interpolation. Our approach enforces structural consistency through a progressive denoising process and preserves temporal dynamics via multimodal guidance. Extensive experiments show that AnchorSync produces coherent, high-fidelity edits, surpassing prior methods in visual quality and temporal stability.

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 · 42 canonical work pages

  1. [1]

    Omri Avrahami, Dani Lischinski, and Ohad Fried. 2022. Blended diffusion for text-driven editing of natural images. In IEEE Conference on Computer Vision and Pattern Recognition. 18208–18218

    work page 2022

  2. [2]

    Sahbi Bahroun, Rahma Abed, and Ezzeddine Zagrouba. 2021. KS-FQA: Keyframe selection based on face quality assessment for efficient face recognition in video. IET Image Processing 15, 1 (2021), 77–90

    work page 2021

  3. [3]

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

    Pith/arXiv arXiv 2023

  4. [4]

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. 2023. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conference on Computer Vision and Pattern Recognition . 18392–18402

    work page 2023

  5. [5]

    Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. 2023. Pix2video: Video editing using image diffusion. In IEEE Conference on Computer Vision and Pattern Recognition. 23206–23217

    work page 2023

  6. [6]

    Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming- Hsuan Yang, et al . 2024. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In IEEE Conference on Computer Vision and Pattern Recognition. 13320–13331

    work page 2024

  7. [7]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34 (2021), 8780–8794

    work page 2021

  8. [8]

    Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. 2024. Ccedit: Creative and controllable video editing via diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition. 6712–6722

    work page 2024

  9. [9]

    Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. 2022. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision . 89–106

    work page 2022

  10. [10]

    Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. 2023. Tokenflow: Consistent diffusion features for consistent video editing.arXiv:2307.10373 (2023)

    Pith/arXiv arXiv 2023

  11. [11]

    Shanyan Guan, Yanhao Ge, Ying Tai, Jian Yang, Wei Li, and Mingyu You. 2025. HybridBooth: Hybrid Prompt Inversion for Efficient Subject-Driven Generation. In European Conference on Computer Vision . 403–419

    work page 2025

  12. [12]

    Feihong He, Gang Li, Mengyuan Zhang, Leilei Yan, Lingyu Si, Fanzhang Li, and Li Shen. 2024. Freestyle: Free lunch for text-guided style transfer using diffusion models. arXiv:2401.15636 (2024)

    Pith/arXiv arXiv 2024

  13. [13]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  14. [14]

    Wenfeng He, Chulong Zhang, Jingjing Dai, Lin Liu, Tangsheng Wang, Xuan Liu, Yuming Jiang, Na Li, Jing Xiong, Lei Wang, et al. 2024. A statistical defor- mation model-based data augmentation method for volumetric medical image segmentation. Medical Image Analysis 91 (2024), 102984

    work page 2024

  15. [16]

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. 2022. Prompt-to-prompt image editing with cross attention control. arXiv:2208.01626 (2022)

    Pith/arXiv arXiv 2022

  16. [17]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. International Conference on Learning Representations 1, 2 (2022), 3

    work page 2022

  17. [18]

    Minghui Hu, Jianbin Zheng, Daqing Liu, Chuanxia Zheng, Chaoyue Wang, Dacheng Tao, and Tat-Jen Cham. 2023. Cocktail: Mixing multi-modality con- trol for text-conditional image generation. In Advances in Neural Information Processing Systems

    work page 2023

  18. [19]

    Saksham Jain, Gautam Seth, Arpit Paruthi, Umang Soni, and Girish Kumar. 2022. Synthetic data augmentation for surface defect detection and classification using deep learning. Journal of Intelligent Manufacturing (2022), 1–14

    work page 2022

  19. [20]

    Ondřej Jamriška, Šárka Sochorová, Ondřej Texler, Michal Lukáč, Jakub Fišer, Jingwan Lu, Eli Shechtman, and Daniel S`ykora. 2019. Stylizing video by example. ACM Transactions on Graphics (TOG) 38, 4 (2019), 1–11

    work page 2019

  20. [21]

    Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Hen- schel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. 2023. Text2video-zero: Text-to-image diffusion models are zero-shot video genera- tors. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 15954–15964

    2023

  21. [22]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  22. [23]

    Max Ku, Cong Wei, Weiming Ren, Huan Yang, and Wenhu Chen. 2024. Anyv2v: A plug-and-play framework for any video-to-video editing tasks. arXiv:2403.14468 (2024)

    Pith/arXiv arXiv 2024

  23. [24]

    Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas Guibas, and Gordon Wetzstein. 2024. Collaborative Video Diffusion: Consistent Multi-video Generation with Camera Control. Advances in Neural Information Processing Systems 37 (2024), 16240–16271

    work page 2024

  24. [25]

    Gihyun Kwon and Jong Chul Ye. 2022. Diffusion-based image translation using disentangled style and content representation. arXiv:2209.15264 (2022)

    Pith/arXiv arXiv 2022

  25. [26]

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. 2018. Learning blind video temporal consistency. In European Conference on Computer Vision . 170–185

    work page 2018

  26. [27]

    Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. 2024. VidToMe: Video Token Merging for Zero-Shot Video Editing. In IEEE Conference on Computer Vision and Pattern Recognition . 7486–7495

    work page 2024

  27. [28]

    Yin Li and Rajalakshmi Nandakumar. 2025. WixUp: A General Data Augmentation Framework for Wireless Perception in Tracking of Humans. In Proceedings of the 23rd ACM Conference on Embedded Networked Sensor Systems . 449–462

    work page 2025

  28. [29]

    Feng Liang, Akio Kodaira, Chenfeng Xu, Masayoshi Tomizuka, Kurt Keutzer, and Diana Marculescu. 2024. Looking Backward: Streaming Video-to-Video Translation with Feature Banks. arXiv:2405.15757 (2024)

    Pith/arXiv arXiv 2024

  29. [30]

    Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al . 2024. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In IEEE Conference on Computer Vision and Pattern Recognition . 8207–8216

    work page 2024

  30. [31]

    Hao Liang, Jiapeng Li, Tianyi Bai, Xijie Huang, Linzhuang Sun, Zhengren Wang, Conghui He, Bin Cui, Chong Chen, and Wentao Zhang. 2024. Keyvideollm: Towards large-scale video keyframe selection. arXiv:2407.03104 (2024)

    Pith/arXiv arXiv 2024

  31. [32]

    Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. 2024. To- wards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing. In IEEE Conference on Computer Vision and Pattern Recognition . 7817–7826

    work page 2024

  32. [33]

    Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. 2022. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision . 423–439

    work page 2022

  33. [34]

    Mingyang Ma, Shaohui Mei, Shuai Wan, Zhiyong Wang, Zongyuan Ge, Vincent Lam, and Dagan Feng. 2020. Keyframe extraction from laparoscopic videos via diverse and weighted dictionary selection. IEEE Journal of Biomedical and Health Informatics 25, 5 (2020), 1686–1698

    work page 2020

  34. [35]

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. 2023. Null-text inversion for editing real images using guided diffusion models. InIEEE Conference on Computer Vision and Pattern Recognition . 6038–6047

    work page 2023

  35. [36]

    Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. 2024. ReVideo: Remake a Video with Motion and Content Control. Advances in Neural Information Processing Systems 37 (2024), 18481–18505

    work page 2024

  36. [37]

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2021. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv:2112.10741 (2021)

    Pith/arXiv arXiv 2021

  37. [38]

    Shuanlong Niu, Yaru Peng, Bin Li, Yuanhong Qiu, Tongzhi Niu, and Weifeng Li. 2024. A novel deep learning motivated data augmentation system based on defect segmentation requirements. Journal of Intelligent Manufacturing 35, 2 (2024), 687–701

    work page 2024

  38. [39]

    Wenqi Ouyang, Yi Dong, Lei Yang, Jianlou Si, and Xingang Pan. 2024. I2VEdit: First-Frame-Guided Video Editing via Image-to-Video Diffusion Models. In SIG- GRAPH Asia 2024 Conference Papers . 1–11

    work page 2024

  39. [40]

    Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. 2023. Zero-shot image-to-image translation. In Association for Computing Machinery Special Interest Group on Computer Graphics and Interactive Techniques. 1–11

    work page 2023

  40. [41]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. 2022. Hierarchical text-conditional image generation with clip latents.arXiv:2204.06125 1, 2 (2022), 3

    Pith/arXiv arXiv 2022

  41. [42]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition . 10684–10695

    work page 2022

  42. [43]

    Liao Shen, Tianqi Liu, Huiqiang Sun, Xinyi Ye, Baopu Li, Jianming Zhang, and Zhiguo Cao. 2024. DreamMover: Leveraging the Prior of Diffusion Models for MM ’25, October 27–31, 2025, Dublin, Ireland. Zichi Liu, Yinggui Wang, Tao Wei, and Chao Ma Image Interpolation with Large Motion. In European Conference on Computer Vision. Springer, 336–353

    work page 2024

  43. [44]

    Roop Singh, Himanshu Mittal, and Raju Pal. 2022. Optimal keyframe selection- based lossless video-watermarking technique using IGSA in LWT domain for copyright protection. Complex & Intelligent Systems 8, 2 (2022), 1047–1070

    work page 2022

  44. [45]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli

  45. [46]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations. arXiv:2011.13456 (2020)

    Pith/arXiv arXiv 2020

  46. [47]

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. 2023. Plug-and-play diffusion features for text-driven image-to-image translation. In IEEE Conference on Computer Vision and Pattern Recognition . 1921–1930

    work page 2023

  47. [48]

    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. arXiv:2305.18264 (2023)

    Pith/arXiv arXiv 2023

  48. [49]

    Wen Wang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. 2023. Zero-shot video editing using off-the-shelf image diffusion models. arXiv:2303.17599 (2023)

    Pith/arXiv arXiv 2023

  49. [50]

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. 2023. Tune-a- video: One-shot tuning of image diffusion models for text-to-video generation. In IEEE Conference on Computer Vision and Pattern Recognition . 7623–7633

    work page 2023

  50. [51]

    Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han

    work page

  51. [52]

    Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. 2023. Rerender A Video: Zero-Shot Text-Guided Video-to-Video Translation. InSIGGRAPH Asia 2023 Conference Papers. 1–11

    work page 2023

  52. [53]

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. 2022. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv:2203.03605 (2022)

    Pith/arXiv arXiv 2022

  53. [54]

    Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xingang Pan. 2024. Diff- Morpher: Unleashing the Capability of Diffusion Models for Image Morphing. In IEEE Conference on Computer Vision and Pattern Recognition . 7912–7921

    work page 2024

  54. [55]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional control to text-to-image diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition. 3836–3847

    work page 2023

  55. [56]

    Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. 2023. Inversion-based style transfer with diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition . 10146– 10156

    work page 2023

  56. [57]

    Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. 2024. Towards consistent video editing with text-to-image diffusion models. Advances in Neural Information Processing Systems 36 (2024)

    work page 2024

  57. [58]

    Jie Zhao, Johan Edstedt, Michael Felsberg, Dong Wang, and Huchuan Lu. 2024. Leveraging the Power of Data Augmentation for Transformer-based Tracking. In IEEE/CVF Winter Conference on Applications of Computer Vision . 6469–6478

    work page 2024

  58. [59]

    Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. 2020. Learning data augmentation strategies for object detection. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16 . 566–583

    work page 2020

  59. [2015]

    In International Conference on Machine Learning

    Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning . 2256–2265

    work page

  60. [2024]

    International Journal of Computer Vision (2024), 1–20

    Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision (2024), 1–20

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.