DFVEdit: Conditional Delta Flow Vector for Zero-shot Video Editing

We extend our sincere gratitude for your visionary assessment, which recognized the theoretical novelty in our unification of editing and sampling. Your constructive suggestions on ablation studies helped crystallize the mathematical essence of our framework. Below are our responses:

Weaknesses

1. On Ablation Study of Number of Diffusion Steps $T$

We thank the reviewer for this suggestion. To analyze the sensitivity of DFVEdit to the total number of diffusion steps (T), we conducted an ablation study on Wan2.1-14B with $T ∈ \{10, 20, 50, 100, 150, 200 \}$. As shown in Table R5, both editing accuracy (measured by CLIP-T↑) and temporal consistency (E_warp↓, CLIP-F↑) stabilize when T≥50. For T < 50, insufficient flow field estimation leads to under-editing, manifesting as blurred or temporally unstable results. In contrast, increasing T beyond 50 yields diminishing improvements at the cost of higher computational overhead, and may result in a slight decrease in performance on background preservation. We therefore set $T=50$ as the default, achieving an optimal trade-off between efficiency and quality. This value is also consistent with common practice in high-fidelity video generation, where moderate step counts suffice under advanced solvers. The quantitative and qualitative visualization results will be included in the final version.

Table R5 Ablation Study on Number of Diffusion Steps $T$

2. On Qualitative–Quantitative Discrepancy

We acknowledge that VideoGrain has good performance on static details (e.g., floor textures) due to its use of fine-grained SAM masks for local optimization. However, this does not imply superior video editing performance. In dynamic consistency—our key focus—DFVEdit outperforms VideoGrain. For example, in the case of the red fox (Figure 4), VideoGrain alters motion pattern details, resulting in unnatural movement (see supplementary video), despite achieving good text alignment. And on stylization tasks ( without SAM masks), VideoGrain's structure preservation ability decreases. In contrast, DFVEdit preserves original motion dynamics while achieving accurate appearance editing, which is supported by lower temporal distortion (E_warp↓) and higher user scores in spatial-temporal consistency (Consistency↑).

3. On ICA vs. Traditional Attention Engineering

We thank the reviewer for this question. It is essential to clarify that ICA is not an attention engineering component, nor does it modify attention weights or feature activations. In contrast to attention-engineering-based approaches that directly manipulate keys, values, or attention maps, ICA only infers an implicit spatial mask from attention responses to guide the Conditional Delta Flow Vector (CDFV)—our core editing mechanism.

This fundamental distinction brings three key advantages:

• Preservation of feature distribution: ICA does not alter attention outputs or introduce residual perturbations, avoiding unintended distribution shifts in unedited regions.
• Computational efficiency: The mask is computed once per timestep via a forward pass. In contrast, attention engineering requires iterative recomputation and caching across layers.
• Architecture compatibility: As a lightweight plugin to CDFV, ICA does not rely on specific attention layouts (can be extracted from both cross-attention in U-Net and full-attention in DiT), enabling seamless deployment on modern Video DiT models where traditional attention editing faces integration and scalability challenges.

In essence, ICA enhances editing accuracy without engaging in attention feature manipulation, making it fundamentally different from and more scalable than conventional attention engineering.

Questions

Q1. On the Originality of Section 3.1

We sincerely thank the reviewer for this thoughtful question. We would like to clarify the motivation and contribution of Section 3.1. The proposed framework, which unifies sampling and editing through a continuous flow perspective, is inspired by the theoretical foundations of Score Matching [1] and Flow Matching [2], particularly the continuous flow interpretation of SDEs discussed in prior work (e.g., in the appendix of the Flow Matching literature). We adapt this idea to the video editing task, offering a new viewpoint that models editing as a latent space flow transformation. Equations 1–4 reframe Score Matching and Flow Matching within a unified continuous flow formulation, providing a clearer theoretical foundation for editing design. Furthermore, Equations 5–6 represent our original derivation, which unifies sampling and editing, explicitly connects flow prediction to practical editing operations, and directly motivates the design of the CDFV. While built upon prior theory, the extension offers a principled and insightful step toward more coherent and controllable video editing.

Q2. On the Term "Unbiased Estimation"

We thank the reviewer for this clarifying question. In our context, "unbiased estimation" means that the expected value of the Conditional Delta Flow Vector (CDFV) equals the true theoretical Delta Flow Vector (DFV) $\Delta v_t$ (as defined in L164). This ensures the CDFV accurately captures the intended conditional flow shift from source to target latents. The unbiasedness of CDFV relies on three key conditions: terminal distribution convergence, shared (coherent) noise, and fine discretization (discussed in Section 3, L170–176).

A biased estimation arises when these conditions are violated, such as:

• Excessive edit strength: When $|\hat{Z}_0 - Z_0| > 3\sigma$, the terminal latent $\hat{Z}_t$ may deviate from the standard Gaussian prior $\mathcal{N}(0, I)$ at $t=T$, breaking the terminal convergence assumption. This aligns with our stated limitation regarding extreme geometric deformations.

• Non-shared noise: If the source and target paths use independent noise realizations ($dW_{\hat{Z}} \ne dW_Z$), the stochastic components do not cancel, introducing noise-induced bias into CDFV.

• Large $\Delta t$: The unbiasedness holds strictly in the continuous-time limit ($\Delta t \to 0$). Large step sizes introduce discretization error, leading to estimation bias.

Limitations

We sincerely thank the reviewer for the valuable feedback. In the final version of the paper, we will add a dedicated discussion in the Limitations section to clarify that DFVEdit is primarily designed for appearance editing, and currently has limited support for non-rigid shape modifications or drag-based interactive editing. We will also cite and discuss the relevant works recommended by the reviewer [1–3] in both the Related Work and Limitations sections to strengthen the paper.

Paper Formatting Concerns

We thank the reviewer for the careful reading. We will revise Fig. 2 in the final version to include explicit annotations (1)–(5) corresponding to the key components in the pipeline. Additionally, the typo at Line 217—'CovideoX-5B'—has been corrected to 'CogVideoX-5B'.

[1] Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2020). Score-based generative modeling through stochastic differential equations.

[2] Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., & Le, M. (2022). Flow matching for generative modeling

We sincerely appreciate your precision-driven review, which highlighted ambiguities in the presentation. Your laser-like focus on clarity – from figure annotations to terminology definitions – has significantly enhanced the paper's technical communicability. Below are our responses:

Weaknesses

Method

1. About the difference from previous attention-engineering works.

• Difference on the conceptual foundation: We unified editing and sampling, and model editing as a continuous flow transformation in latent space via CDFV (Eq.8-12), whereas attention engineering heuristically manipulates feature correlations via Q/K/V modifications.
• Difference on the operational mechanism: CDFV performs end-to-end deformation and ICA only infers an implicit mask from cross-attention or full-attention maps to constrain CDFV’s spatial influence and SAM masks are also applied to CDFV optionally (used in multi-object shape editing), which means ICA and SAM masks are operated on the latent space without engaging in traditional attention engineering.
• Difference on the architectural impact: DFVEdit's flow-based formulation enables model-agnostic operation (have been verified on 2D U-Net in Table R1 and various Video DiTs), while most attention engineering works fail on Video DiTs due to architectural and computational constraints

Experiments

1. On Case Complexity

We appreciate the reviewer’s comment. Our evaluation covers diverse motion regimes. Due to the limitations of static figures, dynamic behaviors may not be fully conveyed. We invite the reviewer to consult the supplementary video, which clearly illustrates temporal coherence, editing fidelity, and motion dynamics. To evaluate motion preservation under diverse conditions, we designed experiments across four categories of motion complexity and articulation level:

• High Dynamics, Low Articulation: e.g., jeep stylization – large motion with minimal structural change.
• Moderate Dynamics, Moderate Articulation: e.g., man → robot – involving limb articulation and pose transitions.
• Low Dynamics, High Articulation: e.g., dogs → foxes – preserving fine, high-frequency non-rigid motions (ear flicks, blinking, facial expressions).
• High Dynamics, High Articulation: e.g., cyclist editing – combining fast motion, limb articulation, and micro-action transfer (see Extensive Results III in the supplementary video).

DFVEdit preserves spatio-temporal coherence from global trajectories to local pose details. In addition, thanks to its resource-efficient design, DFVEdit enables long-sequence editing (beyond typical 8–20 frame limits), enhancing motion continuity. We will include more long-duration, complex cases in the final version.

2. On ICA’s Effect in the Zebra Case

We thank the reviewer for the insightful observation. The zebra case illustrates a fundamental trade-off in ICA: without ICA, higher edit intensity (e.g., mane transformation) but severe background degradation; with ICA, Strong background preservation, yet potentially conservative editing on dynamic, ambiguous boundaries (e.g., flowing mane). This stems from current limitations in modeling diffuse edges under motion. Nevertheless, quantitative results (Table 1) show consistent gains in ‘M.PSNR’ and ‘LPIPS’, and qualitative examples (e.g., bear→polar bear in Figure 5, multi-object edit in Figure F3) confirm ICA’s critical role in visual fidelity and editing accuracy. The benefits of background protection outweigh the minor loss in edge edit strength in most cases. We will explicitly discuss this limitation and explore adaptive masking mechanisms in future work.

Writing/Delivery

We appreciate the reviewer's detailed feedback on the paper's clarity and readability. Below are the revisions we will implement in the final version:

(1) Caption and Text Supplement for Fig 1(c): "Normalized DFV" refers to the original DFV normalized using Min-Max scaling to the range [0,1], illustrating its relative intensity distribution across different sampling steps t and spatial positions. This figure reveals the evolution pattern of DFV from coarse contours to fine details, consistent with the diffusion sampling process, providing an intuitive motivation for a unified perspective on editing and generation.

(2) Definition of Terms at First Mention: We will add the full name and a brief explanation of "DFVEdit" at its first mention in the Introduction (line 69), enhancing readability.

(3) Redrawing of Figure 2 Overview: Given that the current Figure 2 is not conducive to understanding, we will thoroughly redraw it after rebuttal, emphasizing five key stages and ensuring all module locations and data flows are clearly labeled to enhance interpretability. The improved description is as follows:

1. Latent Initialization: Starting from the far left of the diagram, marked as (1), it clearly illustrates the process of obtaining noise-free initial latent variables from input videos through encoding.
2. Forward Flow Mapping: In the middle section of the diagram, marked as (2), it shows how image features are refined through a series of flow transformations from the latent variables in the previous step.
3. Condition Embedding Feeding & CDFV Computation: Parallel to the forward flow and positioned towards the upper right, marked as (3), it highlights how text condition is embedded into the model and the specific process of CDFV computation based on these conditions.
4. Latent Update: Returning to the top left of the diagram, marked as (4), it explains how latent variables are updated based on new information after condition embedding and CDFV computation.
5. Iterative Refinement → Output: At the bottom left of the diagram's first row, marked as (5), it depicts the final step—iterative refinement until generating high-quality output images.

(4) Symbol Consistency and Mechanism Clarification: t: Current diffusion time step; T: Total number of sampling steps (constant). In the ER module, γ represents the reweighting strength (see L205–215). The ER sub-figure shows: input text embeddings → key token identification (red box) → reweighting process (pink box on the right) → influence on Key/Value in ICA (left side), and a comparison of ICA responses before and after embedding enhancement (right side), visually explaining its mechanism.

Questions

Q1.How can you figure out which token is for stylization? (L209 and Fig. 2) By a manual segmentation?

Our token recognition process is fully automated, and it proceeds as follows:

Step 1: Syntax Marking - Users add special delimiters [] around target concepts (such as style words) in their editing instructions, for example, "[Watercolor style painting] of a castle at sunset".

Step 2: Token Mapping and Localization - The text is converted into a token sequence using a tokenizer. Simultaneously, the tokenizer maps the special delimiters to a token sequence, which serves as a matching substring to locate the positions of these delimiters within the token sequence, thereby automatically identifying the corresponding tokens. Once identified, the weights of these corresponding tokens are adjusted through reweighting (Eq.16).

Q2. This is a little minor but will the proposed methods work for other aspect ratios instead of only for squared (1:1) ones?

Regarding aspect ratio support, DFVEdit fully supports video input with any aspect ratio. The method itself does not have a square constraint on the spatial dimension of the input latent variables. We used squares in the experiment only for the consideration of uniform preprocessing and common formats. The final version will include examples of editing results for videos with non-1:1 aspect ratios (such as 16:9).

Limitations

We agree that the reasons for editing failures (such as insufficient fidelity of background details) are partly due to compression loss in the VAE, which is a common bottleneck in current latent space diffusion models, especially affecting the restoration of high-frequency textures and delicate structures. However, we also note that the design limitations of the method itself also play an important role. For example, in Appendix Figure F4, "chimpanzee → chimpanzee with sunglasses" can better preserve the details of the background banner, while "chimpanzee → chimpanzee with red hat" shows texture degradation of the background banner.

We sincerely thank you for your deep technical engagement with our work. Your meticulous analysis of experimental completeness and methodological details has significantly strengthened the paper's empirical foundations. Below are our responses to the concerns that have been raised.

Weakness:

1. On Insufficient Experiments:

(1) User Study: We agree on the importance of user studies for evaluating subjective editing quality. We apologize for not highlighting the user study results before. As detailed in Table 1 (corresponding metric description in p. 9, L 252–255) and Appendix (p. 3, L 48–52), we conducted user studies with 20 participants assessing 80 video-prompt pairs across stylization, shape, and attribute editing (1,600 ratings in total). Results show DFVEdit outperforms baselines in video-prompt alignment ("Edit"), frame quality ("Quality"), and spatio-temporal consistency ("Consistency").

(2) Image Editing Experiments: We appreciate the reviewer’s suggestion on evaluating image editing generalization. We conducted experiments on PIE-Bench [1] using Stable Diffusion 1.5, comparing DFVEdit with Instruct-pix2pix [2], PnP [3], and P2P [4] (PnP and P2P have adopted DirectInversion[1] techniques). Due to time constraints, we evaluated DFVEdit and T2I baselines on two representative subsets: 1_change_object_80 and 9_change_style_80, using standard metrics—Distance (overall structure coherence), PSNR/LPIPS/SSIM (background preservation), and CLIP-T (alignment)—following DirectInversion [1]. Results in Table R1 show the competitive performance and practical cross-modal generalization ability of DFVEdit. Our core contribution remains enabling efficient, zero-shot editing on modern Video DiT models, where DFVEdit overcomes the computational and architectural limitations of attention-based methods without requiring fine-tuning.

Table R1: Quantitative results on PIE-Bench [1] for image editing.

3) Additional Video Editing Baselines: We thank the reviewer for the suggestion. We have added comparisons with CogVideoX-V2V and VidToMe (with the same experiment settings in the comparison session). As shown in Table R2, DFVEdit outperforms both methods across all metrics. These results will be fully presented (including quantitative and qualitative) in the final version.

Table R2: Quantitative comparison with state-of-the-art video editing methods.

(4) Comparison Fairness Across Architectures:

We thank the reviewer for the insightful comment. To ensure fair comparisons:

• We evaluate SDEdit and CogVideoX-V2V on the same Video DiT backbone (CogVideoX-5B, Tables 1 and R2), where DFVEdit still outperforms.
• We have conducted experiments on a unified 2D U-Net (Stable Diffusion 1.5) for image editing (Table R1), demonstrating the effectiveness of DFVEdit independent of Video-DiT-specific advantages.

To the best of our knowledge, DFVEdit is the first method enabling efficient and effective zero-shot video editing on modern Video DiTs—including both score-matching (e.g., CogVideoX) and flow-matching (e.g., Wan2.1) models. Prior attention-based methods are often incompatible due to architectural and computational constraints. DFVEdit overcomes these via a lightweight, flow-field-based formulation.

(5) Results on Wan2.1-14B: We have included qualitative results using Wan2.1-14B in Figure 4 (p. 8) and Figure F5 (Appendix, p. 10). For completeness, we now provide quantitative results in Table R2, showing that DFVEdit achieves better performance when built upon this stronger Video DiT backbone.

Thank you for all your suggestions on experiments. Visualization results, however, can not be shown in this rebuttal phase due to rebuttal policy constraints. We will update all qualitative visualization results in the final version.

2. On Lack of Method Details

(1) DFVEdit Version: We used the full version of DFVEdit, including both the ICA and ER components. The effectiveness of these components is validated in the ablation studies (main text, §4.3), and further implementation details are provided in Appendix B.3, including Figure F1 and Figure F2, which illustrate ER ablation on shape editing and ICA extraction and visualization.

(2) Hyperparameter Selection: For Eq. 16, we conducted ablation studies (Appendix B.3, Figure 5a and Figure F1) to determine optimal values. These hyperparameters are stable and easy to tune, typically requiring only a single sweep across a small range. The parameters in Eq. 6 and Eq. 31 relate to the theoretical unification of editing and sampling dynamics. In practice, DFVEdit introduces only a few key hyperparameters: the fusion timesteps for ICA and the number of sampling iterations (T).

(3) Mask Usage: The mask is applied in Layer 16's cross-attention and is timestep-aware, following FreeMask's design, evolving from coarse to fine during sampling. This allows objects greater flexibility in shape deformation, leading to more natural editing boundaries compared to rigid masks (e.g., in VideoGrain or VideoDirector). While this may cause minor leakage into background regions, resulting in subtle background changes, our CDFV mechanism and ICA module effectively constrain such drift. In most cases, background preservation remains strong.

Questions

Q1&Q2 Additional Experiments and Method Details: Please refer to the weaknesses.

Q3. Seed Sensitivity and Reproducibility: The latent initialization stage is noise-free, meaning we transform a clean source video latent into a clean target latent. When computing the Conditional Delta Flow Vector (CDFV), the same Wiener process $dW$ ( the same random seeds) is applied to both $\hat{Z}_t$ (target latent at t) and $Z_0$ (source latent) In the forward flow (Eq. 1), which leads to the cancellation of stochastic noise in the output difference (Eq. 12). Additionally, each forward pass uses a new random seed, preventing bias accumulation from fixed noise patterns. We tested DFVEdit (on CogVideoX-5B) on 20 video-prompt pairs with five different repetitive experiments to evaluate stability across random seeds. Table R3 reports the mean and standard deviation of key metrics, indicating low variance and consistent performance. The stability arises from our CDFV design—only the deterministic score difference drives editing, and ICA further enhances consistency by anchoring background features.

Table R3: Performance stability across 5 repetitive experiments (mean ± std).

Q4. Limitations in Large-Scale Geometric Editing: We agree with the reviewer that DFVEdit, like most text-driven editing methods, struggles with large-scale geometric transformations (e.g., replacing a jeep with a tank or an elephant). Our method operates via a delta flow vector that models smooth transformations between source and target videos. It excels at stylization, attribute editing (color, texture), local shape changes, and substitutions with similar structure (e.g., car → truck), but is less suited to radical topological changes. This limitation stems from the flow-based formulation and the underlying DiT's reliance on source structure. Generating entirely new, complex geometries while preserving background and motion remains a fundamental challenge in video editing. We focus on high-fidelity, temporally consistent semantic editing with minimal user input. Future work may incorporate explicit structural guidance (e.g., depth, optical flow, layout maps) or layered compositing strategies for extreme shape changes.

[1] Ju, X., Zeng, A., Bian, Y., Liu, S., & Xu, Q. (2023). Direct inversion: Boosting diffusion-based editing with three lines of code. arXiv preprint arXiv:2310.01506.

[2] Brooks, T., Holynski, A., & Efros, A. A. (2023). Instructpix2pix: Learning to follow image editing instructions.

[3] Tumanyan, N., Geyer, M., Bagon, S., & Dekel, T. (2023). Plug-and-play diffusion features for text-driven image-to-image translation.

[4] Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., & Cohen-Or, D. (2022). Prompt-to-prompt image editing with cross attention control.

We are grateful for your incisive conceptual critique that challenged us to refine the philosophical grounding of DFVEdit. Your sharp perspective on zero-shot motivation prompted us to articulate broader value propositions that extend beyond efficiency gains. Below are our responses:

Weaknesses:

1. On Zero-Shot Design and Practical Value

• The strength of zero-shot methods lies not only in eliminating the need for training but also in their plug-and-play flexibility and model-agnostic applicability. This is especially valuable for non-expert users and real-world applications, where retraining for new tasks or a new model (e.g., rapidly evolving Video DiTs) is impractical due to closed-source models or high computational costs.
• DFVEdit offers a unified perspective on generation and editing through continuous manifold transformations in latent space, and proposed CDFV reveals that editing can be interpreted as a conditional flow field shift, providing theoretical insight beyond mere performance.
• While fine-tuning methods (e.g., Tune-A-Video[2], VideoP2P[3]) may achieve high performance on specific tasks, DFVEdit delivers strong editing quality with zero-shot convenience, matching or even exceeding them in temporal consistency and background preservation, as validated in comparison experiments with the fine-tuning method DMT[3] (see Figure 7, Table 1). To further validate this, we include more comparison results with fine-tuning baselines in Table R4.

Table R4: Comparison with fine-tuning-based methods.

2. On Method Design Simplicity and Necessity

We thank the reviewer for bringing this vital point to our attention. While simplicity is desirable, the three core components of DFVEdit—CDFV, ICA, and ER—are distinct and indispensable.

• CDFV establishes the theoretical foundation, modeling editing as flow field transformation in latent space—defining how to edit, and is the core contribution of our method.
• ICA prevents "local distribution drift" by implicitly constraining latent updates to target regions, preserving background and temporal consistency—ensuring what not to change, is a protection module for CDFV.
• ER enhances prompt alignment by adaptively reweighting key tokens in the text prompt, ensuring what to edit aligns with intent.

Ablation studies confirm that removing any component leads to significant degradation: ICA removal causes background artifacts; ER removal reduces prompt alignment and CLIP-T scores. Visual comparisons in the Appendix further validate their necessity. The framework introduces easy-tunable hyperparameters (ICA mask fusion timesteps, ER reweighting strength), which exhibit robustness across tasks. This demonstrates its practicality. In summary, DFVEdit’s design is function-driven: each component solves a specific, empirically validated challenge. The architecture reflects engineering clarity, balancing editing strength, consistency, and instruction alignment, without redundancy.

3. On Experimental Scope and Dynamic Editing

We appreciate the reviewer’s feedback on case diversity. DFVEdit supports not only style transfer but also:

• Single object shape editing (see Figure 3,' man -> robot '; Figure 4)

• Object addition/removal and attribute editing (Figure F4)

• Multi-object editing (Figure 3, 'the right dog -> the right fox'; Figure F3)

Due to the limitations of static figures, dynamic behaviors may not be fully conveyed. We invite the reviewer to consult the supplementary videos, which clearly illustrate temporal coherence and editing fidelity. DFVEdit maintains strong temporal consistency across all categories, from global trajectories to local pose details, outperforming existing zero-shot methods and, in several metrics, rivaling fine-tuned approaches (e.g., DMT, VideoP2P, Tune-A-Video, as shown in Table R4).

We clarify that this work focuses on appearance editing (as stated on p.6, line 221), not motion editing (e.g., changing "walking" to "dancing"). Such tasks require explicit motion control (e.g., via optical flow, 3D poses), and remain a key direction for future work.

4. On Fairness of Experimental Comparisons

We appreciate the reviewers’ insightful comments regarding experimental fairness across model architectures. Indeed, direct comparisons between methods built on different backbones, we rigorously address fairness concerns through:

• In comparison, we have adopted SDEdit (as shown in Figure 3 and Table 1) on the base model CogVideoX-5B, as well as the CogVideoX-V2V pipeline (presented in Table R2), for a fair comparison, which demonstrates that even on the same base model, DFVEdit still outperforms.

• We have conducted controlled experiments using a unified backbone: Stable Diffusion v1.5 (2D U-Net). In this setting, we evaluate DFVEdit’s core components (CDFV, ICA, and ER) in image editing tasks (as shown in Table R1), where the temporal modeling advantages are removed. Results show that DFVEdit significantly improves editing accuracy, detail preservation, and semantic alignment, demonstrating its general effectiveness independent of architectural strength.

To the best of our knowledge, we are the first work enabling effective and efficient zero-shot video editing on modern Video DiT architectures — including both score-matching-based (e.g., CogVideoX) and flow-matching-based (e.g., Wan2.1) formulations. Prior attention-based editing methods are often incompatible with DiTs due to architectural constraints and high computational cost. DFVEdit overcomes these limitations through a lightweight, flow-field-based formulation.

Thus, while comparisons across architectures require careful interpretation, our results highlight two complementary strengths:

• General applicability—validated under controlled 2D settings;
• Forward-looking capability—unlocking high-quality, zero-shot editing on powerful Video DiT models that prior methods cannot support.

[1] Liu, S., Zhang, Y., Li, W., Lin, Z., & Jia, J. (2024). Video-p2p: Video editing with cross-attention control.

[2] Wu, J. Z., Ge, Y., Wang, X., Lei, S. W., Gu, Y., Shi, Y., ... & Shou, M. Z. (2023). Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation.

[3] Yatim, D., Fridman, R., Bar-Tal, O., Kasten, Y., & Dekel, T. (2024). Space-time diffusion features for zero-shot text-driven motion transfer.