acceptodds
Under review as a conference paper at ICLR 2027

ϕ-Noise: Training-Free Temporal Video Conditioning via Phase-Based Noise Manipulation

Abstract

Latent video diffusion models generate videos by progressively transforming Gaussian noise into realistic samples conditioned on text or visual inputs. However, existing conditioning methods often require additional training and computational overhead. Motivated by recent findings on the importance of frequency components in generative models, we propose a simple, training-free approach for motion-conditioned video generation by injecting low-frequency phase information from a reference video directly into the diffusion noise latents. Our method transfers motion cues without modifying the model architecture or inference pipeline. Using several applications, we demonstrate effective control over both appearance and dynamics in generated videos, while achieving competitive or superior results compared to more complex conditioning approaches.

Primary area:
generative models
Keywords:
Generative Models, Video Generation, Motion Conditioning, Motion Control, Frequency Aware, Zero Shot

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8GF4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

φ-Noise is a training-free way to condition a video diffusion model on a reference video by editing only the initial noise. Noise and reference latent are transformed with a DFT, the k lowest-frequency phases of the noise are replaced with the reference, and a mask scales the low band by 1/γ and the rest by β so that total energy is unchanged. The temporal DFT serves image-to-video motion transfer and cut-and-drag; the spatial DFT serves text-to-video motion transfer and image generation. Results are reported on Wan2.2-14B over a 60-video suite, with qualitative transfers to LTX-2 and SDXL.

Critical strengths
  • The method costs one FFT on the initial noise, with no training, inversion, attention surgery or test-time optimisation, and no added latency. Every baseline compared against is more expensive.
  • The presented method is generally useful: three tasks, two video backbones and an image model, with spatial and temporal conditioning in one formulation.
  • The text-to-video result is strong where it matters most for a noise-space method. Among the motion-transfer baselines in Table 1, φ-Noise has the best imaging quality, subject consistency, smoothness and LPIPS-T, and the second-best flow error. The competitors all degrade base-model quality markedly; φ-Noise does not. Reporting the unconditioned Wan rows is the right control, and it carries that claim within a fixed backbone.
Critical weaknesses
  • No variance on a 60-video suite, and k may be tuned per video.
  • The mechanism is observed rather than explained, and two of its descriptions are the wrong way round.
Questions for authors
  1. Are k and γ chosen per video? If so, by what criterion, and what does Table 1 look like with one fixed setting per task?
  2. Can you report variance over seeds for Table 1, and quantitative image-to-video results against MotionClone?
  3. Does phase specifically matter? How does replacing the full low-frequency band, magnitude and phase, under the same Φ compare with phase-only substitution?
  4. For clarification: is the DFT in Eqs. 3, 8 and 9 one-sided or two-sided? As written, substituting for f ≤ k over indices 0…t−1 would break conjugate symmetry, which a one-sided implementation handles automatically.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 8GF4

Dear reviewer 8GF4,

We thank the you for the detailed and constructive evaluation, particularly for recognizing the computational simplicity, broad applicability, and strong text-to-video performance of -Noise. We appreciate the observation that our approach maintains favorable image quality and temporal consistency while achieving competitive motion alignment.

We agree that the main remaining questions concern the robustness of the results and the precise interpretation of the frequency-domain mechanism.

We address some concerns shared with other reviewers in a message above meant for all reviewers: Hyperparameter selection and robustness, I2V quantitative comparisons, and the benefit of using phase only for conditioning.

Variance Across Seeds

We report variance across seeds in the following table:

TaskModelCLIP-T ↑Aes ↑Img ↑│LPIPS-T ↓Flow-E ↓Subj-C ↑Smooth ↑Dyn-D ↑
Cut & Dragbase model0.308 ±0.0430.652 ±0.0180.644 ±0.027│0.116 ±0.075181.10 ±352.620.942 ±0.0080.978 ±0.0030.647 ±0.120
GWTF0.314 ±0.0410.620 ±0.0210.637 ±0.021│0.097 ±0.075152.81 ±336.530.942 ±0.0090.981 ±0.0050.647 ±0.114
TTM0.311 ±0.0420.647 ±0.0210.653 ±0.022│0.110 ±0.109102.39 ±299.170.948 ±0.0120.978 ±0.0050.705 ±0.114
Ours (optimal)0.313 ±0.0380.637 ±0.0180.627 ±0.025│0.171 ±0.113101.49 ±164.600.918 ±0.0140.964 ±0.0060.764 ±0.106
-------------------------------
T2V MTbase model0.312 ±0.0050.604 ±0.0130.705 ±0.009│0.062 ±0.008103.26 ±38.250.955 ±0.0080.979 ±0.0030.645 ±0.095
DiT-Flow0.319 ±0.0050.526 ±0.0120.611 ±0.019│0.112 ±0.01094.60 ±32.010.931 ±0.0080.973 ±0.0030.935 ±0.037
DMT0.314 ±0.0070.530 ±0.0130.581 ±0.019│0.114 ±0.01167.23 ±26.680.914 ±0.0110.963 ±0.0040.871 ±0.062
MotionClone0.304 ±0.0060.548 ±0.0160.646 ±0.021│0.204 ±0.02167.92 ±27.330.864 ±0.0150.919 ±0.0110.903 ±0.051
Ours (optimal)0.302 ±0.0060.546 ±0.0120.683 ±0.014│0.075 ±0.01161.75 ±23.410.952 ±0.0070.980 ±0.0030.709 ±0.090
-------------------------------
I2V MTbase model0.332 ±0.0300.647 ±0.0590.709 ±0.062│0.082 ±0.056119.80 ±315.690.947 ±0.0360.977 ±0.0220.679 ±0.476
MotionClone0.330 ±0.0300.575 ±0.0660.613 ±0.105│0.172 ±0.114108.55 ±266.870.846 ±0.0920.946 ±0.0450.857 ±0.356
Ours (optimal)0.323 ±0.0380.602 ±0.0890.646 ±0.126│0.142 ±0.078167.59 ±329.390.872 ±0.1150.959 ±0.0290.929 ±0.262

We note that as each example uses a different seed the variance is calculated over all examples.

DFT Formulation and Conjugate Symmetry

Thank you for identifying this important mathematical detail. We agree that real-valued noise requires conjugate-symmetric frequency coefficients for a real-valued inverse transform.

In our implementation, we use the real FFT (rFFT), which operates on the non-redundant, one-sided frequency spectrum (such that K=⌊T/2⌋+1K = \lfloor T/2 \rfloor + 1), followed by the corresponding inverse transform (irFFT). Phase substitution is performed on this one-sided representation, and conjugate symmetry is implicitly handled by the inverse transform, ensuring real-valued reconstructed noise. Thus, our implementation avoids the conjugate-symmetry issue raised by the reviewer.

We will revise Eqs. 3, 8, and 9 to make the one-sided frequency indexing and symmetry treatment explicit, resolving the ambiguity in the current notation.

We will also revisit the mechanistic descriptions to distinguish what is directly demonstrated by the experiments from the intuition motivating our design.

We appreciate these detailed suggestions. We believe that clarifying the implementation and providing additional hyperparameter and multi-seed evaluations will strengthen the evidence for the method's robustness.

We welcome further discussion and any additional suggestions or clarifications you might have for our work.

Reviewer_KFG5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces φ-Noise, a training-free framework for motion and structural conditioning in video generation through manipulation of the initial noise. Given a reference video, the method substitutes selected low-frequency Fourier phases of the Gaussian initialization with those extracted from the encoded reference, using spatial or temporal frequency decomposition. It then applies spectral rescaling that preserves total initial energy to mitigate artifacts introduced by phase substitution.

The paper demonstrates three applications: text-conditioned motion transfer, text plus first-frame motion transfer, and cut-and-drag generation from crude animations. Experiments primarily use Wan2.2, with additional demonstrations on LTX and SDXL. The results demonstrate effective motion control and competitive performance, with task-dependent trade-offs in flexibility, consistency, and visual quality.

Critical strengths

1. Simple, training-free conditioning with low computational overhead.

The method requires no fine-tuning, architectural modifications, or additional denoising operations, making it straightforward to integrate into pretrained diffusion models with minimal expected overhead.

2. Useful empirical analysis and component ablations.

The phase and energy analysis provides useful motivation for the proposed approach. Component ablations demonstrate the effectiveness of the complete construction, while parameter studies illustrate the trade-offs between conditioning strength and generation quality.

3. Flexibility across applications and models.

The same framework supports text-conditioned motion transfer, first-frame-conditioned motion transfer, and cut-and-drag generation, including flexible articulation beyond rigid input motion. Additional demonstrations on LTX and SDXL suggest broader applicability.

4. Competitive quantitative performance.

The method achieves strong motion alignment while maintaining favorable image quality and temporal consistency, particularly in text-conditioned motion transfer. Cut-and-drag results are also competitive, demonstrating increased dynamics with some trade-offs in consistency and visual quality.

Critical weaknesses

1. The specific benefit of phase-only conditioning is insufficiently isolated.

Prior work explored low-frequency initialization (FreeInit), reference-phase conditioning (NeuralRemaster), and training-free magnitude-and-phase substitution (Colorful-Noise). While making phase-based conditioning effective on frozen video models is a meaningful contribution, Table 3 does not establish the advantage of transferring phase alone rather than both phase and magnitude. A controlled comparison on the same backbone would better isolate the benefit of this design choice.

2. Backbone differences complicate attribution of comparative improvements.

Competing methods use different video-generation backbones, making it difficult to distinguish improvements due to the conditioning mechanism from those due to the underlying generator. In particular, φ-Noise uses Wan2.2-14B, whereas GWTF uses CogVideoX-5B. While the vanilla Wan comparisons are informative, matched-backbone evaluations and clearer baseline configurations would strengthen the comparative conclusions.

3. Parameter selection and evaluation protocol require clarification.

The frequency cutoff kk is selected empirically, but it is unclear whether parameters are fixed per task or tuned per example. This is particularly relevant given the sensitivity acknowledged in Section 7. Clarifying the parameter-selection procedure, per-task evaluation sets, random seeds, and Flow-Err computation would improve reproducibility and confidence in the reported results.

4. Limited quantitative comparison for text plus first-frame motion transfer.

The paper demonstrates effective conditioning through qualitative comparisons with MotionClone and quantitative ablations against vanilla Wan. However, the absence of a quantitative comparison against MotionClone limits the evidence for comparative performance in this application.

Questions for authors
  1. Phase-only conditioning: Could the authors compare phase-only substitution against transferring both phase and magnitude on the same backbone, using comparable conditioning strengths?

  2. Evaluation and parameter selection: Is the frequency cutoff kk fixed per task or tuned per example, and how is it selected? Please also clarify the evaluation subsets, random seeds, and Flow-Err computation.

  3. Quantitative I2V comparison: Could the authors provide a quantitative comparison with I2V MotionClone to complement the existing qualitative results?

  4. Backbone configurations: Could the authors clarify the backbone and generation settings used for each baseline, particularly GWTF, and consider a matched-backbone comparison?

2 comments
AuthorsOpenReview ↗

Response to Reviewer KFG5

Dear reviewer KFG5,

We thank you for your review, and for recognizing key strengths of our method as well as pointing out weaknesses that can potentially improve our paper. We address concerns in the following.

Reviewer Shared Concerns

Most of the concerns - Benefit of phase-only conditioning, Parameter selection and backbone clarifications, and I2V quantitative results, were raised by other reviewers as well. To prevent repetitiveness, we posted a message above for all reviewers addressing these concerns. We address additional concerns below.

Additional Evaluations Details

We evaluate our method across three settings: Cut-and-Drag (C&D), Image-to-Video Motion Transfer (I2V-MT), and Text-to-Video Motion Transfer (T2V-MT).

C&D comprises 20 examples from the Time-to-Move (TTM) dataset. Each example consists of a still image in which objects are moved along user-defined trajectories to produce a coarse motion-reference video.

I2V-MT comprises 32 examples sourced from DAVIS, Videvo, and in-the-wild videos. For each example, we pair a reference video with a corresponding reference image and text prompt. The reference images and prompts are collected or generated using a pipeline that combines an LLM and an image-editing model.

T2V-MT comprises 30 examples from LOVEU-TGVE-2023. We use videos from this benchmark together with object-replacement prompts to evaluate the transfer of motion to new subjects while preserving the original video dynamics.

For random seeds, we initialized the initial random seed to 4242 and generated noise inputs for each example iteratively without reinitializing it during the process to prevent the usage of the same seed multiple times which might cause bias.

We appreciate any further suggestions and clarifications you might have following our response.

Reviewer_KFG5OpenReview ↗

Response to the rebuttal

Thanks to the authors for the detailed response and additional experiments. I think the phase-only ablation and new I2V comparisons help address two of my main concerns.

I still have reservations about the reliance on manual per-video parameter selection, particularly given the mixed results with fixed parameters. The lack of matched-backbone comparisons also remains unresolved, although the additional clarifications are helpful.

Nevertheless, I find the simplicity of the training-free approach, its flexibility across applications, and the demonstrated results meaningful. I therefore maintain my weak accept (3), although with some reservations.

Reviewer_fXg2OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ϕ-Noise, a training-free method for reference-conditioned video generation. It injects reference-video information into the initial noise through low-frequency phase substitution, followed by spectral rescaling to preserve total energy. Experiments demonstrate motion transfer and Cut & Drag on Wan2.2, with additional qualitative results on LTX and SDXL.

Critical strengths
  • Interesting conditioning idea: The paper explores frequency-domain manipulation of initial noise for video motion control. NeuralRemaster uses a similar phase-replacement operation with finetuning; making such conditioning effective on unchanged pretrained video generators is an interesting and useful contribution.

  • Simple and practical: The method requires no finetuning, architectural changes, or interventions during denoising, making it easy to integrate into pretrained generators.

  • Promising results and versatility: Table 1 reports competitive motion alignment, with the lowest Flow-E for Cut & Drag and second-lowest for T2V motion transfer among the compared methods. The additional LTX and SDXL examples suggest applicability beyond one backbone.

Critical weaknesses
  • Missing comparisons between conditioning signals from different frequency component: The paper does not compare phase-only, magnitude-only, and full magnitude-and-phase replacement within the same frequency band. These comparisons are needed to establish the specific benefit of using reference phase for motion control.

  • Overstated distributional claims: The claims that energy balancing “re-whitens” the noise and maintains compatibility with the training distribution are stronger than the evidence supports. Total-energy preservation does not gurantee white Gaussian statistics, although the experiments support improved denoising stability.

  • Mismatch between analysis and experiments: Section 3 analyzes temporal FFT, while the main T2V results use spatial FFT. The temporal analysis does not directly explain the spatial variant’s results. Separate analyses and comparisons are needed, including justification for using spatial or temporal FFT rather than joint spatiotemporal 3D FFT.

  • Unclear attribution of quality improvements: Competing methods use different backbones, making it difficult to determine whether the reported visual-quality advantages come from the proposed conditioning method or the underlying generator. More controlled comparisons are needed to separate these effects.

Questions for authors
  • How do phase-only, magnitude-only, and full magnitude-and-phase replacement compare on the same backbone and frequency band, with appropriately normalized inputs?

  • What evidence supports the re-whitening and distributional claims beyond total-energy preservation and improved denoising stability?

  • Can the authors separately analyze spatial and temporal conditioning and compare them with joint spatiotemporal 3D FFT? What motivates the choice for each task?

  • Which backbone and sampling settings are used for each baseline? Are comparisons on a common backbone available, where feasible, to isolate the benefit of the conditioning method?

2 comments
AuthorsOpenReview ↗

Response to Reviewer fXg2

Dear reviewer fXg2,

We thank you for your careful evaluation and constructive feedback. We appreciate the recognition of the novelty of phase-based conditioning, the simplicity of our training-free approach, and its strong motion-alignment performance. We address the concerns in the following.

Reviewer Shared Concerns

Some raised concerns - Benefit of phase-only conditioning, and backbone clarifications, and are shared with other reviewers. To prevent repetitiveness, we posted a message above for all reviewers addressing these concerns. We address additional concerns below.

Re-Whitening Claims:

We thank the reviewer for pointing out the distinction between preserving spectral energy and maintaining Gaussian white-noise statistics. The manuscript has been revised accordingly.

We agree that energy preservation alone does not guarantee that the modified noise follows the original Gaussian distribution or retains statistical whiteness.

To further illustrate our claim that the input noise need not be purely white, we conducted an additional experiment in which we explicitly violated the whiteness property of the Gaussian input while preserving its marginal statistics. Specifically, we colored (statistically) the white Gaussian noise by manipulating its amplitude distribution in the frequency domain, inducing different spectral profiles with slopes f∈{−1,−2,1,2}f\in\lbrace -1, -2, 1, 2\rbrace. We then applied the same denoising process to these colored Gaussian signals.

For fractional (negative) slopes (stronger low frequencies and attenuated high frequencies), the model produced reasonable results. In contrast, for linear/polynomial (positive) slopes, the model collapsed, yielding meaningless or heavily distorted outputs.

To explain this behavior, in Figure 7 in the Appendix we plot the PSD of the initial noise used. As can be seen, colored noise signals with amplified low-frequencies lead to model collapse. Moreover, our proposed noise manipulations (dashed lines) ensure that the spectrum remains relatively flat while preserving the total energy (norm) of the initial white Gaussian noise.

This demonstrates two key points:

  1. Strict whiteness is not a mandatory requirement for the diffusion model to function properly.
  2. The capacity to produce meaningful results under a positive-slope amplitude distribution aligns directly with our energy-balanced normalization. Since this normalization scales down low frequencies while scaling up high frequencies, it effectively "shifts" the amplitude distribution toward a positive slope.

Temporal vs. Spatial vs Spatio-Temporal

Since -noise is used to introduce motion conditioning, the temporal dimension is a natural candidate for this purpose. However, as discussed in Section 3.4, our experiments showed that temporal phase information alone is insufficient to transfer structural consistency. We therefore introduce an additional source of structural information. For I2V and Cut & Drag tasks, we found that first-frame conditioning provides sufficient structural guidance to ensure consistency, enabling clean motion transfer. In T2V, where no explicit structural information is available, we explored using the spatial dimensions' phase information as an alternative structural source. Interestingly, we found that spatial phase information alone can provide not only spatial structure but also temporal consistency, as each consecutive frame is implicitly connected to its preceding one. Although this might suggest that the temporal dimension is unnecessary for I2V and Cut & Drag, using spatial phase information in these settings proves ineffective, as it introduces unwanted spatial information unrelated to the first-frame condition.

Motivated by these observations, we further investigated the use of the spatio-temporal domain. However, for first-frame-conditioned tasks, this approach is unsuitable because it introduces additional structural information that conflicts with the intended conditioning. For T2V, we observed that it over-constrains the initial noise, causing the model to largely disregard the conditioning signal. The quantitative results are presented in the table below and will be included in the paper's appendix.

FFT DimsCLIP-T ↑Aes ↑Img ↑│LPIPS-T ↓Flow-E ↓Subj-C ↑Smooth ↑Dyn-D ↑
Spatio-Temporal0.2750.2680.217│0.12682.620.6940.8691.000
Temporal0.2460.5190.737│0.15844.340.9480.9870.833
Spatial (Ours)0.3020.5460.683│0.07561.750.9520.9800.709

Bold and italic denote best and second-best, respectively. For each experiment we report the optimal k and combinations.

We appreciate any further suggestions and clarifications you might have following our response.

Thanks the authors for their detailed and clear rebuttal. My main concerns are well-addressed. I would be happy to increase my score to weak accept.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to the Reviewers' Main Concerns

Dear reviewers,

We thank you for your effort, constructive feedback, and patience, as running experiments for video generation is a lengthy process. Your suggestions are important for improving our work. We are publishing an initial response addressing most reviewer concerns to initiate a discussion. Additional experiments regarding other concerns and responses to individual reviewers will be posted soon. We invite the reviewers to view the revised manuscript following the additional experiments (all changes are marked in blue font).

I2V-MT Quantitative Comparisons

We agree with the need for I2V-MT quantitative comparisons for the evaluations completeness. The updated comparison contains an additional evaluation of -Noise against Wan2.2-I2V (our base model) and MotionClone-I2V. We also invite the reviewers to observe the qualitative comparisons and results presented in the paper and the supplemental video for a broader aspect. The additional quantitative comparisons can be observed below in Table 1.

Hyper-parameter Selection and Robustness

We wish to clarify that our method does require light tuning of parameters per video to achieve optimal balance between motion conditioning and aesthetic results. This is directly linked to the fact that different conditioning videos have different low-frequency energy levels, which are a central part of our method. To represent the best of our method in the quantitative comparison we chose to manually select the parameters for each video out of 8-10 constant parameter options for each task.

As we agree with the importance of the matter, we conducted an evaluation of various k and combinations across the evaluation set. Full comparisons can be found in a dedicated table in Appendix D.1. Additionally, to reflect the abilities of -noise using a constant parameter decision we add our best single-parameter evaluation to the main comparison table. The added values can be seen below in Table 1. As can be observed, while using constant values is less optimal than the per video selection, they still show competitiveness in most metrics against other methods.

Phase-only vs. Magnitude Only vs. Magnitude + Phase

To address the reviewers' concerns, we first clarify our motivation for using phase information as guidance. As established in the literature, phase primarily carries structural information, whereas magnitude captures more general signal properties, such as color, which may hinder clean motion transfer when used for conditioning. Thus, focusing on phase allows us to isolate temporal structure while avoiding unwanted information.

To validate this intuition, we evaluated the two additional variants suggested by the reviewers: Phase + Magnitude and Magnitude Only, comparing against Phase Only. As shown in Table 2 below, phase-only conditioning achieves the best results. The reported are on the T2V-MT task, as it provides the most direct and independent setting for evaluating our conditioning mechanism.

Backbone and Experimental Setting Clarifications

Our quantitative evaluations are conducted using Wan-2.2. Since some baselines rely on different backbones with potentially varying visual quality, this may introduce noise into the image-based metrics evaluation. To provide a more comprehensive assessment of conditioning abilities, both temporally and aesthetically, we included qualitative comparisons against other methods and across prompts and seeds (Appendix D4, D5), as well as a supplemental video showing full generated sequences. While each method has its tradeoffs, these three complementary evaluations provide a clearer picture of their respective strengths and limitations.

To improve clarity, we added a base-model column to Table 1 and clarified this in the table caption.

Changes Added to the Manuscript (Highlighted in Blue)

We invite the reviewer to examine the revised manuscript. So far, we have:

  • Explicitly identified and highlighted the backbone used by each competing method (Table 1).
  • Clarified how the FFT operation is applied to real-valued signals.
  • Clarified and justified our parameter-selection procedure for the evaluation set (Section 6.3).
  • Conducted an extensive evaluation across a wide range of kk and γ\gamma combinations to better understand the trade-offs associated with parameter selection (Appendix, Section D.1 and Table 4).
  • Added a detailed discussion of how our proposed normalization relates to noise whiteness, including the motivation behind it and its effect on the frequency power spectrum (Appendix, Section E and Figure 7).
  • Added quantitative comparisons for the I2V-MT task (Table 1).
AuthorsOpenReview ↗

Supporting Tables

Updated Full Comparison to Prior Work:

TaskModelBackboneCLIP-T ↑Aes ↑Img ↑│LPIPS-T ↓Flow-E ↓Subj-C ↑Smooth ↑Dyn-D ↑
Cut & Dragbase modelWan-I2V0.3080.6520.644│0.116181.100.9420.9780.647
GWTFCogVideoX0.3140.6200.637│0.097152.810.9420.9810.647
TTMWan-I2V0.3110.6470.653│0.110102.390.9480.9780.705
Ours (k=4, γ=30k{=}4,\ \gamma{=}30)Wan-I2V0.3100.6510.621│0.178109.590.9140.9720.800
Ours (optimal)Wan-I2V0.3130.6370.627│0.171101.490.9180.9640.764
----------------------------------
T2V MTbase modelWan-T2V0.3120.6040.705│0.062103.260.9550.9790.645
DiT-FlowCogVideoX0.3190.5260.611│0.11294.600.9310.9730.935
DMTZeroScope v20.3140.5300.581│0.11467.230.9140.9630.871
MotionCloneAnimateDiff (SD 1.5)0.3040.5480.646│0.20467.920.8640.9190.903
RopeCraftWan-T2V0.3130.5140.573│0.12453.030.9090.9690.925
Ours (k=3, γ=12k{=}3,\ \gamma{=}12)Wan-T2V0.2980.5910.688│0.024103.840.9830.9930.067
Ours (optimal)Wan-T2V0.3020.5460.683│0.07561.750.9520.9800.709
----------------------------------
I2V MTbase modelWan-I2V0.3320.6470.709│0.082119.800.9470.9770.679
MotionCloneAnimateDiff (SD 1.5)0.3300.5750.613│0.172108.550.8460.9460.857
Ours (k=2, γ=30k{=}2,\ \gamma{=}30)Wan-I2V0.3320.6180.655│0.09996.560.9120.9790.850
Ours (optimal)Wan-I2V0.3230.6020.646│0.142167.590.8720.9590.929

The base model is included for reference.

For a clearer demonstration of our method's performance in I2V-MT, we encourage the reviewers to refer to the additional qualitative comparisons presented in Figures 13–16 of the Appendix.

Phase / Magnitude Comparison (T2V):

VariantCLIP-T ↑Aes ↑Img ↑│LPIPS-T ↓Flow-E ↓Subj-C ↑Smooth ↑Dyn-D ↑
Magnitude only0.2210.2730.237│0.3112381.070.6980.9041.000
Magnitude + Phase0.2090.2550.208│0.12270.800.8650.9730.444
Phase only (ϕ-noise)0.3020.5460.683│0.07561.750.9520.9800.709

Bold and italic denote best and second-best, respectively.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.