acceptodds
Under review as a conference paper at ICLR 2027

FReD: Frequency-Resolved Distillation for Video Generation

Abstract

Few-step distillation accelerates video generation by reducing the number of denoising evaluations needed at inference. Distribution-matching approaches primarily achieve this by aligning teacher and student distributions in latent space, but such alignment does not explicitly assign teacher guidance to individual spatial frequency bands. To investigate this issue, we analyze generated and real videos in the spatial frequency domain and observe distinct discrepancies in their low- and high-frequency components. These discrepancies motivate frequency-resolved distillation with dedicated supervision for low- and high-frequency predictions. Specifically, we decompose the outputs of a frozen teacher model and a trainable fake score model into two frequency bands while retaining the complete input context, and the difference within each band supervises the corresponding component of the same native student prediction, extracted using a non-learnable filter. We further introduce overlap-aware loss weighting to regulate supervision in the transition region where the low- and high-frequency components overlap. Our method adds band-residual losses to the training code and incurs little training overhead, while preserving the original sampler and leaving the number of denoising steps and inference cost unchanged. Experiments across text-to-video, image-to-video, and autoregressive video generation tasks demonstrate improved generation quality and closer spectral agreement with real videos.

Primary area:
generative models
Keywords:
video generation, distillation, frequency decomposition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_U247OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper builds FReD on top of the distribution-matching distillation framework of DFD (Data-Forcing Distillation, Chen et al. 2026). Using a pair of non-learnable complementary Gaussian masks, it splits the student's full latent prediction — together with the predictions of the frozen teacher and the fake-score model — into a low-frequency and a high-frequency path in the spatial frequency domain, and constructs a separate residual loss for each. A shared exponential weight, driven by the degree of overlap between the two masks, then down-weights the loss contribution of the transition band. The method only changes how the loss is computed at student-update time: the sampler, the number of denoising steps, and inference cost are entirely unchanged.

Validation covers text-to-video, autoregressive video generation, and image-to-video, all with 1.3B-scale Wan models, compared against DMD2, DP-DMD, and DFD. The paper reports six VBench quality metrics, visual and camera-pose diversity, training and inference time, and three ablations — overlap weighting, band composition, and mask bandwidth.

Critical strengths

S1. The paper is well organized with clear demonstration of the motivaiton. The gap between reference and generated frames is first quantified in the frequency domain, showing that preserving magnitude and energy is not sufficient to preserve decoded appearance. Every design choice at the decomposition end has a corresponding ablation. The validation end covers three tasks and three ablation groups.

S2. The proposed method yielded modest improvements without much extra overhead. Under the same checkpoint and the same 40-step budget, continued DFD training does not move at all (86.46 → 86.44) while this paper reaches 87.23, with Imaging Quality +2.42 while all three baselines sit near 68 (Table 5). Inference time is unchanged, and only 5 lines of code replace DFD's 2. The appendix also volunteers two restrictive derivations.

Critical weaknesses

W1. Novelty is insufficient: band splitting and frequency reweighting are both precedented, and a near-concurrent, more complete version exists Designing separate losses per frequency band is routine — SDKD (arXiv 2507.02939, 2025) decouples teacher latents into high/low bands and distills them separately; an AAAI 2026 cross-modal distillation work decomposes features with FFT high/low-pass masks, using MSE on low frequencies and LogMSE on high. Frequency reweighting in diffusion distillation also has a 2023 precedent that this paper misses: CVPR 2023 Diffusion Probabilistic Model Made Slim (arXiv 2211.17106) applies a DFT to teacher and student predictions and weights by a negative power of spectral amplitude to amplify the high-frequency loss, motivated by exactly the same "high frequencies are missing" argument. More importantly, the paper almost fully overlaps with SAP-DMD (arXiv 2609.29116, 2026-09-24): same quantity acted upon, same per-frequency reweighting, same zero inference overhead — two instances of one idea, differing only in that SAP-DMD is adaptive while this paper is fixed. This paper's PDF is dated two days later, so I do not count it as a missed citation, but this paper is the weaker of the two, and their conclusions conflict: SAP-DMD suppresses low-frequency dominance, whereas this paper leaves low frequencies entirely unsuppressed.

W2. Soft-Overlap is a heuristic trick with no theoretical grounding By the paper's own two appendix identities, its gradient equals the DFD gradient times a coefficient that is strictly positive and never exceeds 1 — no new information, no change of direction. Under this premise, SO merely picks the shape of one scaling function. Its justification is only three design constraints (equals 1 when there is no overlap, decays monotonically with overlap, invariant to swapping the two masks), which admit infinitely many solutions; no derivation is given for this particular form, and the decay exponent is tested only at 0 and 1. No objective function that SO optimizes is stated anywhere, nor any optimality argument. Worse, it points the wrong way: the overlap region is in fact under-counted (about 0.5×), so restoring it would require the reciprocal (about 2), but SO takes 0.5 and pushes the coefficient down to 0.25.

Questions for authors

Q1 (method stability assessment) The three gains (+0.77 / +0.47 / +0.43) currently have no variance estimate at all — all ablations share one training seed, and the appendix states that generation seeds do not constitute an independent run. Please report mean ± std over at least 3 independent training runs, with significance tests for the text-to-video average score and for Imaging Quality (+2.42); and explain why metrics degrade monotonically with training steps (Imaging 70.78@40 → 68.98@100), plus the basis for reporting the 40-step point. If the favorable window is as narrow as 8–20 student updates, how is stability guaranteed?

Q2 (positioning against baselines) Please add at least one non-DMD-family 2026 baseline to position this work relative to the current SOTA. The related work already lists the following candidates, yet none of them is compared anywhere: rCM (Zheng et al., ICLR 2026, pp. 2582–2603), TMD (Nie et al., CVPR 2026, pp. 4645–4655), Causal-rCM (Zheng et al., arXiv:2606.25473), and One-Forcing (Feng et al., arXiv:2605.23458). Note that rCM and TMD are already formally published at ICLR and CVPR 2026 respectively, and neither appears in any comparison in this paper, so a reader has no way to judge where the work actually stands.

2 comments

We thank the reviewer for the careful assessment and constructive suggestions.

1. Contribution and related work (W1)

Our motivation was DFD-generated videos with hazy backgrounds, poor color rendition and unstable fine details. Section 3 diagnoses these failures through spatial-spectral analysis and latent interventions. Building on frequency-aware supervision in SDKD, FD-CMKD and Diffusion Probabilistic Model Made Slim, our contribution is a diagnosis of DFD-based video distillation and explicit control of its guidance allocation, evaluated across T2V, AR and I2V.

SAP-DMD addresses dominant low-frequency amplitudes in the DMD directional field using sample- and channel-adaptive modulation. We examine preservation across both frequency bands relative to reference videos. FReD uses a fixed frequency profile defined by complementary masks, band residual losses and overlap weighting. Both updates have a spectral-preconditioning interpretation; FReD focuses on guidance allocation across both frequency bands in DFD-based video distillation.

2. Soft-overlap weighting (W2)

SO attenuates the transition region shared by both loss paths. Its student-output gradient is

∇gLSO=N−1F−1[KF(d)],K=W(ML2+MH2).\nabla_g\mathcal L_{\mathrm{SO}}=N^{-1}\mathcal F^{-1}[K\mathcal F(d)],\quad K=W(M_L^2+M_H^2).

Equation (10) defines the weighted residual objective:

LSO=12∑b∈{L,H}mean⁡B,C,T,ω[W∣F(rb)∣2].\mathcal L_{\mathrm{SO}}=\frac12\sum_{b\in\{L,H\}}\operatorname{mean}_{B,C,T,\omega}[W|\mathcal F(r_b)|^2].

We use the overlap-dependent attenuation profile W=2−γOW=2^{-\gamma O}, O=4MLMHO=4M_LM_H, with γ=1\gamma=1, evaluated empirically in Table 4. Positive frequency-dependent weights preserve individual Fourier phases but can change the complete gradient's direction, redistributing existing guidance. At the crossover, SO reduces the coefficient from 1/2 to 1/4. Recovering DFD instead requires W=1/(ML2+MH2)W=1/(M_L^2+M_H^2), which cancels the reweighting.

At 40 continuation steps, SO reduces DD from 67.61% (768/1136) to 66.20% (752/1136), a 1.41-point decrease, while TF/MS remain comparable. Table 4 reports the accompanying quality gains.

3. Independent training runs and training length (Q1)

FReD is reported as mean ±\pm sample SD over three independent training runs. DFD uses its officially released checkpoint as a fixed reference. Avg. is the mean of SC, BC, AQ, IQ, TF and MS; DD is separate.

MethodT2V Avg.AR Avg.I2V Avg.
DMD285.82783.57384.003
DP-DMD86.11783.55786.887
DFD86.4684.0487.83
FReD87.243±0.12187.243\pm0.12184.617±0.10584.617\pm0.10588.360±0.08988.360\pm0.089

For T2V Avg., the mean gain over fixed DFD is 0.783: a two-sided one-sample t-test gives t(2)=11.208t(2)=11.208, p=0.00787p=0.00787, and a 95% CI of [0.482, 1.084] for the gain. FReD's IQ runs are 70.78, 69.63 and 69.82 (main-run gain: 2.42), giving 70.077±0.61670.077\pm0.616 versus fixed DFD 68.36: mean gain 1.717, t(2)=4.823t(2)=4.823, p=0.0404p=0.0404, CI [0.185, 3.248]. These tests condition on fixed DFD scores, assume approximately Gaussian FReD run scores, and use unadjusted p-values (n=3n=3).

The updated DD results are below; parentheses give dynamic/total videos. The T2V FReD entry is the 40-step main result.

MethodT2VARWan I2V
DMD282.04 (932/1136)93.75 (480/512)27.51 (383/1392)
DP-DMD80.37 (913/1136)88.28 (452/512)26.01 (362/1392)
DFD62.32 (708/1136)80.66 (413/512)14.44 (201/1392)
FReD66.20 (752/1136)82.42 (422/512)16.02 (223/1392)
Real70.69 (803/1136)85.35 (437/512)22.99 (320/1392)

FReD improves DD over DFD by 3.87/1.76/1.58 percentage points, computed from unrounded counts. This full-method comparison is distinct from the SO ablation.

We chose 40 continuation steps empirically. With SO, IQ at 40/60/80/100 steps is 70.775/70.797/69.803/68.982 (Table 4). We hypothesize that the inherited DFD query mismatch, where fake scores use noised student predictions and teacher scores use noised real data, contributes to less reliable guidance during continued refinement.

4. Additional baselines (Q2)

The four added baselines use 500 training iterations. FReD rows are main-experiment references; Avg. excludes DD.

TaskMethodSCBCAQIQTFMSDDAvg.
T2VrCM95.4994.9761.1669.9196.9598.5173.8386.17
T2VTMD96.4695.7661.6269.8798.0298.9242.8786.78
T2VFReD95.9295.5364.0170.7898.0399.0866.2087.23
ARCausal-rCM93.8594.0057.7167.2196.5298.6469.2484.66
AROne-Forcing48.0079.8729.5868.8490.2393.42100.0068.32
ARFReD91.5593.2961.2766.5196.3498.0882.4284.51

FReD has the highest T2V Avg. in this comparison. In AR, Causal-rCM has a slightly higher Avg. (84.66 vs. 84.51), while FReD has higher AQ and DD, showing metric-dependent trade-offs.

Thank you again for your detailed assessment and constructive suggestions. Our response clarifies the relationship to prior work and the role of soft-overlap weighting, and provides results from independent training runs together with additional baseline comparisons. We would be grateful for your feedback on these explanations and results, particularly whether any of your original concerns require further clarification. We would be happy to address any remaining questions.

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

The paper claims treating high frequency and low frequency components equally in DMD style few step video generation might degenerate the performance. To address this issue, the authors propose Frequency-Resolved Distillation, which trains the model using separate guidance for high- and low-frequency components

Critical strengths

Strength one: The paper tries to address an important challenge: training a stable, real-time autoregressive video generation model.

Critical weaknesses

Weakness one The intuition behind the proposed approach is not sufficiently clear. The analysis on the left side of Figure 1 operates at the level of pixel-space frequency components. However, generation quality may depend more strongly on high-level semantic features, whose relationship to this frequency decomposition is not established. The right side of Figure 1 shows that perturbing frequency components can distort images, but this observation does not clearly motivate providing separate guidance for the two frequency bands. As I understand the method, the two components are ultimately added back together. The approach therefore appears to remove cross terms between the components and then partially reintroduce them through soft-overlap loss weighting. Why this treatment should improve distillation remains unclear.

Weakness two The evaluation might be critically flawed. The ablation results in Table 4 raise concerns about the role of the overlap term. Without soft-overlap loss weighting, the model performs poorly on temporal flicker and motion smoothness. Does this suggest that the overlap term is particularly important for modeling motion?

Furthermore, only six of the seven VBench dimensions discussed are reported; the dynamic degree metric is missing. How does removing the overlap term affect this metric? This is particularly relevant because, as discussed above, the proposed method appears to effectively downweight the overlap term. The reported improvements could therefore involve a trade-off with motion dynamics that the current evaluation does not reveal.

Questions for authors

Question one Could you report the complete set of VBench metrics Question two Could you clarify your motivation? Question three Line 1017 states that the evaluation uses 348 VBench image–prompt pairs with four generation seeds each, yielding 1,392 videos per method. How were these 348 image–prompt pairs selected?

6 comments

We thank the reviewer for the careful assessment and constructive questions. We explain the motivation and supervision mechanism and report Dynamic Degree results together with a matched soft-overlap (SO) ablation.

1. Motivation and frequency supervision (W1, Q2)

Our motivation came from grayish or hazy backgrounds and unstable fine details observed during DFD distillation. This prompted us to examine how supervision across spatial frequencies affects broad appearance and local detail. Frequency bands describe spatial scales; their contribution is assessed through controlled interventions and ablations.

Section 3 combines two diagnostics: Figure 1 (left) measures the spatial spectra of decoded frames, while Figure 1 (right) intervenes directly on generated latents and decodes them with the same VAE. Our training-time frequency operators also act in VAE latent space. Figure 3 provides qualitative video comparisons.

Although gL+gH=gg_L+g_H=g, both loss paths update one student output and backbone, and inference uses the complete prediction gg. Summing band losses differs from computing a loss after recombination:

\mathcal L_{\mathrm{band}}=\frac{\|r_L\|_2^2+\|r_H\|_2^2}{2N}.$$ The full loss contains the cross-residual term $2\langle r_L,r_H\rangle$ inside the squared norm. Equation (7) shows that the band loss modifies the student-output gradient through $M_L^2+M_H^2$. SO further weights this update: $$O=4M_LM_H,\quad W=2^{-\gamma O},\quad K_{\mathrm{SO}}=W(M_L^2+M_H^2).$$ Both band losses contribute in the transition region; SO attenuates their shared contribution without reintroducing the cross-residual term. At $M_L=M_H=1/2$, the update factor is $1/2$ without SO and $1/4$ with the default $\gamma=1$, versus $1$ for the full-prediction loss. Disabling SO sets $W=1$ while retaining the overlapping Gaussian masks. This makes the spectral allocation of guidance explicit and adjustable. Table 6 evaluates band-specific supervision; Table 7 varies bandwidth with SO disabled; Table 4 isolates SO. These ablations provide empirical evidence for how the supervision design affects generation quality and consistency. ### 2. Complete evaluation and SO's effect on motion (W2, Q1) Together with SC, BC, TF, MS, AQ and IQ in Tables 1-3, the following DD results cover all seven dimensions discussed in the review. Avg. remains the six-metric mean; DD is reported separately. DD (%) is calculated from the number of videos classified as dynamic divided by the total evaluated. Parentheses show these counts. |Method|T2V|AR|Wan I2V| |---|---:|---:|---:| |DMD2|82.04 (932/1136)|93.75 (480/512)|27.51 (383/1392)| |DP-DMD|80.37 (913/1136)|88.28 (452/512)|26.01 (362/1392)| |DFD|62.32 (708/1136)|80.66 (413/512)|14.44 (201/1392)| |FReD|66.20 (752/1136)|82.42 (422/512)|16.02 (223/1392)| |Real reference|70.69 (803/1136)|85.35 (437/512)|22.99 (320/1392)| Relative to DFD, FReD increases DD by 3.87, 1.76 and 1.58 percentage points on T2V, AR and Wan I2V, computed from unrounded counts. Its DD is also closer to the corresponding real-video reference in all three settings. Thus, the quality improvements over DFD accompany a higher proportion of dynamic videos. The matched SO ablation at 40 continuation steps is: |Configuration|TF|MS|DD (%)| |---|---:|---:|---:| |Without SO|98.031|99.101|67.61 (768/1136)| |With SO|98.026|99.081|66.20 (752/1136)| The SO-enabled entry is our main T2V setting. Enabling SO reduces DD by 1.41 percentage points (16 videos), while TF and MS remain comparable. Alongside the SC, BC and AQ improvements in Table 4, this identifies a trade-off between visual quality and the proportion of dynamic videos at this training length. The full FReD-versus-DFD DD gain of 3.87 points measures the combined effect of the method's components, distinct from SO's isolated effect. ### 3. Image-prompt pairs (Q3) We use 348 image-prompt pairs from the VBench release. Four different generation seeds per pair yield $348\times4=1{,}392$ videos per method, as stated at line 1017.

Thanks for the detailed rebuttal from the author. For Q3, VBench has 944 prompts. Did you mean VBench-I2V? But VBench-I2V has 355 prompt-image pair, not 348. Could you clarify which exact dataset are you using? For Q2, as shown in tables, SO hurt the DD by 1.41, and FReD underperforms DMD2 and DP-DMD a lot. My concern is not fully addressed by the results. At this point. I would increase my rating to 2.

Thank you for reconsidering our work and increasing your rating. We appreciate your rigorous and thoughtful review, particularly your careful examination of the evaluation protocol and motion results.

Q3: Dataset and the 348-pair evaluation subset.

We used the official VBench-I2V dataset, which contains 355 image-prompt pairs. DFD reports a VBench test set containing 348 images in Sec. 5.1.2 of its paper. To match that evaluation-set size, we randomly selected 348 of the 355 pairs and used four generation seeds per pair, yielding 1,392 videos per method. All compared methods were evaluated on this selected subset. The number 348 therefore denotes our evaluation subset; the source dataset is VBench-I2V with 355 pairs.

Q2: Dynamic Degree and the quality-motion trade-off.

In the matched SO ablation at 40 continuation steps, enabling SO reduces DD from 67.61% (768/1136) to 66.20% (752/1136), a decrease of 1.41 percentage points. TF/MS remain comparable, alongside the visual-quality improvements reported in Table 4. We interpret this result as a quality-motion trade-off.

A higher DD does not necessarily imply more plausible motion. The official implementation uses RAFT optical flow to classify each video as dynamic or static and reports the fraction classified as dynamic. It does not directly assess whether the detected motion is physically plausible or appropriate for the scene. Texture instability and artifacts may contribute to estimated motion, so DD is informative about detected dynamics but cannot establish motion quality by itself.

In our qualitative comparisons, some DMD2 and DP-DMD outputs exhibit flickering textures, artifacts, and implausible movement. These observations suggest that part of their measured dynamics may reflect undesirable temporal variation.

We evaluated real videos to provide additional context for the motion levels:

TaskDMD2DP-DMDFReDReal
T2V82.0480.3766.2070.69
AR93.7588.2882.4285.35
Wan-I2V27.5126.0116.0222.99

All values are DD percentages. Relative to this reference, FReD is closer to Real in T2V, has the same absolute gap as DP-DMD in AR, and has a larger gap than both baselines in Wan-I2V.

We hypothesize that, within the real-data-based training inherited from DFD, teacher guidance on noised real data helps suppress implausible temporal variation, while the student's denoising-based training on noised real samples can also attenuate legitimate motion.

Reviewer_r1iKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes the difference between generated and real videos in the spatial frequency domain and proposes frequency-resolved distillation for video generation. The method uses non-learnable spatial Gaussian filters to constructe separate losses for low- and high-frequency components and introduces Soft-Overlap Loss Weighting (SO). The paper demonstrates improved quality over Data-Forcing Distillation (DFD).

Critical strengths
  1. The proposed method is simple and practical, requiring only additional filtering and loss weighting during training to achieve significant performance improvements, while maintaining training and inference times comparable to those of DFD.
  2. Table 5 compares FReD with continued DFD training from the same checkpoint under an identical additional training budget. While DFD shows little change in average quality, FReD improves the metrics. This controlled comparison provides evidence of FReD's effectiveness.
  3. The authors provide a clear rationale for the frequency weighting design, along with a detailed mathematical derivation.
Critical weaknesses
  1. The paper applies the same weighting scheme to both frequency bands without comparing it against asymmetric weighting schemes.
  2. DFD reports scores on metrics such as Dynamic Degree, VisionReward, and Instruction Following, but this paper does not provide comparisons with DFD on these metrics.
Questions for authors
  1. Please provide theoretical justification or experimental evidence for adopting symmetric weighting of high- and low-frequency components. Could asymmetric weighting of high- and low-frequency components lead to better performance?
  2. Please provide a comparison with DFD on metrics such as Dynamic Degree, VisionReward, and Instruction Following.
  3. Please provide more high-resolution close-up comparison with DFD to better substantiate the claimed improvements in fine details and temporal coherence.
6 comments
AuthorsOpenReview ↗

response to r1

We thank the reviewer for the constructive comments on frequency weighting, complementary metrics, and qualitative comparisons.

1. Symmetric and asymmetric weighting (W1, Q1)

The diagnostics in Section 3 (Figure 1), together with the qualitative comparisons in Figure 3, motivate supervising both spatial-frequency bands. The band-specific ablation (Table 6) shows complementary effects: High-only supervision improves TF/MS relative to Low-only, whereas Low-only yields higher SC, BC, AQ, and IQ. These are endpoint weight choices (0,1) and (1,0), respectively. Joint supervision achieves the highest scores on five of the six metrics, supporting its use for both consistency and visual quality.

We additionally evaluated asymmetric weights on Wan2.1 T2V. The overall loss scale is matched across all settings below; (λL,λH)(\lambda_L,\lambda_H) specifies the relative weighting of the low/high bands. The joint reference is from Table 6.

WeightsSCBCTFMSAQIQ
(0.4,0.6)95.832195.435297.752298.768863.885270.8533
(0.6,0.4)95.886295.433297.763298.780162.735970.7032
Joint (1,1)95.921695.500697.783098.786563.900170.8702

Joint supervision scores highest on all six metrics in this comparison. The asymmetric settings remain close in SC, BC, TF, MS, and IQ, while AQ varies more. These results support our choice of equal band coefficients.

2. Dynamic Degree, VisionReward, and instruction following (W2, Q2)

We now report the additional metrics, using DFD's officially released pretrained checkpoint. Dynamic Degree (DD, %) is calculated from the number of videos classified as dynamic. Parentheses show dynamic/total counts; Wan-I2V uses 348 pairs with four seeds each (1,392 videos).

MethodT2VARWan-I2V
DMD282.04 (932/1136)93.75 (480/512)27.51 (383/1392)
DP-DMD80.37 (913/1136)88.28 (452/512)26.01 (362/1392)
DFD62.32 (708/1136)80.66 (413/512)14.44 (201/1392)
FReD66.20 (752/1136)82.42 (422/512)16.02 (223/1392)
Real reference70.69 (803/1136)85.35 (437/512)22.99 (320/1392)

Relative to DFD, FReD increases DD by 3.87, 1.76, and 1.58 percentage points on T2V, AR, and Wan-I2V, respectively, computed from unrounded counts. Its DD is also closer to the real-video reference in all three settings. DD complements the quality and temporal-consistency metrics by measuring the proportion of dynamic videos.

We also report VisionReward (VR) and VideoAlign Text Alignment (TA), using TA to assess instruction following through adherence to the prompt. Both are higher-is-better. We verified sample coverage, means, and paired intervals across all 13,872 scoring records. Below, Δ=FReD−DFD\Delta=\mathrm{FReD}-\mathrm{DFD}; CIs are 95% paired confidence intervals for the difference.

TaskMetricDFDFReDDelta95% paired CI
T2VVR0.1150200.116599+0.001578[-0.000433,+0.003590]
T2VTA1.9601331.940693-0.019440[-0.044967,+0.003590]
Wan-I2VVR0.1390520.138327-0.000725[-0.001606,+0.000120]
Wan-I2VTA0.7385840.741911+0.003327[-0.004860,+0.011350]
ARVR0.1030140.106948+0.003934[+0.002159,+0.005735]
ARTA1.9553761.979405+0.024029[+0.008947,+0.039768]

AR improves on both metrics, with both intervals entirely above zero. T2V and Wan-I2V show small differences with mixed directions, and all four intervals include zero. Thus, the strongest evidence of improvement on these complementary metrics is in AR.

3. High-resolution qualitative comparisons (Q3)

We provide high-resolution close-ups against DFD and corresponding video comparisons in the supplementary material (response-all3_dfd_vs_fred_with_prompts.mp4) for closer inspection of fine details and temporal coherence.

Thank you again for your careful review and constructive feedback. In our response, we provided additional asymmetric-weighting experiments, Dynamic Degree, VisionReward, and text-alignment evaluations, as well as qualitative video comparisons in the supplementary material. We would appreciate your thoughts on these additional results and whether they address your concerns. Please let us know if any aspect of our response would benefit from further clarification.

Thank you for the authors' detailed responses to my comments. The responses to W1 & Q1 and W2 & Q2 have addressed my concerns regarding symmetric weighting, the evaluation metrics, and comparisons with DFD and other methods. However, these results raise a few additional questions. I would appreciate your clarification on the following points to help me decide whether to maintain my current rating.

  1. How to explain the observation that high-only supervision yields higher TF and MS scores than low-only supervision, whereas low-only supervision yields higher SC, BC, AQ, and IQ scores?
  2. How to explain the larger gap between FReD’s DD and the real-video reference compared with the corresponding gaps for DMD2 and DP-DMD in the Wan-I2V setting?
  3. How to explain the observation that FReD improves both VR and TA over DFD in the AR setting, whereas the changes in T2V and Wan-I2V are small and mixed in direction?

Comments on the paper

1 comment

We thank the reviewers and the Area Chair for their time. We summarize the evidence on three central issues: why FReD supervises both frequency bands, how it affects motion, and how its gains hold up across training runs and baselines. FReD addresses the loss of both coarse appearance and fine detail in DFD-based video distillation. SAP-DMD addresses an imbalance within the DMD update, where low-frequency components dominate the guidance. Our study asks how to recover both the scene layout and color and the fine textures that DFD can lose relative to real videos. Table 6 shows why both bands matter. High-only supervision yields better temporal flickering and motion smoothness scores, while low-only supervision yields better subject consistency, background consistency, aesthetic quality, and imaging quality. In our additional Wan2.1 T2V ablation, equal weighting scored highest on all six metrics compared with low/high ratios of 0.4/0.6 and 0.6/0.4 at the same overall loss scale. See Table 6 and Symmetric and asymmetric weighting (W1, Q1). FReD increases Dynamic Degree over DFD across all three tasks, while soft-overlap weighting introduces a measurable trade-off. FReD increased DD over DFD by 3.87 percentage points on T2V, 1.76 on AR, and 1.58 on Wan-I2V. In the matched T2V ablation at 40 continuation steps, enabling soft overlap lowered DD from 67.61% to 66.20%, alongside the gains in subject consistency, background consistency, and aesthetic quality reported in Table 4. On the complementary metrics, FReD improved both VisionReward and text alignment over DFD in AR, with both paired 95% confidence intervals above zero. Changes in T2V and Wan-I2V were small and mixed. See Complete evaluation and SO’s effect on motion (W2, Q1) and Dynamic Degree, VisionReward, and instruction following (W2, Q2). Three independent FReD training runs and four additional baselines provide a broader comparison. Across the three runs, FReD achieved a T2V six-metric average of 87.243 ± 0.121 (mean ± sample SD), compared with 86.46 for the released DFD checkpoint. We also evaluated four additional baselines at 500 training iterations each. On T2V, the main FReD run scored 87.23, compared with 86.17 for rCM and 86.78 for TMD. On AR, FReD scored 84.51, compared with 84.66 for Causal-rCM and 68.32 for One-Forcing. See Independent training runs and training length (Q1) and Additional baselines (Q2). Our responses to the reviewers address the remaining questions and provide further details on the benchmark configurations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.