acceptodds
Withdrawn from ICLR 2027

RoDCache: Routing-Drift-Aware Caching for Accelerating Sparse MoE Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models incur substantial inference cost due to repeated execution of the denoising network across multiple timesteps. Caching methods reduce this cost by reusing results from previous denoising steps based on estimates of cache validity. However, in sparse MoE diffusion models, existing cache-validity signals overlook how routing drift in expert assignments and routing weights alters the active computation path and affects reuse reliability. To address this limitation, we present RoDCache, a training-free caching method that accounts for routing drift in cache-validity assessment. It comprises two components: Routing Drift Quantification (RDQ) and Routing-Drift-Constrained Scheduling (RDCS). RDQ quantifies model-level routing drift by aggregating changes in expert assignments and routing weights across tokens and MoE layers. Based on this signal, RDCS selects the longest reuse interval whose predicted routing drift does not exceed a tunable threshold. Across three sparse MoE diffusion models, \method achieves – speedups and consistently outperforms existing caching methods in PSNR, SSIM, and LPIPS, with PSNR gains of up to dB. Further analysis shows that RoDCache adapts computation allocation to prompt difficulty and improves quality consistency across prompts. Code will be released at https://anonymous.4open.science/r/RoDCache.

Primary area:
generative models
Keywords:
Sparse MoE Diffusion Models, Cache-based Acceleration, Routing Drift, Adaptive Caching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DyRjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose RoDCache, a method for accelerating sparse MoE diffusion models through adaptive caching. The method measures changes in expert assignments and routing weights between full computation steps and uses them to predict how long cached results remain valid. It requires no additional training and uses a single threshold to control the level of acceleration. Experiments show improved similarity to outputs obtained without caching while maintaining competitive generation quality and latency.

Critical strengths

Simple and effective design. RoDCache requires no additional training and uses a single tunable threshold to control cache reuse. It consistently improves PSNR, SSIM and LPIPS over evaluated caching baselines across three sparse MoE models while maintaining competitive generation quality and latency.

Critical weaknesses
  1. Unclear practical advantage. Better PSNR, SSIM and LPIPS show that generated images and videos are closer to those obtained without caching. These gains do not by themselves demonstrate better generation quality. CLIP/VBench Score and ImageReward remain largely comparable to SeaCache. Although RoDCache achieves lower latency in the reported settings, it is unclear whether this advantage would hold at comparable generation quality under other parameter choices. The evidence for a better tradeoff between generation quality and inference time therefore remains limited.

  2. Weak evidence for choosing routing drift. Figure 2 removes shared correct detections and normalizes the remaining counts to 100%. This could make a small difference between two nearly identical indicators look substantial. Without absolute counts, it is unclear how much routing drift actually improves on residual drift. This weakens the evidence behind a central design choice.

Questions for authors
  1. Figure 2 reports the relative proportions of prompts uniquely identified by each indicator, but does not provide the underlying counts. Could you report the total numbers of correctly identified prompts, ∣Qdev∩Qroute∣|Q_{\mathrm{dev}} \cap Q_{\mathrm{route}}| and ∣Qdev∩Qres∣|Q_{\mathrm{dev}} \cap Q_{\mathrm{res}}|, including their shared correct detections?

    The two indicators could identify almost identical sets of prompts, differing only in a small number of cases. Excluding shared detections and normalizing the remaining counts to sum to 100% could then make a small absolute advantage appear substantial. Reporting the shared and unique detection counts would help establish the magnitude of the improvement.

  2. Could you clarify whether the main objective is to reproduce the images and videos generated without caching or to preserve generation quality while accelerating inference? Are the improvements in PSNR, SSIM and LPIPS intended as the main empirical contribution? If so, please explain why closer similarity to outputs obtained without caching is practically important and distinguish these gains from improvements in generation quality in your claims. If the main objective is to preserve generation quality while reducing inference time, could you compare RoDCache with strong baselines such as SeaCache at the same latency using CLIP Score, ImageReward and VBench?

3 comments

Dear Reviewer DyRj,

Thank you very much for your thorough review and constructive feedback. We greatly appreciate the opportunity to address your concerns and questions.

W2: Weak evidence for choosing routing drift. Figure 2 removes shared correct detections and normalizes the remaining counts to 100%. This could make a small difference between two nearly identical indicators look substantial. Without absolute counts, it is unclear how much routing drift actually improves on residual drift. This weakens the evidence behind a central design choice.

Q1: Figure 2 reports the relative proportions of prompts uniquely identified by each indicator, but does not provide the underlying counts. Could you report the total numbers of correctly identified prompts, ∣Qdev∩Qroute∣|\mathcal{Q}_{\mathrm{dev}} \cap \mathcal{Q}_{\mathrm{route}}| and ∣Qdev∩Qres∣|\mathcal{Q}_{\mathrm{dev}} \cap \mathcal{Q}_{\mathrm{res}}|, including their shared correct detections?

The two indicators could identify almost identical sets of prompts, differing only in a small number of cases. Excluding shared detections and normalizing the remaining counts to sum to 100% could then make a small absolute advantage appear substantial. Reporting the shared and unique detection counts would help establish the magnitude of the improvement.

Following this suggestion, we report the absolute detection counts below. Figure 2 presents the equal-gg normalized exclusive shares, whereas the following counts aggregate all eligible settings. We have added these absolute counts and the corresponding analysis to Appendix A of the revised manuscript.

Detection setCountPercentage of Qdev\mathcal{Q}_{\mathrm{dev}}
Qdev\mathcal{Q}_{\mathrm{dev}}288,000100.00%
Qdev∩Qroute∩Qres\mathcal{Q}_{\mathrm{dev}} \cap \mathcal{Q}_{\mathrm{route}} \cap \mathcal{Q}_{\mathrm{res}}64,40322.36%
Qdev∩Qroute\mathcal{Q}_{\mathrm{dev}} \cap \mathcal{Q}_{\mathrm{route}}125,83543.69%
Qdev∩Qres\mathcal{Q}_{\mathrm{dev}} \cap \mathcal{Q}_{\mathrm{res}}102,28235.51%

Representative results at different cache-reuse lengths are:

ssTotalSharedRouting correctRouting rateResidual correctResidual rate
117,5203,8666,91139.45%6,39836.52%
615,1203,6807,24947.94%6,05040.01%
1212,2402,8365,57445.54%4,53937.08%
189,3601,8513,72939.84%2,79729.88%
246,4809411,98330.60%1,32820.49%

Based on these results:

  • Including shared detections, routing drift correctly identifies 125,835 instances compared with 102,282 for residual drift, providing 23,553 additional detections and an overall detection rate 8.18 percentage points higher. Its advantage is 2.93, 7.93, 8.46, 9.96, and 10.11 percentage points at s=1s=1, 66, 1212, 1818, and 2424, respectively.

Therefore, the advantage in Figure 2 is not an artifact of excluding shared detections and renormalizing a small number of remaining cases.

W1: Unclear practical advantage. Better PSNR, SSIM and LPIPS show that generated images and videos are closer to those obtained without caching. These gains do not by themselves demonstrate better generation quality. CLIP/VBench Score and ImageReward remain largely comparable to SeaCache. Although RoDCache achieves lower latency in the reported settings, it is unclear whether this advantage would hold at comparable generation quality under other parameter choices. The evidence for a better tradeoff between generation quality and inference time therefore remains limited.

Q2: Could you clarify whether the main objective is to reproduce the images and videos generated without caching or to preserve generation quality while accelerating inference? Are the improvements in PSNR, SSIM and LPIPS intended as the main empirical contribution? If so, please explain why closer similarity to outputs obtained without caching is practically important and distinguish these gains from improvements in generation quality in your claims. If the main objective is to preserve generation quality while reducing inference time, could you compare RoDCache with strong baselines such as SeaCache at the same latency using CLIP Score, ImageReward and VBench?

  • Objective. RoDCache aims to preserve generation quality while accelerating inference and minimizing deviations from full-computation outputs. Improvements in PSNR, SSIM, and LPIPS are a main empirical contribution in terms of fidelity, showing that RoDCache preserves full-computation outputs more faithfully than existing caching methods. We have revised the manuscript to distinguish fidelity from generation quality in our claims (e.g., Line 26 and Lines 101–102) and separately group PSNR, SSIM, and LPIPS under “Fidelity to Vanilla” and CLIP Score, ImageReward, and VBench under “Generation Quality” in the result tables.

  • Practical importance. Caching is typically used as a drop-in acceleration technique and should ideally preserve model behavior under the same prompt, random seed, and sampling configuration. Cache-induced changes in object count, spatial layout, rendered text, subject identity, local details, or video dynamics may break reproducibility and require existing generation and content-review pipelines to be revalidated. Such changes may not be fully captured by CLIP Score, ImageReward, or VBench because perceptually different outputs can receive similar aggregate scores.

  • Consistency with prior work. The empirical advantages of TeaCache [1], DPCache [2], SeaCache [3], and DiCache [4] are also primarily reflected by PSNR, SSIM, and LPIPS, while their CLIP Score, ImageReward, and VBench results generally remain comparable rather than showing consistent improvements.

  • Comparison with SeaCache. Our main table already compares the two methods at their closest latency settings. Across all slow and fast settings, RoDCache achieves lower latency, higher PSNR and SSIM, and lower LPIPS, while maintaining CLIP Score, ImageReward, and VBench comparable to SeaCache.

[1] Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model, Liu et al., CVPR 2025.

[2] Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache, Cui et al., CVPR 2026.

[3] SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models, Chung et al., CVPR 2026.

[4] DiCache: Let Diffusion Model Determine Its Own Cache, Bu et al., ICLR 2026.

Dear Reviewer DyRj,

Thank you again for your valuable feedback and for taking the time to review our work.

We would like to kindly follow up and ask whether our responses have adequately addressed your concerns. If there are any remaining questions or aspects that you would like us to clarify further, we would be more than happy to provide additional explanations.

We would also greatly appreciate any further thoughts you may have regarding our responses.

Thanks,

The Authors

Reviewer_utt4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper quantifies routing drift and uses it as a criterion for deciding whether to reuse cached features in video diffusion models.

Critical strengths

None

Critical weaknesses
  1. The finding is not novel, as MoECa [1] also uses routing drift as a metric for making cache reuse decisions. Although this paper quantifies routing drift at full-computation steps, this distinction appears to constitute only an incremental improvement.

  2. The lack of a performance comparison with MoECa, which takes a highly similar approach, is another major weakness.

[1] Li, Maoliang, et al. “MoECa: Aligning Feature Reuse with Expert Decomposition in Diffusion Transformers.” arXiv preprint arXiv:2606.15615 (2026).

Questions for authors
  1. Does the proposed method improve performance over MoECa?

  2. What are the key novel aspects and technical contributions that distinguish the proposed method from MoECa?

5 comments

Dear Reviewer utt4,

Thank you very much for your constructive feedback. We appreciate the opportunity to clarify the differences from MoECa and provide an additional comparison.

W1: The finding is not novel, as MoECa [1] also uses routing drift as a metric for making cache reuse decisions. Although this paper quantifies routing drift at full-computation steps, this distinction appears to constitute only an incremental improvement.

Q2: What are the key novel aspects and technical contributions that distinguish the proposed method from MoECa?

MoECa was already discussed in the “Caching for Sparse MoE Diffusion Models” paragraph of Section 2 (Related Work) in our original submission. Although both methods use routing information, their technical designs differ substantially:

  • Routing signal. Although both methods use an L1L_1-type difference to measure routing-weight changes for matched expert branches, MoECa uses this difference only as one component of a branch-level recomputation score and handles assignment changes separately through unmatched-branch recomputation. RoDCache instead jointly encodes expert assignments and routing weights in normalized routing vectors and aggregates their changes across all tokens and MoE layers. This design captures both discrete expert switches and continuous changes in routing strength on a common scale, producing a single network-wide signal for assessing overall cache validity. While both methods apply to token-choice routing, RoDCache’s routing-signal definition additionally supports expert-choice routing.

  • Adaptive scheduling. RoDCache uses its model-level routing signal to predict accumulated drift and adaptively determine the next full-computation step using a single threshold. MoECa instead retains a fixed mandatory full-refresh interval (N=4N=4 by default), with routing dynamics influencing branch-level recomputation rather than the refresh schedule.

  • Reuse overhead. At every reuse step, MoECa executes the router, matches current and cached expert branches, computes branch scores, and selectively recomputes branches. Its branch-level caches also incur additional runtime and memory overhead. RoDCache caches only the transformer residual and routing state, requiring no router execution, branch matching, or denoising-network computation during reuse steps.

Overall, RoDCache is not simply MoECa’s routing metric evaluated at full-computation steps. It introduces a unified model-level routing signal that jointly captures changes in expert assignments and routing weights across tokens and MoE layers. This distinction underlies RoDCache’s broader routing compatibility, adaptive scheduling, and lower reuse overhead, making it a distinct framework rather than an incremental improvement.

W2: The lack of a performance comparison with MoECa, which takes a highly similar approach, is another major weakness.

Q1: Does the proposed method improve performance over MoECa?

We additionally evaluated MoECa on HiDream-I1. Since the paper does not provide a code link, we used the public repository released under the authors’ organization as our reference. The repository does not include support for HiDream-I1. We therefore ported MoECa’s expert-cache architecture, branch-matching mechanism, and scoring mechanism to HiDream-I1 and used its default configuration. MoECa assumes token-choice routing and therefore cannot be directly applied to the expert-choice routing used by Nucleus-Image. Our port also runs out of GPU memory on LingBot-Video because its branch-level caches exceed the available GPU memory. We consequently report the comparison on HiDream-I1. We have incorporated the comparison into Table 2, with implementation and configuration details provided in Section 6.1 and the appendix of the revised manuscript.

MethodLatency (s) ↓PSNR ↑SSIM ↑LPIPS ↓
Vanilla144.41∞1.00000.0000
MoECa-default60.8515.1180.62360.3637
RoDCache-slow37.4821.1340.75630.2147
RoDCache-fast28.1416.9000.63960.3598

Based on these results, the following observations can be made:

  • Under both the slow and fast settings, RoDCache achieves substantially lower latency and consistently better PSNR, SSIM, and LPIPS than MoECa-default, demonstrating a clearly better fidelity–latency trade-off.

Dear Reviewer utt4,

Thank you again for your valuable feedback and for taking the time to review our work.

We would like to kindly follow up and ask whether our responses have adequately addressed your concerns. If there are any remaining questions or aspects that you would like us to clarify further, we would be more than happy to provide additional explanations.

We would also greatly appreciate any further thoughts you may have regarding our responses.

Thanks,

The Authors

  1. The authors present the use of routing drift itself as their first contribution in the paper, while acknowledging in the rebuttal that both methods use routing drift. I therefore believe that the revised paper does not adequately position this work relative to prior work, and that the first contribution cannot be claimed in its current form. Although I acknowledge the technical differences described in the authors’ response to Weakness 1, I consider these differences incremental.

  2. Table 3 of the MoECA paper reports lower latency and TFLOPs than TeaCache on HiDream, achieving a PSNR of 29 at a 2.93× speedup. However, the MoECA results added by the authors show substantially higher FLOPs and latency than TeaCache, with a PSNR of only 15. This discrepancy raises concerns about whether MoECA was implemented correctly and whether the comparison was conducted fairly.

For these reasons, I maintain my original score of 1.

Reviewer_dq2nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies caching for sparse MoE diffusion models, motivated by the limitation that existing cache-validity indicators overlook changes in expert routing. The authors observe that expert assignments and routing weights can change significantly across denoising steps, and that routing drift is a more reliable indicator of cache validity than residual drift. Based on this observation, they propose RoDCache, which quantifies routing drift across tokens and MoE layers and uses historical drift to adaptively determine cache reuse intervals.

Critical strengths
  1. The paper addresses a clearly defined problem with well-motivated observations.

  2. The experimental evaluation is comprehensive, covering multiple models, baselines, and evaluation metrics.

  3. The paper is well written and clearly organized. The proposed method also demonstrates noticeable improvements over existing baselines.

Critical weaknesses
  1. Ideally, routing drift should be both predictable, allowing future drift to be estimated from historical observations, and informative, reflecting differences in cache risk across prompts and denoising steps. However, RDQ aggregates routing changes across all tokens and MoE layers into a single scalar through uniform averaging. While this provides a simple and stable signal, it is unclear how much information relevant to cache validity is preserved after such aggressive aggregation. In particular, the resulting signal may primarily capture a common timestep-dependent trend rather than meaningful prompt-specific routing changes.

  2. Routing drift does not necessarily reflect the actual cache error, as changes in expert assignments or routing weights do not directly correspond to changes in expert outputs. Large routing drift may result in little output difference, while small routing drift does not guarantee accurate cache reuse. The paper does not sufficiently explain why routing drift should reliably indicate residual error, especially when expert outputs can change even with stable routing.

Questions for authors
  1. Could the authors compare different routing drift aggregation strategies (e.g., mean vs. top-k mean or other strategies)? More importantly, could they replace each prompt's routing drift with the average drift measured from other prompts over the same denoising interval, and compare performance at matched compute budgets?

  2. Could you provide a more direct analysis or experiment of the correlation between routing drift and actual residual error?

8 comments

Dear Reviewer dq2n,

Thank you very much for your constructive feedback. We appreciate the opportunity to further examine whether RDQ preserves prompt-specific information and how directly routing drift reflects cache error.

W1: Ideally, routing drift should be both predictable, allowing future drift to be estimated from historical observations, and informative, reflecting differences in cache risk across prompts and denoising steps. However, RDQ aggregates routing changes across all tokens and MoE layers into a single scalar through uniform averaging. While this provides a simple and stable signal, it is unclear how much information relevant to cache validity is preserved after such aggressive aggregation. In particular, the resulting signal may primarily capture a common timestep-dependent trend rather than meaningful prompt-specific routing changes.

Q1: Could the authors compare different routing drift aggregation strategies (e.g., mean vs. top-k mean or other strategies)? More importantly, could they replace each prompt's routing drift with the average drift measured from other prompts over the same denoising interval, and compare performance at matched compute budgets?

Thank you for this insightful suggestion. We conducted two additional experiments on HiDream-I1 to examine the aggregation strategy and the contribution of prompt-specific routing information. We have incorporated the corresponding results and analysis into Lines 500–509 and Appendix E of the revised manuscript.

  • Mean vs. Top-10% aggregation. We replaced the standard mean aggregation in RDQ with Top-10% Mean, which averages only the largest 10% of token-level routing changes, and compared the two methods at similar latency.
AggregationLatency (s) ↓CLIP Score ↑ImageReward ↑PSNR ↑SSIM ↑LPIPS ↓
Top-10% Mean29.8735.05731.206116.1440.61030.3781
Mean (RDQ)28.1435.05591.215016.9000.63960.3598

Mean (RDQ) achieves lower latency while increasing PSNR by 0.756 dB and SSIM by 0.0293 with LPIPS reduced by 0.0183. Top-10% Mean overemphasizes a small number of local routing changes, whereas mean aggregation captures changes distributed across the model and provides a more stable scheduling signal.

  • Shared Mean. We replace prompt-specific routing drift with the interval-wise average from 20 randomly sampled other prompts over the same denoising interval. This preserves the common timestep-dependent trend while removing routing information specific to the current prompt.
Routing signalLatency (s) ↓CLIP Score ↑ImageReward ↑PSNR ↑SSIM ↑LPIPS ↓
Shared Mean29.7034.99251.209916.1380.60400.3809
Mean (RDQ)28.1435.05591.215016.9000.63960.3598

At comparable compute budgets, Shared Mean performs worse in PSNR, SSIM, and LPIPS despite its higher latency. We further compare computation allocation under Mean (RDQ) and Shared Mean on the Hard-10% and Easy-10% prompts:

Routing signalHard FLOPs (T)Easy FLOPs (T)H/E ↑Hard PSNR ↑Hard SSIM ↑Hard LPIPS ↓Easy PSNR ↑Easy SSIM ↑Easy LPIPS ↓PSNR Gap ↓SSIM Gap ↓LPIPS Gap ↓
Shared Mean1555.381555.381.00013.2940.57610.404718.3710.63660.30905.0770.06050.0957
Mean (RDQ)1555.381384.291.12415.9800.63840.338319.1640.66420.31973.1840.02580.0186
  • Shared Mean assigns identical computation to hard and easy prompts because its routing signal depends only on the denoising interval rather than the current prompt. In contrast, Mean (RDQ) allocates more computation to hard prompts and less to easy prompts.
  • Mean (RDQ) achieves better overall fidelity, substantially improves the hard-prompt results, and produces smaller Hard–Easy gaps. These results demonstrate that RDQ preserves meaningful prompt-specific cache-risk information rather than merely capturing a common denoising-step trend.

W2: Routing drift does not necessarily reflect the actual cache error, as changes in expert assignments or routing weights do not directly correspond to changes in expert outputs. Large routing drift may result in little output difference, while small routing drift does not guarantee accurate cache reuse. The paper does not sufficiently explain why routing drift should reliably indicate residual error, especially when expert outputs can change even with stable routing.

Q2: Could you provide a more direct analysis or experiment of the correlation between routing drift and actual residual error?

We agree that routing drift is not equivalent to residual error. RDQ is intended to indicate cache risk rather than exactly estimate residual error.

Following the reviewer’s suggestion, we directly computed the Spearman correlation between routing drift and actual subsequent residual error using the traces of all 200 HiDream-I1 prompts, obtaining 207,000 valid samples. The confidence interval was calculated using 1,000 prompt-level bootstrap samples.

Valid samplesSpearman ρ ↑Bootstrap 95% CI
207,0000.968[0.962, 0.973]
  • Routing drift has a strong positive correlation with actual subsequent residual error. Although it does not fully determine residual error, it reliably identifies when cached residuals are more likely to become inaccurate, supporting its use as the cache-risk signal in RDCS.

Thank you for the additional experiments, which have addressed most of my concerns. However, I would like to clarify how the Spearman correlation of 0.968 was computed. Was it calculated by pooling samples across different timesteps and reuse intervals? If so, could the authors also report the correlation across prompts at fixed timesteps and interval lengths, to rule out the influence of common temporal trends?

Comments on the paper

1 comment

Dear Reviewers,

As we are approaching the end of reviewer-author discussions, please go through authors' responses and raise questions that you would like authors to reply.

Thanks,

AC

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

Related papers

Discussion (0)

Sign in to comment.