acceptodds
Withdrawn from ICLR 2027

PropCache: Propagation-Calibrated Step Caching for Diffusion Transformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Step caching accelerates diffusion transformers by reusing model outputs, but which steps to reuse determines fidelity at a fixed compute budget. Existing schedulers rely on runtime triggers, candidate replay, or budget-specific search instead of pricing each reuse's terminal impact. We propose PropCache, a training-free compiler built on a propagation price. It factorizes terminal impact into a local defect and a remaining-trajectory gain, then uses one calibration to price every candidate reuse and compile exact-budget schedules for all tiers. Across six flow-matching diffusion transformers spanning 1.3B to 14B parameters, dense and dual-expert backbones, and video and image generation, PropCache establishes the strongest evaluated fidelity–compute point at every backbone's aggressive tier. It lowers Learned Perceptual Image Patch Similarity (LPIPS) by 8–29% against LeMiCa across five video models, leads LeMiCa at all nine HunyuanVideo-13B budgets by 13.5–31.8%, and reaches 4.83× end-to-end speedup. On the full 943-prompt VBench suite, its matched-latency gain is significant at p = 0.000002. Anonymized code is available at https://anonymous.4open.science/r/propcache-review-5CA2/.

Primary area:
generative models
Keywords:
computer vision, video generation, diffusion models, diffusion transformers, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2BjMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PropCache, a step-caching scheduler for visual diffusion models. The key insight is that the final influence of a local caching error should take the amplification effects of the subsequent denoising steps into account. PropCache then measures the separated propagation gain for each timestep (per NFE scheduler setting), and use a dynamic programming algorithm to calculated a calibrated caching configuration, which aims in minimizing the final accumulated error under a fixed NFE. The paper demonstrates competitive fidelity-computation trade-offs across many image and video generation models.

Critical strengths

1. Clear presentation and a well-motivated problem formulation

  • The paper is clearly stated. The idea of measuring error propagation rate and then use it to search for a better step-caching scheduler is natural and easy to follow.

2. Accounting for error propagation is a valuable idea

  • Instead of inspecting on only local properties (e.g. adjacent frame differences / similarity), the method modeling the generation progress in a whole. trajectorial view. Which is more direct to the central point of preserving generation fidelity.

3. The offline, fixed-budget implementation is practically appealing.

  • The method only requires a pre-calibration measure for the propagation gain and a optimal programming for the time scheduler. Which introduce almost no extra computation cost in the following inference stage. Which is practically useful for large-scale deployment acceleration.
Critical weaknesses

1. The main advantage of PropCache is calibration-time efficiency

  • It seems that the propagation gain is close to a curve that decreases broadly over the time. Which results in a strategy that allocates more computation early and makes more cache later. Table 4 is particularly relevant: the searched exponential and power-law alternatives achieve lower mean LPIPS than the measured curve. This suggests that the main contribution may be obtaining useful temporal weights at low calibration cost, instead of an efficient-and-effective solution.

2. The high calibration cost of parametric baselines may depend on the search implementation

  • The calibration-efficiency comparison in Table 4 is made between directly measured propagation profiles and parameterized profiles selected through extensive LPIPS-based searching. It could be more convincing with a baseline that first fit an parameterized profile directly with the same measured propagation probes, and then compile these schedules using the DP algorithm. Such a comparison can help demonstrate the gain of the measured one against the parameterized one.

3. Connection to continuous timestep scheduling and missing baselines

  • The PropCache directly caches the velocity output of the flow matching model (instead of any intermediate states inside the network), which makes the method more like a grid-aligned timestep schedule. In fact, given a fitted A(t) and d(a,t), we can directly calculate a continuous timestep schedule under certain extended PropCache cost function. A comparison between PropCache and some few-step timestep scheduling methods (e.g., GITS [1], which also minimizes a truncation error of a series of trajectory intervals, and AYS [2] & OYS [3], which directly optimize a continuous timestep schedule) on performance and budgets can better disclose this.

[1] Chen et al., On the Trajectory Regularity of ODE-based Diffusion Sampling, ICML 2024.

[2] Sabour et al., Align Your Steps: Optimizing Sampling Schedules in Diffusion Models, ICML 2024.

[3] Zhange et al., Optimize Your Sampling: Tuned Diffusion Sampling with Bayesian Optimization, arxiv 2026.

Questions for authors

I'm still confused since though PropCache is formulated as a step-caching method, its main algorithmic contribution lies in selecting the denoising steps at which model evaluations are performed, rather than introducing a new caching operator. Which is more similar to a timestep scheduling method. I think this connection should be discussed more explicitly.

1 comment

Thank you for suggesting a direct fit to the measured probes. We ran this control and clarify the Table 4 cost comparison and the connection to timestep scheduling below.

1. Quality and calibration cost in Table 4

The searched exponential and power-law curves have lower mean LPIPS at k14/k18; EXP improves over measured by 9.0%/9.4%. The incremental EXP/measured generation-count ratios, 8.31x/3.80x, apply to that LPIPS slope search. Fitting directly to the probes shares the propagation measurements used by measured interpolation.

PropCache's contribution is the propagation-aware reuse price and segment objective. Endpoint gain weights the local reuse defect and sampler step size, so defects with different downstream effects receive different prices. Segment aggregation and exact-budget minimax DP then determine refresh locations. Table 4's A=1 ablation supports propagation weighting within the tested pooled protocol.

2. Direct fitting from the same measured probes

We fitted log(gain)=b0+b1*step by ordinary least squares (OLS), then compiled schedules with the same DP. OLS determines the EXP slope from probes; calibration LPIPS selects the pricing resolution and exponent. In the FLUX pooled control, both families share defects, pricing resolution, solver, exponent grid, and selection rule; each family selects its own exponent.

For 16 historical FLUX test prompts and two seeds, define Delta = LPIPS(EXP) - LPIPS(measured); negative values favor EXP:

BudgetMeasured / EXP LPIPSDelta
k120.129200 / 0.130213+0.001013
k140.102347 / 0.093178-0.009169
k180.052103 / 0.053355+0.001252
k220.029136 / 0.027141-0.001994

The primary comparison, weighted equally across four budgets, favors EXP: -0.002225, 95% CI [-0.004161, -0.000295]. Bootstrap resampling keeps both seeds and all budgets together within each prompt. Measured has lower mean LPIPS at k12/k18; k18's marginal CI is [+0.000217, +0.002523]. In a separate control that selects pricing resolution and p on calibration data within each family, measured achieves 3.52% lower LPIPS at k18; k14/k22 select identical masks.

Video-model outcomes also vary:

  • HYV1.0, k16/18/25: measured/EXP means are 0.090542/0.096496, 0.068166/0.071492, and 0.024238/0.025424. These ten-prompt results are descriptive; evaluation prompts contributed to gain pooling.
  • HYV1.5, k14/20/25: EXP reduces LPIPS by 3.04%/11.17%/11.59% on 13 calibration-disjoint prompts, comparing calibration-selected pooling/exponent configurations.
  • Wan2.1, k14/20: the same-input control favors EXP by 7.27%/3.61%, with 14/16 and 15/16 wins. Two-budget Bonferroni-adjusted 97.5% intervals for Delta are [-0.011653, -0.005540]/[-0.003505, -0.000368].
  • Wan2.2, k18/24: measured has lower means, with Delta +0.005911/+0.000440 and 95% CIs [-0.001498, +0.012595]/[-0.002300, +0.003196]. The 16 units include unseen prompts and calibration-seen prompts with fresh seeds; both families share gain/defect inputs and canonical pricing.

The preferred gain representation varies by model and budget within the same propagation-pricing framework. Both representations share probe acquisition; candidate generation and selection incur separate costs. We will distinguish OLS fitting from the original slope search in the revision.

3. Connection to timestep scheduling

For FLUX with Euler updates and a held final velocity, each reuse interval satisfies x(b)=x(a)+(σ(b)−σ(a))v(a)x(b)=x(a)+(\sigma(b)-\sigma(a))v(a) in real arithmetic. The computed steps therefore form a grid-aligned coarse Euler schedule. The propagation price supplies an anchor-dependent cost for selecting those steps. GITS also optimizes interval-based trajectory costs, while AYS/OYS optimize sampling grids. We will make these connections explicit in the revision.

The audited Wan runtime reuses block-stack residuals added to the current input, allowing the final velocity to vary at cached steps. Its operator and feasible schedules differ from pure final-velocity holding.

We also tested a shared-schedule variant of OSS (Optimal Stepsize for Diffusion Sampling), selected using only 12 calibration prompt-seed pairs and evaluated on 176 pairs at matched NFE. OSS-minus-PC LPIPS is +0.000654 [-0.000187, +0.001524] at k14 and -0.001717 [-0.002873, -0.000730] at k18. OSS has lower LPIPS at k18; its ImageReward intervals include zero at both budgets. Direct comparisons with GITS/AYS/OYS remain pending.

Complete calibration timing and the extended continuous-cost control remain follow-up items. We will upload the updated code and experiment records and provide an anonymous access link.

Reviewer_uSQiOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PropCache, a step-level caching method to reduce the inference time required to generate images and videos using DiTs. It estimates how caching errors propagate to the final result using calibration prompts, replacing one network forward at a time with reuse of a previous output. The measured propagation gain, local error, and sampler step size form a reuse-cost table, which a dynamic program uses to compile exact-budget schedules. Calibration is offline and training-free, producing shared schedules for different compute budgets without runtime search.

Evaluation spans 6 backbones and includes metrics for fidelity to full-compute outputs, image-text alignment, predicted human preference, and video quality. The authors report improved fidelity at matched compute budgets and up to 4.83x end-to-end speedup over full-compute generation.

Critical strengths

Strong speedups are demonstrated across six backbones. The quantitative results look SOTA at several tested operating points, although corrected MagCache results outperform PropCache on HunyuanVideo-1.5 at k=20 and k=25 (Table 10).

The algorithm is interesting and supported by substantial empirical evidence -- removing propagation weighting worsens held-out fidelity (Table 4), and the joint pricing/exponent ablation explains an important design choice (Appendix J.7). Probe-density and calibration-pool studies further support the approach. The released code is easy to set up and start generating videos (had an agent do this, I just verified its steps).

Critical weaknesses

From viewing the videos generated myself, there is a very very strong flicker for moving objects, such as a person's skin when moving to place an object on a table. As such, I find quality validation to remain incomplete. LPIPS-based calibration and dense-reference LPIPS/PSNR/SSIM establish fidelity rather than sufficient perceptual or temporal quality. Table 8 adds useful VBench evidence, but its custom 88-prompt x two-seed protocol seems non-standard, esp. the dimension-specific prompt coverage and static filtering for flickering. The additional 200-pair audit in Appendix E is ambiguous. Broader preference and motion-quality evidence would strengthen the comparisons in Table 13. In particular, I request ImageReward and/or HPSv3, esp for the FLUX evaluation.

Organization and missing protocol details make the design choices difficult to assess and reproduce. Exponent selection is spread across Section 4 and Appendices B6, G, and J.3-J.7; Tables 16–18 interrupt the derivations on page 25-27. Exact probe/interpolation settings, affine surrogate fitting, and compulsory step handling are insufficiently specified. The exponent grid’s lower bound/density have limited justification, calibration costs are incremental rather than complete, and the PDF omits hardware and timing-protocol details.

Additionally, this work is missing ECAD comparison and related-work discussion (ICLR 2026, https://proceedings.iclr.cc/paper_files/paper/2026/hash/e68274fc4f158dbcbd4dddc672f7ee9c-Abstract-Conference.html)

ECAD already optimizes model-specific caching schedules using a small calibration prompt set and evaluates FLUX.1-dev. The authors should discuss its relationship to PropCache and compare on FLUX.1-dev at matched latency or compute, reporting consistent quality metrics and calibration costs, though obv not on video as that paper did not seem to show video. This comparison would clarify PropCache’s added value because both methods address budget-constrained caching schedule optimization and the resulting quality–compute frontier.

More broadly, a controlled FLUX.1-dev comparison against one or two representative recent methods, such as HyCa, SpeCa, or ClusCa, would help establish PropCache’s practical tradeoff beyond the evaluated scheduling baselines. Comparisons should use matched end-to-end TMACs and consistent quality metrics, accounting for the methods' different caching granularities.

Questions for authors
  • Please specify Table 8’s prompt selection, per-dimension coverage, VBench version/mode, static filtering, and retained sample counts, and clarify how it relates to the standard VBench protocol. Please provide standard-protocol quality results for the compared video arms, or explain the scope supported by the custom evaluation. Also identify the backbone, arms, and cohort of Appendix E’s 200-pair audit and its relationship to Table 8.
  • Please consolidate the deployed settings and their selection rationale: probe nodes/interpolation, affine-surrogate coefficients and fitting, mandatory computed steps, and the exponent grid. Please distinguish complete calibration acquisition cost from the reported incremental cost, and specify hardware, precision, software versions, and timing procedure.
  • Please report ImageReward or HPSv3 for FLUX.1-dev, including the dense baseline, to assess preference beyond agreement with dense outputs.
  • Please discuss ECAD in related work and distinguish its approach from PropCache. A matched-budget FLUX.1-dev comparison would be helpful if feasible; related-work discussion is the minimum requested clarification.
  • How does PropCache compare with recent feature-forecasting and token-clustering approaches such as HyCa, SpeCa, and ClusCa? Please discuss their relationship and provide a controlled FLUX.1-dev quality–latency comparison against representative methods, to establish whether PropCache’s practical advantage extends beyond the evaluated scheduling baselines.

HyCa: Let Features Decide Their Own Solvers: Hybrid Feature Caching for Diffusion Transformers (https://arxiv.org/abs/2510.04188)

SpeCa : Accelerating Diffusion Transformers with Speculative Feature Caching (https://arxiv.org/abs/2509.11628)

ClusCa: Compute Only 16 Tokens in One Timestep: Accelerating Diffusion Transformers with Cluster-Driven Feature Caching(https://arxiv.org/abs/2509.10312)

ECAD: Evolutionary Caching to Accelerate Your Off-the-Shelf Diffusion Model (https://arxiv.org/abs/2506.15682)

If my weakness concerns / questions above are responded to, I will consider raising my score appropriately. I do not expect the authors to materially reorganize the supplement fully during the rebuttal period generally, as that would be excessively burdensome, so I leave that as a strong suggestion for camera-ready / future manuscripts. However, given that much of this manuscript appears to be written with the support of AI, it would greatly benefit readers and other reviewers, should the appendix be properly formatted to be easy to follow and justify decisions made in the paper.

I would very much like to see the additional human preference standins from above, as well as comparisons to other important caching methods left out as also mentioned above.

3 comments

Thank you for inspecting the generated videos and identifying the missing comparisons. We added motion diagnostics, ImageReward scores, and ECAD and ClusCa comparisons. Reported CIs are paired 95% intervals clustered by prompt.

1. Moving-object flicker

We generated 72 Wan2.2 videos using eight action and four static prompts, two seeds, and dense/PropCache (PC)/LeMiCa-style arms. On the action prompts, residual discrepancy in flow-defined moving regions (lower is better) is 4.492734/4.680543 for PC/LeMiCa-style: difference -0.187809 [-0.459526, +0.099044]. The interval includes zero. Please share the backbone, prompt, and seed of the moving-skin example so we can inspect that case directly.

On four official static prompts, VBench 0.1.5 temporal_flickering gives dense/PC/LeMiCa-style scores of 0.982552/0.980018/0.978578. PC improves over LeMiCa-style by +0.001440 [+0.000855, +0.002149] and trails dense by -0.002534 [-0.004534, -0.000556]. Static flickering and moving-region stability are evaluated separately.

2. ImageReward, including dense

We scored all 880 original FLUX samples: 88 prompts, two seeds, five arms. Mean ImageReward is dense 1.001891, PC14 0.960469, PC18 0.989674, LeMiCa14 0.975048, and BudCache14 0.975312. PC-minus-dense is -0.041423 [-0.070806, -0.018284] at k14 and -0.012217 [-0.029131, +0.009745] at k18. PC14-minus-LeMiCa14/BudCache14 is -0.014579/-0.014844, with both intervals including zero. These scores add a learned preference measure to the fidelity evaluation and show a loss relative to dense at k14.

3. Original Table 8 and Appendix E

The archived Table 8 specification records VBench 0.1.5. The evaluation uses Wan2.2 dense40/PC18/LeMiCa-style18, 84 VBench plus four adaptation prompts, two seeds, and all 176 clips per arm in each of the six dimensions. It uses custom_input mode, with no separate static filtering. The new static test applies the official metric to a fixed four-prompt subset.

Original Appendix E concerns HunyuanVideo-13B dense/PC18/LeMiCa18: 100 prompts x two seeds, 200 pairs/600 videos, with 176 unseen pairs as the primary cohort.

4. ECAD and representative recent methods

ECAD searches layer-level caching schedules; PropCache compiles whole-step schedules from propagation-aware segment costs. Under ECAD's released 20-step/1024-pixel/guidance-5 protocol, we recalibrated PC on six prompts/two seeds and tested on 12 disjoint prompts/two seeds. PC11/ECAD slow use 409.116/402.521 T Transformer MACs (+1.64%). PC-minus-ECAD LPIPS is -0.350953 [-0.414809, -0.291364]; ImageReward is +0.367753 [+0.100775, +0.651136]. The 20-step surrogate-consistency diagnostic exceeds its thresholds: median/p90 0.312540/0.477350 versus 0.20/0.35.

HyCa/SpeCa use feature prediction/caching; ClusCa uses token clustering. We tested ClusCa on 88 prompts/two seeds:

PC / ClusCaTransformer MAC fraction, PC / ClusCaOwn-dense LPIPS, PC / ClusCa
k12 / c2048,t50.429 / 0.4270.110918 / 0.350124
k14 / c3072,t50.500 / 0.4970.082608 / 0.360372
k18 / c4096,t30.643 / 0.6330.038326 / 0.217445

Each method uses its own dense reference. All paired LPIPS intervals favor PC. ImageReward differences are +0.018412/+0.009087/-0.028879, with all intervals including zero. MAC fractions are relative to dense Transformer compute; PC counts full forwards, while ClusCa traces cached paths. End-to-end TMAC/latency accounting remains pending.

5. Settings, cost, and appendix organization

The implementation stores squared prices, so its exponent is p=r/2p=r/2, with grid {1,1.1,...,2,2.5,3,4,6,infinity}. Ten audited tiers select p in [1,2.5]. Calibration LPIPS jointly selects pricing resolution and p. Original FLUX k12/14/18/22 use pooled/p1.1, per-unit/p2, per-unit/p1.1, and per-unit/p1. Zero-based probe nodes: FLUX [1,3,5,8,11,14,17,20,22,24,26,27]; Wan2.1 [1,4,8,13,19,25,31,38,43,46,48,49]; Wan2.2 [1,3,6,10,15,20,25,30,34,37,38,39]. Wan2.2 canonical pricing uses log-linear interpolation over step index; the full layout requires step 0, and the tested expert-window configuration also computes step 17. I1 uses velocity defects; I2 uses an affine latent-displacement surrogate.

For the 20-step ECAD control, recorded acquisition includes 12 captures (30.80 s) and 240 probes (562.27 s). Complete historical acquisition/selection/scoring/setup costs, original I2 fit coefficients, and hardware/timing records remain to be consolidated; ECAD's original search cost is unavailable.

We have reorganized the appendix to bring protocol details together and move result and provenance tables out of the derivations. We will upload the revised PDF, updated code, and playable video comparisons and provide an anonymous access link.

Reviewer_uSQiOpenReview ↗

Reply to Author's Rebuttal (1/2)

Thank you for your reply.

My remaining concerns are below:

Flickering

As requested, here is the prompt and seed used to generate the video which shows the hand flickering and deforming considerably:

  • Model: Wan-AI/Wan2.1-T2V-1.3B-Diffusers (revision 0fad780a534b6463e45facd96134c9f345acfa5b), bfloat16
  • Prompt: "The 3D house model is being constructed from particles." (t2vcb_011, the first fixed holdout in your released data/wan2.1-1.3B/inputs/h1_manifest_fps2.json), seed 30087855, using the released official negative prompt
  • Sampling: 50 steps, UniPCMultistepScheduler (flow_shift 3.0), guidance 5.0, 81 frames at 832x480, max_sequence_length 512
  • PropCache k14: your released data/wan2.1-1.3B/masks/masks_k14.json:rhoDP_r25, computed steps [0,1,2,3,4,5,6,8,10,13,18,26,38,47], run with the unmodified released scripts/wan2.1-14B/runtime/lib_wan21.py (Fn0/Bn0 residual replay)
  • LeMiCa k14: your released data/inputs/wan21_lemica_official_masks.json:14, computed steps [0,1,2,3,4,5,7,10,17,24,31,38,45,49], same runtime
  • Dense: official WanPipeline, same settings
  • Our LPIPS vs. dense: PropCache 0.124, LeMiCa 0.144. The hand and arm flicker and deform in the PropCache output relative to dense.

I am also not convinced that PC has a meaningful improvement over LeMiCa with respect to flickering, as its improvements are in the third decimal place over only 4 prompts, and on those prompts PC still trails dense. The moving-region test reports no comparison to dense.

Human Preference Proxy

PC14 is significantly worse than dense on ImageReward: -0.041, CI [-0.071, -0.018]. This, combined with the qualitative samples, leads me to hypothesize that PC overfits to its calibration metric by making changes that impact LPIPS less than they impact visual quality. PC14 doesn't beat LeMiCa, BudCache or ClusCa (every CI includes zero), and in Table 3, PC's HPSv2.1 is slightly lower than LeMiCa's. PC does win on LPIPS, but this is its calibration metric, not a preference proxy.

VBench

Please use the official VBench protocol with the VBench-provided prompts and 5 samples per prompt (short or extended prompts are fine, as long as they are applied consistently), and report the official per-dimension scores. Were the 4 "adaptation" prompts you listed seen during calibration, or are they fully held out?

ECAD and ClusCa Comparison

I believe there are mistakes or configuration issues in these baselines.

  • The ECAD slow configuration in the authors' response is listed at 402.5 TMACs, while ECAD's Table 4 (1024px) and its released slow schedule show 644 TMACs, with fast at 377 and dense at 1190. Please clarify.

  • Given that ECAD slow has a -0.09 ImageReward change relative to dense on held-out PartiPrompts, I find the 0.37 gap between PC and ECAD slow to be implausible. Further, since ECAD slow is -0.09 vs. dense, this gap implies PC is about +0.28 above dense. However, PC is below dense on FLUX elsewhere (-0.041 at k14).

  • Which 12 disjoint prompts were used? Please use a standard prompt set for fairness. Please do the same for ClusCa.

  • I am unable to understand this sentence in the reply: "The 20-step surrogate-consistency diagnostic exceeds its thresholds: median/p90 0.312540/0.477350 versus 0.20/0.35." Could the authors please clarify? Does this mean that PC's own validity check has failed?

  • The ClusCa configuration appears to deviate significantly from the ClusCa paper: the rebuttal uses K=2048-4096, but the paper never goes above 32. The intended way to increase compute is to decrease N rather than increase K. In ClusCa's code, a larger K only adds double-stream MLP tokens. Attention and the 38 single-stream blocks remain forecast, and K values this large leave many clusters empty in the 512-token text stream, which likely explains the large LPIPS of 0.35.

  • Please rerun with a corrected configuration: K=16, Taylor order 2 (the Taylor order used in the rebuttal was not stated), and a sweep over N.

  • Please use standard prompt sets and report which ones are used.

Continued in next comment..

Reviewer_uSQiOpenReview ↗

Reply to Author's Rebuttal (2/2)

Gains

The clear win for PC is against A=1, i.e., no weighting. Simple parametric decaying curves match or beat the measured gain:

  • Searched EXP/power curves have 6-9% lower LPIPS in Table 4.
  • A 1-prompt profile is within the CI of the 12-prompt profile.
  • An OLS EXP fit to the same probes beats the measured gain overall on FLUX, Wan2.1, and HYV1.5, per your reply to Reviewer 2BjM.

There are multiple instances of "pending" in the reply. Can the authors please clarify whether these will be included in the rebuttal or the manuscript? These messages appear to be intended for the authors from an AI agent rather than for reviewers:

  • "End-to-end TMAC/latency accounting remains pending."
  • "Complete historical acquisition/selection/scoring/setup costs, original I2 fit coefficients, and hardware/timing records remain to be consolidated"

Finally, the replies indicate that video links will be provided and that the PDF has been or will be revised, but neither is currently visible on OpenReview.

The rebuttal's new results weaken the paper's central claims. The new evidence suggests that PropCache's gains are specific to LPIPS, its calibration metric, and have not been shown to carry over to preference or temporal quality, and parametric fits match or beat its measured gain. The added baseline comparisons do not yet support a fair evaluation. I am therefore lowering my score to 1.

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

This paper proposes PropCache, which is used to (systematically) decide in which diffusion steps should call DiT outputs (, and which should reuse cached outputs.)

More specifically, this method combines local cache errors, with measured propagation gains. To find the best diffusion steps (to call the DiTs), the authors propose a dynamic programming algorithm minimizing a hand-crafted error function, to produce a schedule, given a computation budget.

Critical strengths
  • Overall I feel the idea is reasonable. I like the approach of accounting for how local caching errors affect the final output. For each model, after the offline DP algorithm, the store of fixed schedule is straightforward, and easy to deploy.

  • Overall, the experiments seem to cover most of the key points.

Critical weaknesses
  • My main concern is regarding how much the difference is comparing PropCache and DPCache in terms of method design. Propagation-aware pricing is a meaningful distinction, but the DP algorithm, offline calibration, and budget-constrained scheduling seems to already have been introduced in DPCache...

  • The calibration setup needs further explanation. As shown in the paper, different T2V models use different prompt sets and calibration sizes, making it difficult to assess robustness to these choices. Why were different prompt pools (for calibration) chosen for the difference T2V models, and how sensitive are the results to those prompts? Results using a shared calibration pool should make the evidence more convincing.

  • Since the main contribution is when to refresh the cache, it would be interesting (and useful) to examine whether the gains still persist when the method is combined with a stronger cache predictor, such as TaylorSeer.

It might also be helpful to include qualitative comparisons at the same speedup in Table 3. Adding PropCache's outputs for the same examples would help show whether the fidelity gains are visually apparent (under the same efficiency setting).

Video comparisons would also help. The sampled frames in Figures 1, 7 and 15 are helpful, but they make it a bit difficult to assess motion quality. I could not find playable videos in the submitted supplementary archive.

Questions for authors

How sensitive are the results to the calibration prompts, and why were different calibration pools chosen for the T2V models?

1 comment

Thank you for engaging closely with our mechanism. We clarify the contribution and provide results for shared calibration and predictor composition.

Difference from DPCache (W1)

Offline calibration and budget-constrained dynamic programming (DP) are shared framework components. In the inspected DPCache implementation, summed path-transition costs use a feature-L1/Taylor proxy over dense histories and a synthetic terminal-feature sentinel.

PropCache contributes an operator-aware propagation price. We inject a bounded reuse defect at one step, compute the remaining trajectory, and measure its endpoint effect. The gain A(t)A(t) captures the downstream amplification of the local defect. Combining it with the anchor-dependent defect d(a,t)d(a,t) and sampler step size gives price=gain×local defect×step size\text{price}=\text{gain}\times\text{local defect}\times\text{step size}. Thus, equally sized local defects can receive different prices depending on their downstream effects. We aggregate these prices within reuse segments and minimize the largest segment cost; the DP solves this discrete proxy exactly at the specified budget.

Table 4 of the submitted paper directly tests propagation weighting: setting A=1A=1 increases held-out LPIPS by 84.0% at k14 and 129.0% at k18. This connects the measured propagation weights to the quality of the resulting schedules. The methodological distinction lies in the measurement target, reuse-price construction, and segment objective.

Shared calibration pool (W2)

The original calibration pools were constructed separately during each model's evaluation. To test a common pool, we re-priced and re-selected Wan2.1 and HunyuanVideo-1.0 schedules using the same four calibration prompt texts and seeds, then evaluated them on three separate held-out prompts. Each backbone used its own gain and defect measurements; schedule selection used calibration outputs.

Mean held-out LPIPS against dense references is:

Modelk16k18k25
Wan2.10.0915000.0701550.035800
HunyuanVideo-1.00.0765930.0592650.024137

Both backbones show decreasing dense-reference error as the compute budget increases under this common four-prompt calibration pool. This small-scale shared-pool control complements the paper's pool-size ablations and directly addresses the choice of calibration prompts across models.

Composition with Taylor prediction (W3)

We evaluated 120 Wan2.2 generations: 12 prompts, two seeds, and five arms comprising dense generation and PC/LeMiCa-style schedules under HOLD and order-1 residual Taylor prediction. The HOLD-calibrated masks remained frozen. Both schedules used the same predictor settings and executed-block budget, k18. Predictor histories follow each mask, with order-zero startup fallback.

The table reports PC-minus-LeMiCa-style LPIPS; negative values favor PC. Confidence intervals use paired prompt-cluster resampling, keeping both seeds together.

PredictorLPIPS difference95% CI
HOLD-0.007976[-0.014525, -0.000941]
Order-1 residual Taylor-0.010661[-0.018192, -0.002704]

PC has lower LPIPS under both predictors, with both intervals below zero. The frozen-schedule comparison shows that its refresh-placement advantage persists when cached residuals use the tested Taylor predictor. This directly addresses whether the scheduling gain carries over to a stronger cache predictor.

Qualitative comparisons and videos (W4/W5)

We will provide an anonymous access link to playable videos and the corresponding qualitative comparisons.

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.