acceptodds
Under review as a conference paper at ICLR 2027

Accelerating Video World Models at the Edge via Defect-Adaptive Reanchoring

Abstract

Video world models require many expensive neural evaluations along their sampling trajectories. Caching skips some of them, but each approximate output alters the subsequent trajectory, and its error is unobservable precisely when the evaluation is skipped. We introduce the Defect-Adaptive Reanchoring Controller (DARC), a training-free method that uses every fresh evaluation twice: as an anchor for predicting complete model outputs, and as feedback on the forecast it replaces. DARC predicts outputs from recent anchors using latent-state and timestep geometry, calibrates this predictor from pre-call forecast errors, and reanchors once accumulated exposure indicates that further prediction is unreliable. It adds no neural evaluation on skipped steps and leaves the native schedule and solver unchanged. Its routing signal tracks the true reuse error with Spearman , versus at most for prior caches. Across four world-model families and seven benchmarks, DARC is faster and more energy-efficient than prior caches on every family, reaching 2.4-3.0 end-to-end speedup with 59-70% lower energy on the edge-class Jetson AGX Thor and up to 3.8 across five GPU platforms, with near-reference quality. Our code and models will be publicly released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video world models, Diffusion model acceleration, Training-free caching, Adaptive computation, Efficient inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Z25mOpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DARC, a training-free framework for accelerating world models (video generation) through steps caching. DARC combines state-conditioned output extrapolation, online calibration, and an accumulated reanchoring strategy. Experiments across four world-model families demonstrate substantial inference speedups while maintaining competitive generation quality. Speedup is measured on five GPU platforms, further enhancing their result.

Critical strengths
  1. Its observation that caching errors affect subsequent generation states, making both prediction accuracy and refresh timing important, is well motivated.

  2. The paper includes component-wise ablations for baseline comparisons, sensitivity studies, and energy measurements.

Critical weaknesses
  1. The speed-quality advantage is less conclusive under matched computational budgets. More experiments are needed to justify this.

  2. DARC uses benchmark-specific values for hyperparameters, whereas competing methods are primarily evaluated using their default parameters. This might lead to unfair comparisons with baselines.

Questions for authors

Please refer to the weakness section.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Z25m (1/2)


We thank Reviewer Z25m for the review. We are glad the reviewer found our core motivation convincing, and we appreciate the recognition of our component-wise ablations, sensitivity studies and energy measurements, and of evaluating speed across five GPU platforms. The reviewer's two concerns, about matched compute and about how hyperparameters were set, are addressed below with additional experiments.


W1. Speed-quality advantage under matched computational budgets.


We thank the reviewer for this comment and agree that a comparison at matched speed is essential. In the paper, the speed-matched analysis of Appendix D was an ablation on a small subset of SANA-Bench. To address this concern thoroughly, we re-ran all baselines and DARC under a common protocol on the full benchmarks of all four model families (Table R1). For each benchmark, every baseline, including the recent RACER, is tuned to run at DARC's speed: its speed knob is set on calibration samples disjoint from the evaluation set, by measured speed alone and never by quality. Quality is measured as paired PSNR, SSIM and LPIPS against the unaccelerated model's output for the same sample and seed, which isolates the error introduced by caching.

Under this protocol, DARC achieves the best PSNR, SSIM and LPIPS on all seven benchmarks. Since every method runs at the same speed, these gains come from how DARC chooses which calls to skip and how it predicts their outputs, not from a different operating point. Two further experiments, reported in our responses to the other reviewers, support this. At WorldDynCache's operating point, with the same share of model calls and the same speedup, DARC is closer to the unaccelerated model on Aether-5B (response to Reviewer BmSZ, W6, Table R5). In long autoregressive rollouts of up to 1,021 frames, it remains the closest to the unaccelerated model at every stage of the video (response to Reviewer CEXN, Q2 and Q6, Tables R4 and R5). We have added Table R1 to the revised paper and based the claims of Sec. 4.1 on this matched-speed comparison.


Table R1. Comparison at matched speed, laid out as Table 1. Every baseline is tuned separately for each benchmark to run at DARC's speed (target next to each benchmark name). Its settings are chosen by measured speed alone, on calibration samples disjoint from the evaluation set (revised Appendix B). Unlike Table 1, which compares PSNR/SSIM/LPIPS against the real video, this table compares them against the unaccelerated model's video for the same sample and seed, which isolates the error introduced by the cache. Bold: best; italics: second best among accelerated methods.

Model / methodPSNR ↑ / SSIM ↑ / LPIPS ↓PSNR ↑ / SSIM ↑ / LPIPS ↓
SANA-WMVBench-I2V (target 3.35×)SANA-Bench (target 2.28×)
+ EasyCache26.09 / 0.851 / 0.06920.10 / 0.623 / 0.182
+ WorldCache20.54 / 0.666 / 0.15317.79 / 0.473 / 0.258
+ FlowCache19.19 / 0.607 / 0.26615.42 / 0.294 / 0.393
+ NaviCache27.15 / 0.876 / 0.06020.79 / 0.699 / 0.159
+ RACER27.40 / 0.844 / 0.06920.60 / 0.663 / 0.174
+ DARC (Ours)28.52 / 0.885 / 0.05721.88 / 0.710 / 0.139
Cosmos-Predict2.5-2BVBench (T2V) (target 2.78×)PAI-Bench-G (I2W) (target 2.38×)
+ EasyCache17.62 / 0.559 / 0.36118.03 / 0.585 / 0.286
+ WorldCache14.35 / 0.457 / 0.52716.02 / 0.528 / 0.380
+ FlowCache14.41 / 0.315 / 0.67514.93 / 0.389 / 0.567
+ NaviCache12.90 / 0.263 / 0.64511.93 / 0.270 / 0.654
+ RACER18.11 / 0.601 / 0.30018.88 / 0.621 / 0.237
+ DARC (Ours)18.98 / 0.631 / 0.25419.58 / 0.661 / 0.207
MiniWorld-1BDROID (target 2.26×)RealEstate10K (target 2.35×)
+ EasyCache33.35 / 0.964 / 0.011523.68 / 0.800 / 0.132
+ WorldCache38.69 / 0.985 / 0.004929.91 / 0.896 / 0.071
+ FlowCache28.74 / 0.922 / 0.027722.83 / 0.817 / 0.149
+ NaviCache33.25 / 0.963 / 0.011924.62 / 0.824 / 0.117
+ RACER38.14 / 0.983 / 0.005029.70 / 0.909 / 0.066
+ DARC (Ours)39.99 / 0.991 / 0.003931.33 / 0.920 / 0.057
AstraSekai (target 2.47×)
+ EasyCache28.80 / 0.833 / 0.150
+ WorldCache26.97 / 0.785 / 0.215
+ FlowCache19.50 / 0.503 / 0.545
+ NaviCache30.05 / 0.854 / 0.171
+ RACER31.02 / 0.864 / 0.165
+ DARC (Ours)32.35 / 0.885 / 0.109

AuthorsOpenReview ↗

Response to Reviewer Z25m (2/2)


W2. Benchmark-specific hyperparameters.


We thank the reviewer for raising this point and would like to separate the two settings clearly. In the submitted Tables 1 and 2, DARC uses the budget ε and span cap m of Table 4, chosen per setting from small sweeps, while all other constants are shared. For all rebuttal experiments, we fixed one configuration for every model and benchmark (span cap m = 10, unit weights, qmax⁡q_{\max} = 1) and set only ε per benchmark, so that DARC runs at its Table 1 speed. Every baseline in Table R1 receives the same treatment: its threshold is set per benchmark to reach that speed, and where the threshold alone could not reach the target speed, one secondary knob was also changed (revised Appendix B). DARC therefore needs no more per-benchmark tuning than any baseline. Under this configuration, DARC remains the closest to the unaccelerated model on all seven benchmarks (Table R1). Table R2 further shows that the span cap and qmax⁡q_{\max} need no tuning, and that the weights only shift the operating point in the same direction as ε. The per-benchmark choice therefore reduces to a single number that sets the speed target, as a threshold does for any cache.


Table R2. Robustness of DARC to its hyperparameters, one parameter changed at a time. All rebuttal experiments use a single configuration shared across model families (m = 10, all weights 1, qmax⁡q_{\max} = 1); only the budget ε differs between benchmarks, and it plays the same role as the threshold of a baseline cache by selecting the operating point on the speed-quality trade-off. Changing the span cap or the weights over a wide range shifts the operating point smoothly and never causes a failure, so DARC's results do not depend on fine-tuning these parameters. ε is given as a multiple of each model's default. Calls: share of native model calls. Speed: relative to the unaccelerated model. PSNR: paired to the unaccelerated model (dB). Each cell: Calls ↓ / Speed ↑ / PSNR ↑.

ParameterValueSANA-WM (VBench-I2V)MiniWorld (DROID)Astra (Sekai)
ε×1/868% / 1.48× / 36.3165% / 1.54× / 44.1874% / 1.31× / 42.75
×1/447% / 2.03× / 34.6562% / 1.63× / 43.0357% / 1.60× / 35.38
×1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
×416% / 4.33× / 27.0737% / 2.47× / 39.1719% / 3.46× / 28.35
×1615% / 4.39× / 26.2135% / 2.56× / 38.6916% / 3.85× / 26.54
m339% / 2.31× / 30.4560% / 1.64× / 42.8341% / 2.09× / 35.33
528% / 2.95× / 30.0743% / 2.25× / 39.3734% / 2.42× / 34.13
10 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
2024% / 3.32× / 29.7543% / 2.25× / 39.3733% / 2.43× / 33.97
none24% / 3.39× / 29.7543% / 2.25× / 39.3733% / 2.42× / 33.97
wβqw_{\beta q}024% / 3.44× / 29.7343% / 2.25× / 39.3729% / 2.62× / 33.16
0.2524% / 3.42× / 29.7343% / 2.25× / 39.3729% / 2.66× / 33.15
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
424% / 3.30× / 29.7344% / 2.23× / 39.4448% / 1.84× / 34.87
1624% / 3.42× / 30.6157% / 1.76× / 42.5760% / 1.56× / 35.12
wchw_{\text{ch}}021% / 3.52× / 28.9542% / 2.28× / 39.4932% / 2.53× / 32.39
0.2522% / 3.47× / 28.9443% / 2.27× / 39.5133% / 2.48× / 31.89
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
430% / 2.86× / 31.7651% / 1.93× / 40.5937% / 2.24× / 33.99
1646% / 2.12× / 34.8762% / 1.62× / 43.5247% / 1.87× / 35.10
wqw_q023% / 3.38× / 29.6943% / 2.26× / 39.4232% / 2.48× / 33.46
0.2523% / 3.39× / 29.7043% / 2.26× / 39.4233% / 2.47× / 33.52
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
425% / 3.27× / 30.7944% / 2.22× / 39.5636% / 2.29× / 34.59
1629% / 2.97× / 30.3853% / 1.88× / 40.4743% / 2.03× / 34.65
qmax⁡q_{\max}0.2524% / 3.36× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
0.524% / 3.35× / 29.7343% / 2.25× / 39.3733% / 2.42× / 33.97
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
none24% / 3.48× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97

Dear Reviewer,

We really thank you again for your detailed review of our paper. We know this is a busy period. We have posted a detailed response addressing each of your comments and the requested experiments, and we would be grateful if you could please take a look when convenient. If you feel your concerns have been addressed, we would sincerely appreciate your considering an updated evaluation. If anything remains unconvincing, we would be glad to discuss it together. We are following this thread closely and will respond promptly. We would greatly welcome any additional feedback you might have.

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

This paper introduces the Defect-Adaptive Reanchoring Controller (DARC), a training-free method to accelerate video world model inference by intelligently caching and predicting model outputs. The key insight is that each fresh neural evaluation can serve dual purposes: as an anchor for predicting future outputs, and as feedback on the forecast it replaces. DARC retains the two most recent fresh outputs and compact signatures of their input states, then uses the current state's displacement relative to these anchors (combined with timestep progress) to extrapolate the change between stored outputs via a secant prediction. A hold-secant blend, calibrated online from measured pre-call forecast errors, moderates the extrapolation. The controller tracks an exposure score that accumulates within the current segment, combining the last measured forecast defect with current state-time geometry, and requests a fresh evaluation when this score reaches a budget. DARC adds no neural evaluation on skipped steps and leaves the native schedule and solver unchanged. Its routing signal tracks the true reuse error with Spearman rho = 0.95, versus at most 0.50 for prior caches. Across four world-model families (SANA-WM, Cosmos, MiniWorld, Astra) and seven benchmarks, DARC achieves 2.4-3.0x end-to-end speedup with 59-70% lower energy on Jetson AGX Thor, and up to 3.8x across five GPU platforms, with near-reference quality.

Critical strengths

The dual-use of fresh evaluations — as both prediction anchors and forecast feedback — is a clever and well-motivated design that addresses a fundamental information asymmetry in caching: the error of a skipped step is unobservable exactly when the evaluation is skipped. The defect-exposure control loop (Section 3.4) that combines lagged measured errors with current state geometry is a principled solution to this problem, and the Spearman rho = 0.95 result demonstrates that the routing signal genuinely tracks true reuse error — a significant improvement over prior methods (rho <= 0.50). The training-free nature of DARC is a major practical advantage: it preserves the pretrained checkpoint, native timestep grid, and solver, making it a drop-in accelerator. The evaluation is comprehensive: four world-model families covering camera-controlled, robot-action-conditioned, and text-conditioned generation; seven benchmarks; five GPU platforms ranging from datacenter (H200) to edge (AGX Thor); and reporting of both latency and energy consumption. The consistent gains across all model families (2.4-3.0x on edge) and the near-reference quality (e.g., PSNR 24.6 vs. 22.2 for WorldCache on SANA-WM) are strong empirical results. The ablation in Figure 2 showing that "when reuse happens matters as much as how often" is an insightful finding.

Critical weaknesses

(1) The method has numerous hyperparameters (segment budget epsilon, max anchor span m, controller weights w_beta_q, w_ch, w_q, q_max, T_norm) whose sensitivity is not systematically analyzed. While the paper reports results across model families, it does not study how performance varies with these parameters, whether they need per-model tuning, or what happens at extreme settings. For a training-free method intended as a drop-in accelerator, hyperparameter robustness is critical for adoption. (2) The exposure score (Eq. 4) is described as "an empirical criterion for reanchoring" that "does not provide a certified trajectory-error bound." This means there is no theoretical guarantee that DARC's predictions stay within an acceptable error bound over long generation horizons. The paper does not discuss failure modes — what happens when the predictor diverges significantly from the true output, or when the scene undergoes a rapid transition that the secant extrapolation cannot capture. (3) The secant prediction uses only the two most recent anchors. For complex scenes with non-monotonic dynamics (e.g., camera revisits, scene cuts), a two-point extrapolation may be insufficient. The paper shows that DARC handles revisits better than WorldCache (Figure 1), but the improvement is attributed to better routing rather than better prediction. Whether a higher-order predictor (e.g., using three or more anchors) would improve quality at the same skip rate is not explored. (4) The quality metrics are somewhat inconsistent across benchmarks. On some (e.g., SANA-Bench, VBench), DARC shows quality improvements over the baseline (+0.7, +0.4), which is surprising for a caching method and suggests the baseline may have its own quality issues. On others (e.g., DROID-100 FVD), DARC shows a degradation (+9.1 FVD). The paper does not explain these inconsistencies or discuss which benchmarks are most sensitive to caching artifacts. (5) While the paper reports energy savings (59-70% on edge), it does not provide a breakdown of where the energy savings come from (reduced GPU compute, reduced memory traffic, reduced cooling). Understanding the energy profile would help practitioners predict gains on other hardware. (6) The comparison with prior caches (EasyCache, WorldCache, FlowCache, NaviCache) uses each method with its default settings. Whether these defaults are optimal for the evaluated models is not discussed. Some prior methods (e.g., WorldCache) have their own tunable parameters, and a tuned baseline could narrow the gap. (7) The paper claims DARC "leaves the native schedule and solver unchanged," but the controller effectively creates a non-uniform evaluation schedule (some steps are fresh, others predicted). This means the effective solver trajectory differs from the reference, and the long-term stability of this modified trajectory over many autoregressive chunks (for AR-based world models) is not analyzed.

Questions for authors

Q1. The method has numerous hyperparameters (segment budget epsilon, max anchor span m, controller weights w_beta_q, w_ch, w_q, q_max, T_norm). Can you provide a systematic sensitivity analysis — how does performance (speedup and quality) vary with each parameter? Are the default values universal across all four model families, or do they require per-model tuning? What happens at extreme settings (e.g., very aggressive or very conservative budgets)? Q2. The exposure score (Eq. 4) does not provide a certified trajectory-error bound. Can you characterize the failure modes: what happens when the secant predictor diverges significantly from the true output? Have you observed catastrophic error accumulation over long generation horizons (e.g., 1000+ steps)? Is there a recovery mechanism beyond reanchoring, such as error detection or fallback to full evaluation? Q3. The secant prediction uses only two anchors. Have you experimented with higher-order predictors (e.g., three-anchor quadratic extrapolation, or learned predictors)? Would a more expressive predictor improve quality at the same skip rate, or is the two-anchor limitation a fundamental information constraint? Q4. On some benchmarks (SANA-Bench, VBench), DARC shows quality improvements over the baseline (+0.7, +0.4), which is surprising for a caching method. Can you explain this? Is the baseline itself suboptimal, or does DARC's non-uniform schedule act as an implicit regularizer? On DROID-100 FVD, DARC degrades (+9.1 FVD) — what is special about this benchmark that makes it sensitive to caching? Q5. The baselines (EasyCache, WorldCache, FlowCache, NaviCache) are used with default settings. Have you tuned these baselines for the evaluated models? Specifically, WorldCache has its own routing parameters — does a tuned WorldCache narrow the gap to DARC? Q6. For AR-based world models that generate many autoregressive chunks, the modified trajectory from non-uniform evaluation could compound errors across chunks. Have you tested DARC on long video generation (e.g., 60+ seconds)? Does the quality degradation grow with the number of AR chunks, and is there a practical limit on generation length? Q7. Can you provide a breakdown of the energy savings (59–70% on edge)? How much comes from reduced GPU compute vs. reduced memory traffic vs. reduced cooling? This would help practitioners predict energy gains on other hardware platforms.

10 comments
AuthorsOpenReview ↗

Response to Reviewer CEXN (1/8)


We thank the Reviewer CEXN for the thoughtful review. We are glad the reviewer found DARC’s central idea, using each fresh evaluation both as a prediction anchor and as feedback on the forecast it replaces, clever and well motivated, as it addresses a basic difficulty of caching, i.e., the error of a skipped step cannot be observed at the moment it is skipped. We appreciate that the reviewer considers the defect-exposure loop principled and the training-free, drop-in design a practical advantage, as it leaves the checkpoint, timestep grid and solver untouched. We are also grateful that the reviewer highlighted the finding that “when reuse happens” matters as much as how often, which shaped much of the method’s design. We answer each of the reviewer’s questions below.

For reference, Table R1 compares all methods at matched speed on the full benchmarks and several of the answers below refer to it.


Table R1. Comparison at matched speed, laid out as Table 1. Every baseline is tuned separately for each benchmark to run at DARC's speed (target next to each benchmark name). Its settings are chosen by measured speed alone, on calibration samples disjoint from the evaluation set (revised Appendix B). Unlike Table 1, which compares PSNR/SSIM/LPIPS against the real video, this table compares them against the unaccelerated model's video for the same sample and seed, which isolates the error introduced by the cache. Bold: best; italics: second best among accelerated methods.

Model / methodPSNR ↑ / SSIM ↑ / LPIPS ↓PSNR ↑ / SSIM ↑ / LPIPS ↓
SANA-WMVBench-I2V (target 3.35×)SANA-Bench (target 2.28×)
+ EasyCache26.09 / 0.851 / 0.06920.10 / 0.623 / 0.182
+ WorldCache20.54 / 0.666 / 0.15317.79 / 0.473 / 0.258
+ FlowCache19.19 / 0.607 / 0.26615.42 / 0.294 / 0.393
+ NaviCache27.15 / 0.876 / 0.06020.79 / 0.699 / 0.159
+ RACER27.40 / 0.844 / 0.06920.60 / 0.663 / 0.174
+ DARC (Ours)28.52 / 0.885 / 0.05721.88 / 0.710 / 0.139
Cosmos-Predict2.5-2BVBench (T2V) (target 2.78×)PAI-Bench-G (I2W) (target 2.38×)
+ EasyCache17.62 / 0.559 / 0.36118.03 / 0.585 / 0.286
+ WorldCache14.35 / 0.457 / 0.52716.02 / 0.528 / 0.380
+ FlowCache14.41 / 0.315 / 0.67514.93 / 0.389 / 0.567
+ NaviCache12.90 / 0.263 / 0.64511.93 / 0.270 / 0.654
+ RACER18.11 / 0.601 / 0.30018.88 / 0.621 / 0.237
+ DARC (Ours)18.98 / 0.631 / 0.25419.58 / 0.661 / 0.207
MiniWorld-1BDROID (target 2.26×)RealEstate10K (target 2.35×)
+ EasyCache33.35 / 0.964 / 0.011523.68 / 0.800 / 0.132
+ WorldCache38.69 / 0.985 / 0.004929.91 / 0.896 / 0.071
+ FlowCache28.74 / 0.922 / 0.027722.83 / 0.817 / 0.149
+ NaviCache33.25 / 0.963 / 0.011924.62 / 0.824 / 0.117
+ RACER38.14 / 0.983 / 0.005029.70 / 0.909 / 0.066
+ DARC (Ours)39.99 / 0.991 / 0.003931.33 / 0.920 / 0.057
AstraSekai (target 2.47×)
+ EasyCache28.80 / 0.833 / 0.150
+ WorldCache26.97 / 0.785 / 0.215
+ FlowCache19.50 / 0.503 / 0.545
+ NaviCache30.05 / 0.854 / 0.171
+ RACER31.02 / 0.864 / 0.165
+ DARC (Ours)32.35 / 0.885 / 0.109

Q1. Hyperparameter sensitivity; universal defaults or per-model tuning?


We thank the reviewer for this question, which is central to using DARC as a drop-in accelerator. To address it directly, all rebuttal experiments use a single configuration shared across the four model families. The only setting chosen per benchmark is the budget ε, which sets the speed-quality operating point in the same way the threshold does for every baseline cache.

To show how DARC responds to each parameter mentioned in the comment, Table R2 varies one parameter at a time across three model families (SANA-WM, MiniWorld, and Astra), including both budget extremes. Four observations follow.

The budget behaves predictably at both extremes. As ε decreases, DARC evaluates more steps and converges to the unaccelerated model. As ε increases, speed rises and then saturates, because the always-fresh steps and the span cap bound how many calls can be skipped. DARC therefore degrades gracefully rather than collapsing, even at a budget sixteen times the default.

The span cap and qmax⁡q_{\max} need no tuning. Loosening either one, or removing it entirely, leaves calls and fidelity essentially unchanged on all three models. Only tighter span caps add calls, and they do so in exchange for higher fidelity.


AuthorsOpenReview ↗

Response to Reviewer CEXN (2/8)


The weights act consistently and safely. Each weight shifts the operating point in the same direction as the budget, as a larger weight never leads to fewer calls, and a smaller one never to more. Across a 64-fold range of values, no setting causes a failure.

TnormT_{\text{norm}} is not a separate parameter. It only rescales the quantity that is compared with ε, so changing it is equivalent to changing ε, which the budget rows already cover.

Together, these results show that a single shared configuration is transferable across model families and that adapting DARC to a new model or benchmark amounts to choosing ε to achieve the desired speed. We added Table R2 and this discussion to the revised Appendix G.


Table R2. Sensitivity of DARC to each hyperparameter, one parameter changed at a time. All rebuttal experiments share one configuration across model families, with the budget ε as the only per-benchmark setting; here, ε is varied from one eighth to sixteen times its default on each model, and every other parameter over a wide range around its default. TnormT_{\text{norm}} only rescales the quantity compared with ε and is therefore covered by the ε rows. Calls: share of native model calls. Speed: relative to the unaccelerated model. PSNR: paired to the unaccelerated model (dB). Each cell: Calls ↓ / Speed ↑ / PSNR ↑.

ParameterValueSANA-WM (VBench-I2V)MiniWorld (DROID)Astra (Sekai)
ε×1/868% / 1.48× / 36.3165% / 1.54× / 44.1874% / 1.31× / 42.75
×1/447% / 2.03× / 34.6562% / 1.63× / 43.0357% / 1.60× / 35.38
×1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
×416% / 4.33× / 27.0737% / 2.47× / 39.1719% / 3.46× / 28.35
×1615% / 4.39× / 26.2135% / 2.56× / 38.6916% / 3.85× / 26.54
m339% / 2.31× / 30.4560% / 1.64× / 42.8341% / 2.09× / 35.33
528% / 2.95× / 30.0743% / 2.25× / 39.3734% / 2.42× / 34.13
10 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
2024% / 3.32× / 29.7543% / 2.25× / 39.3733% / 2.43× / 33.97
none24% / 3.39× / 29.7543% / 2.25× / 39.3733% / 2.42× / 33.97
wβqw_{\beta q}024% / 3.44× / 29.7343% / 2.25× / 39.3729% / 2.62× / 33.16
0.2524% / 3.42× / 29.7343% / 2.25× / 39.3729% / 2.66× / 33.15
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
424% / 3.30× / 29.7344% / 2.23× / 39.4448% / 1.84× / 34.87
1624% / 3.42× / 30.6157% / 1.76× / 42.5760% / 1.56× / 35.12
wchw_{\text{ch}}021% / 3.52× / 28.9542% / 2.28× / 39.4932% / 2.53× / 32.39
0.2522% / 3.47× / 28.9443% / 2.27× / 39.5133% / 2.48× / 31.89
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
430% / 2.86× / 31.7651% / 1.93× / 40.5937% / 2.24× / 33.99
1646% / 2.12× / 34.8762% / 1.62× / 43.5247% / 1.87× / 35.10
wqw_q023% / 3.38× / 29.6943% / 2.26× / 39.4232% / 2.48× / 33.46
0.2523% / 3.39× / 29.7043% / 2.26× / 39.4233% / 2.47× / 33.52
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
425% / 3.27× / 30.7944% / 2.22× / 39.5636% / 2.29× / 34.59
1629% / 2.97× / 30.3853% / 1.88× / 40.4743% / 2.03× / 34.65
qmax⁡q_{\max}0.2524% / 3.36× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
0.524% / 3.35× / 29.7343% / 2.25× / 39.3733% / 2.42× / 33.97
1 (default)24% / 3.30× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97
none24% / 3.48× / 29.7343% / 2.25× / 39.3733% / 2.43× / 33.97

AuthorsOpenReview ↗

Response to Reviewer CEXN (3/8)


Q2. No error bound; failure modes; long horizons.


We thank the reviewer for this question. The exposure score is a control signal rather than a certified bound (Appendix I.1), so DARC relies on three safeguards to limit a wrong forecast. First, the span cap allows at most m − 1 consecutive predicted steps. Second, every fresh call measures the error of the forecast it replaced, and a large error shortens the next span. Third, the first two steps, the first calibration step and the final step are always evaluated fresh, and no state is carried between trajectories.

Fallback to full evaluation. This feedback acts as an automatic fallback: the measured error sets how soon the next fresh call comes. After the largest 10% of measured errors, DARC makes its next fresh call 1.4 steps later on average, so nearly every step is evaluated, whereas after a below-median error it comes 6.0 steps later. Over 2,109 fresh calls on SANA-Bench, a larger error is followed by a shorter span.

When the predictor is wrong. To test this directly, we replaced DARC's forecast with raw extrapolation, without the calibrated blend, and left the controller unchanged (Table R3). Where the raw forecast is less accurate, DARC compensates with more fresh calls (up to 43%, 55% and 41% of the steps on the three models), and no run diverges, although PSNR drops by up to 4 dB on MiniWorld without the calibrated blend.


Table R3. DARC with a weaker forecast. Share of native model calls (↓) / PSNR paired to the unaccelerated model in dB (↑) when the calibrated blend is replaced by raw extrapolation through two or three anchors, with the controller unchanged. Same settings and samples as Table R2.

Forecast (calls ↓ / PSNR in dB ↑)SANA-WMMiniWorldAstra
DARC (Ours)24% / 29.7343% / 39.3733% / 33.97
Raw two-anchor extrapolation31% / 28.9246% / 38.0133% / 33.37
Raw three-anchor extrapolation43% / 31.2955% / 35.3841% / 31.91

No catastrophic accumulation (Table R4). Table R4 compares the deviation from the unaccelerated model at the start and the end of each video, on 60-second SANA-WM rollouts and on autoregressive generation with MiniWorld and Astra. The deviation grows along the rollout for every method, as expected once any cache alters the trajectory, yet DARC has the lowest deviation of all methods at both the start and the end of every video.

Long rollouts (Table R5). To test longer horizons, we extended MiniWorld rollouts to 509 and 1,021 frames, two and four times the length used in the paper, running EasyCache, WorldCache, RACER and DARC. Fidelity decreases smoothly with length, with no sudden breakdown at any point. DARC has the highest PSNR in every stretch of both rollouts, with WorldCache close behind and RACER and EasyCache further back. Within the longest horizon we tested, we therefore observe no catastrophic error accumulation.


Table R4. Error along the rollout. Paired LPIPS to the unaccelerated output (lower is better) at the start and end of each video, each method at a Table R1 setting. Bold: lowest. SANA-Bench: 60 s videos; MiniWorld: 253 frames, autoregressive; Astra: 3 autoregressive windows.

MethodSANA-Bench first 7.5 s ↓SANA-Bench last 7.5 s ↓MiniWorld first 1/8 ↓MiniWorld last 1/8 ↓Astra window 1 ↓Astra window 3 ↓
EasyCache0.1110.1640.00370.01600.0720.223
WorldCache0.1620.2400.00080.00740.0810.260
FlowCache0.2660.3730.00590.07310.3090.668
NaviCache0.0980.1340.00340.01710.0730.240
RACER0.0940.1360.00090.01490.0700.235
DARC (Ours)0.0900.1190.00030.00560.0410.171

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

DARC is a training-free complete-output cache for video world models. It keeps the two most recent fresh outputs as anchors, predicts skipped outputs by a secant extrapolation (placed by a blend of timestep progress and latent-state displacement) damped toward a zero-order hold with an online-calibrated coefficient, and uses each fresh call both as a new anchor and to measure the pre-call forecast error. An accumulated “exposure” score, fed by the last measured defect and current state geometry, triggers the next fresh call at budget ε or span cap m. Solver and schedule are unchanged. Evaluated on 4 world models × 7 benchmarks × 5 GPUs (incl. Jetson Thor): 2.3–3.3× speedup on H200, 59–70% lower energy on Thor, VBench/PAI within 0.7% of exact; appendix adds speed-matched comparisons, routing-signal correlation (ρ=0.95), a cumulative ablation, and a few-step companion (DARC-FS).

Critical strengths
  1. Broad, careful measurement: four model families, seven benchmarks, five hardware platforms, measured energy, bootstrap CIs on SANA-Bench, and a speed-matched analysis.

  2. Candid appendix, including the observation that revisit PSNR can reward loss of fine detail.

Critical weaknesses
  1. The main-text claims are contradicted by the paper’s own speed-matched analysis. Sec. 4.1 states that prior caches approaching DARC’s speed lose more quality, but baselines run at released, conservative thresholds while DARC’s ε and m are tuned per setting. When thresholds are matched to DARC’s ~19 fresh calls (Table 5), paired LPIPS to EXACT is EasyCache 0.241, DARC 0.242, WorldCache 0.243, NaviCache 0.248 — a tie, and Appendix D itself says NaviCache and EasyCache stay closest to EXACT. DARC’s only matched-speed advantage is ΔRevisit PSNR, which Appendix D acknowledges grows with distance from EXACT and may reward blur. The headline advantage is therefore an operating-point difference, not a better routing rule; energy savings equal 1−1/speedup by the authors’ own account and add no independent evidence.
  2. The novel controller components contribute little, and the closest baselines are missing. In Table 3 the predictor (secant + online calibration) accounts for 2.6 dB; the two new components, defect routing and geometry risk, add 0.5 dB each (five scenes). The headline ρ=0.95 is the own-target, 10-prompt estimate; under a common target it falls to 0.61, and Appendix F concedes DARC’s signal is built from the same kind of forecast error it is scored on. WorldDynCache (risk-controlled) and Disagree-to-Accelerate are cited as mechanistically closest but not compared.
Questions for authors
  1. Report Table 1 with all baselines at speed-matched thresholds (paired PSNR/LPIPS to EXACT, not only VBench), and tune baselines per setting as DARC is.
  2. Add WorldDynCache and Disagree-to-Accelerate as baselines.
  3. Explain the OOM entries in Table 2 given Appendix I says peak memory was not measured and offloading configs differ.
8 comments
AuthorsOpenReview ↗

Response to Reviewer BmSZ (1/6)

We thank Reviewer BmSZ for a detailed review of the paper, including the appendix, which made the review especially valuable. We are glad the reviewer found our evaluation broad and carefully executed, covering four model families, seven benchmarks and five hardware platforms, with measured energy, bootstrap confidence intervals and a speed-matched comparison. We also appreciate that the reviewer valued the candor of our appendix, including the observation that revisit PSNR can reward the loss of fine detail. We address each question and comment below:


W1. Main-text claims and the speed-matched analysis.


We thank the reviewer for the comment. The speed-matched comparison in Appendix D (Table 5) was designed as an ablation on the SANA-Bench subset to examine how each method behaves once pushed to a higher speed and was not meant as a full benchmark evaluation. Even there, the matched baselines did not quite reach DARC’s operating point: DARC remained the fastest method in the group while using the fewest fresh calls, and on those ablation samples, the differences in paired LPIPS were too small to reliably separate the methods.

To further address this question and as suggested by the reviewer, we re-ran all baselines and DARC under the same protocol on the full benchmarks of Table 1 (Table R1), adding RACER (Disagree-to-Accelerate) as a further baseline. Every baseline is tuned separately for each benchmark to run at DARC's speed, with its speed knob set on calibration samples disjoint from the evaluation set, by measured speed alone and never by quality. Quality is measured as paired PSNR, SSIM and LPIPS against the unaccelerated model’s output for the same sample, rather than against the real video as in Table 1. This isolates the error introduced by caching, and it does not reward blur, since any loss of detail moves a video away from the reference. Under this protocol, DARC achieves the best PSNR, SSIM and LPIPS on all seven benchmarks and all four model families. Since every method runs at the same speed, these gains come from which calls are skipped and how their outputs are predicted, that is, from the routing and prediction rules themselves rather than from a different operating point.


Table R1. Comparison at matched speed, laid out as Table 1. Every baseline is tuned separately for each benchmark to run at DARC's speed (target next to each benchmark name). Its settings are chosen by measured speed alone, on calibration samples disjoint from the evaluation set (revised Appendix B). Unlike Table 1, which compares PSNR/SSIM/LPIPS against the real video, this table compares them against the unaccelerated model's video for the same sample and seed, which isolates the error introduced by the cache. Bold: best; italics: second best among accelerated methods.

Model / methodPSNR ↑ / SSIM ↑ / LPIPS ↓PSNR ↑ / SSIM ↑ / LPIPS ↓
SANA-WMVBench-I2V (target 3.35×)SANA-Bench (target 2.28×)
+ EasyCache26.09 / 0.851 / 0.06920.10 / 0.623 / 0.182
+ WorldCache20.54 / 0.666 / 0.15317.79 / 0.473 / 0.258
+ FlowCache19.19 / 0.607 / 0.26615.42 / 0.294 / 0.393
+ NaviCache27.15 / 0.876 / 0.06020.79 / 0.699 / 0.159
+ RACER27.40 / 0.844 / 0.06920.60 / 0.663 / 0.174
+ DARC (Ours)28.52 / 0.885 / 0.05721.88 / 0.710 / 0.139
Cosmos-Predict2.5-2BVBench (T2V) (target 2.78×)PAI-Bench-G (I2W) (target 2.38×)
+ EasyCache17.62 / 0.559 / 0.36118.03 / 0.585 / 0.286
+ WorldCache14.35 / 0.457 / 0.52716.02 / 0.528 / 0.380
+ FlowCache14.41 / 0.315 / 0.67514.93 / 0.389 / 0.567
+ NaviCache12.90 / 0.263 / 0.64511.93 / 0.270 / 0.654
+ RACER18.11 / 0.601 / 0.30018.88 / 0.621 / 0.237
+ DARC (Ours)18.98 / 0.631 / 0.25419.58 / 0.661 / 0.207
MiniWorld-1BDROID (target 2.26×)RealEstate10K (target 2.35×)
+ EasyCache33.35 / 0.964 / 0.011523.68 / 0.800 / 0.132
+ WorldCache38.69 / 0.985 / 0.004929.91 / 0.896 / 0.071
+ FlowCache28.74 / 0.922 / 0.027722.83 / 0.817 / 0.149
+ NaviCache33.25 / 0.963 / 0.011924.62 / 0.824 / 0.117
+ RACER38.14 / 0.983 / 0.005029.70 / 0.909 / 0.066
+ DARC (Ours)39.99 / 0.991 / 0.003931.33 / 0.920 / 0.057
AstraSekai (target 2.47×)
+ EasyCache28.80 / 0.833 / 0.150
+ WorldCache26.97 / 0.785 / 0.215
+ FlowCache19.50 / 0.503 / 0.545
+ NaviCache30.05 / 0.854 / 0.171
+ RACER31.02 / 0.864 / 0.165
+ DARC (Ours)32.35 / 0.885 / 0.109

AuthorsOpenReview ↗

Response to Reviewer BmSZ (2/6)


Regarding tuning, setting hyperparameters per model is standard practice for training-free caches in the literature. NaviCache, for example, uses NalignN_{\text{align}} = 5 for HunyuanVideo and 10 for Wan2.1, with error thresholds adapted to each model and mode (0.04-0.07 on Wan2.1, 0.025-0.04 on HunyuanVideo, and 0.15-0.55 on Open-Sora). Furthermore, they also have a different set of parameters for their Fast, Medium, and Slow variants.

We nevertheless agree with the reviewer that the comparison should not depend on how DARC was tuned. In Table R1, every baseline receives the same treatment as DARC: its speed knob is set separately for each benchmark to reach DARC's speed, and where its threshold alone could not reach that speed, one secondary knob was also changed (revised Appendix B). DARC itself sets only its budget ε per benchmark. We have added Table R1 to Appendix D.1 and its summary against the best prior method to Sec. 4.1, revised the claims of Sec. 4.1 so that they rest on this matched-speed comparison, and updated the summary in Appendix D accordingly.

Relation to Table 5 (Table 10 in the revision). Table 5 and Table R1 are different experiments. Table 5 was an ablation that differs from Table R1 in two ways. First, it covers eight SANA-Bench scenes, whereas Table R1 covers all 160 SANA-Bench videos and the full sets of the other six benchmarks. Second, in Table 5 each baseline's threshold was raised until its number of fresh calls approached DARC's, and the baselines remained slower than DARC, whereas in Table R1 every baseline runs at DARC's speed on each benchmark. Absolute values therefore differ between the two tables. We state these differences in Appendix D of the revised paper.


W2. Revisit PSNR at matched speed.


We thank the reviewer for engaging closely with this passage. The remark in Appendix D was a cautious reading of the eight-scene ablation subset: across thresholds, revisit gains rose as methods moved further from Exact, and we noted reduced fine detail as a possible confound, not as a finding about DARC. We have now tested it in two ways.

Fidelity on the full benchmark (Table R1). At matched speed, DARC has the best paired PSNR, SSIM, and LPIPS on SANA-Bench, the benchmark on which Table 1 reports its largest revisit gain, as well as on all other benchmarks. A loss of detail would move DARC away from the unaccelerated model, so its revisit consistency comes together with fidelity to the reference, not at its expense.

Sharpness, measured directly (Table R2). In ablation analysis runs, separate from the evaluation runs of Tables 1 and R1, we measured sharpness as the variance of the Laplacian of luminance, which is a standard focus measure, for each video relative to the unaccelerated model. DARC retains 97.4% of the unaccelerated model's sharpness. To test whether this small difference could explain the revisit gain, we blurred the unaccelerated videos themselves, one by one, to DARC's sharpness and recomputed revisit PSNR on the benchmark's own frame pairs. Blur alone raises revisit PSNR by 0.009 dB, against +0.20 dB for DARC, so it accounts for less than 5% of DARC's revisit gain. DARC's gain exceeds the blur effect in 87% of the videos, and on the 86 videos where DARC keeps at least 98% of the sharpness its gain is still +0.16 dB, against +0.0005 dB from blur. Reduced detail therefore does not explain DARC's revisit gain. We have revised Appendix D accordingly and added Table R2 to it.


Table R2. Does DARC blur? From ablation analysis runs, separate from the evaluation runs of Tables 1 and R1. Sharpness: variance of the Laplacian of luminance (a standard focus measure), per video, relative to the unaccelerated model. Blur control: the unaccelerated output is blurred, video by video, to DARC's sharpness, and revisit PSNR is recomputed on the benchmark's frame pairs. 95% bootstrap intervals over videos.

SANA-Bench
Sharpness kept by DARC97.4% [96.7, 98.1]
Revisit PSNR gain of DARC+0.20 dB [+0.17, +0.24]
Revisit PSNR gain from blur alone+0.009 dB [+0.007, +0.011]
86 videos where DARC keeps ≥ 98% of the sharpness
Revisit PSNR gain of DARC+0.16 dB
Revisit PSNR gain from blur alone+0.0005 dB

AuthorsOpenReview ↗

Response to Reviewer BmSZ (3/6)


W3. Energy versus speedup.


We thank the reviewer for this comment and are happy to clarify. Energy is measured directly, not derived from the speedup: for every run, we sample power (board power at 10 Hz with nvidia-smi, and every module power rail at 1 Hz with tegrastats on the Jetson AGX Thor) and integrate it over the run. The expression 1 − 1/speedup in Appendix A.3 serves only as a reference, namely the saving one would expect if average power stayed constant.

To examine what energy actually follows, we performed a dedicated energy analysis on a separate set of runs (Tables E1 and E2); since these runs are distinct from the evaluation sets of Tables 1 and 2, their absolute values differ slightly from those tables. Table E1 shows that energy follows the number of model calls rather than wall-clock time. The measured saving is larger than the constant-power reference 1 − 1/speedup on both models (56.4% against 52.6% on SANA-WM, and 69.5% against 67.0% on MiniWorld-1B), so the energy measurements carry information that the speedup does not. The reason is that a fresh model call costs the same energy with and without DARC, while DARC's predicted steps add a little time at low power and almost no energy. The saving is therefore set by the calls that DARC skips: the share of Exact energy spent in model calls, multiplied by the fraction of calls removed, predicts the measured saving to within one percentage point (57.1% and 70.4%). Energy measurements also reveal overheads that latency hides: NaviCache, for instance, gives no speedup on Cosmos yet uses up to 6.0% more energy (Table 2).

Table E1 also splits the saved energy by power rail, and Table E2 breaks down the power of a model call on the H200 and the Thor. Since model loading, the refiner and decoding are left unchanged, DARC's saving comes from the model calls it skips. Most of a skipped call's energy is GPU compute, about a quarter is idle power that the shorter run no longer pays, and memory accounts for only a small share on both devices. We have added Tables E1 and E2 and this analysis to the revised Appendix A.3.


Table E1. Energy of Exact and DARC on the Jetson AGX Thor: the saving follows model calls, not wall time. From the per-rail tegrastats logs of dedicated energy-analysis runs, separate from the evaluation sets of Tables 1 and 2. Energy per call averaged over every call of each run. Energy saving: measured, relative to Exact. Share of the saved energy: Exact minus DARC, split by power rail; losses are unmonitored supplies and conversion losses. †The 21.5 W idle power times the time saved; it is contained in the four rail columns, not added to them. The last five columns give the share of the saved energy.


MethodFresh calls ↓Energy per callSpeed ↑Energy saving ↑GPU railCPU, SoC, memory5 V railLossesIdle floor†
SANA-WM
Exact6011.13 kJ1.00×------
DARC (Ours)21.211.14 kJ2.11×56.4%55.6%16.2%16.5%11.7%22.1%
MiniWorld-1B
Exact23031.1 J1.00×------
DARC (Ours)5832.1 J3.03×69.5%46.6%22.5%18.7%12.1%29.0%

Table E2. Where the power of a model call goes. From dedicated energy-analysis runs, separate from the evaluation sets of Tables 1 and 2. H200: GPU board power with hardware counters; compute is the SM and tensor cores, memory is memory traffic. Jetson AGX Thor: module input power split by the rails that tegrastats reports, above a measured 21.5 W idle floor; compute is the GPU rail. The Thor has no per-unit counters, and its memory subsystem shares a rail with the CPU and SoC, so its memory entry is an upper bound on memory power. Shares are relative to the power of one model call, except the last column, which is relative to the full model's energy per video.

DevicePower per model callComputeIdle floorMemory5 V rail and conversion lossesUnchanged stages
H200422-590 W64-79%17-24%≤ 4-16%-2-29%
Jetson AGX Thor74-97 W43-65%22-29%≤ 5-14%8-15%2-12%

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers


We are grateful to the reviewers for their time and helpful comments. We are glad that Reviewer BmSZ found the evaluation broad and careful, Reviewer CEXN found the idea of DARC clever and well motivated, and Reviewer Z25m found the motivation convincing and valued our ablations and energy measurements. A short summary of the new results follows.


  • Comparison at matched speed: We re-ran all baselines, including the recent RACER, on the full benchmarks of all four model families, with each baseline tuned per benchmark to run at DARC's speed (2.26× to 3.35×). Measured against the unaccelerated model's output for the same sample and seed, DARC is the closest on all seven benchmarks and all three metrics, exceeding the best baseline by +0.7 to +1.4 dB PSNR.
  • Comparison with WorldDynCache: At WorldDynCache's published operating point on Aether-5B (same share of model calls, 2.15× speedup), DARC reaches 35.10 vs. 33.21 dB PSNR and 0.035 vs. 0.061 LPIPS.
  • Blur and revisit consistency: DARC keeps 97.4% of the unaccelerated model's sharpness. Blurring the unaccelerated output to the same sharpness raises revisit PSNR by only +0.009 dB, against +0.20 dB for DARC, so the revisit gain does not come from lost detail.
  • Measured energy: On the Jetson AGX Thor, DARC reduces energy by 56.4% (SANA-WM, 2.11×) and 69.5% (MiniWorld, 3.03×), more than the constant-power reference 1 − 1/speedup (52.6% / 67.0%). A model call costs the same energy with and without DARC, so the saving follows the number of calls skipped.
  • Routing versus fixed schedules: With the same predictor and fewer fresh calls (17.1 vs. 18), DARC's routing outperforms the best of three fixed refresh schedules (0.145 vs. 0.175 LPIPS), showing that where calls are placed matters, not only how many are made.
  • One configuration across models: All new experiments use a single configuration shared by the four model families; only the budget ε is set per benchmark, as each baseline sets its own threshold. Changing the span cap or the routing weights moves DARC smoothly along the speed-quality curve, a weaker forecast is compensated by more fresh calls without divergence, and higher-order predictors bring no gain.
  • Long rollouts and rapid transitions: In MiniWorld rollouts of up to 1,021 frames (128 autoregressive chunks), fidelity decays smoothly, and DARC has the highest PSNR over every stretch, with WorldCache close behind. On the 10% fastest-changing frames, DARC leads the closest baseline by +1.43/+0.54/+1.91 dB on Sekai/DROID/RealEstate10K (the DROID lead is not significant).
  • Changes to the paper: We now report both routing correlations together (ρ = 0.95 against each method's own error; 0.61 against a common error, versus at most 0.36 for prior caches), describe the tuning protocol of every method, clarify that the OOM entries are actual out-of-memory crashes, add a table of benchmark characteristics, and discuss RACER and WorldDynCache in the related work. All changes are included in the revised manuscript. Table numbers in our responses refer to the submitted version, whereas tables numbered R are rebuttal tables, numbered separately in each response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.