EpaCache: Error-Propagation-Aware Caching for Accelerating Diffusion-Based Visual Generation

Yuhan Liu; Zongwei Hong; Jinglun Li; Linze Li; Shen Zhang; Yao Tang

cs.AIAug 29, 2026

EpaCache: Error-Propagation-Aware Caching for Accelerating Diffusion-Based Visual Generation

Authors: Yuhan Liu, Zongwei Hong, Jinglun Li, Linze Li, Shen Zhang, Yao Tang

Organizations: JIIOV Technology

Abstract

Diffusion-based visual generative models deliver strong image and video synthesis quality but incur high inference costs because sequential samplers repeatedly evaluate large networks. Caching-based methods reduce inference latency by reusing intermediate computations across adjacent timesteps. However, existing cache controllers rely primarily on local temporal variation and overlook the trajectory-level consequences of cache reuse. We introduce Error-Propagation-Aware Cache (EpaCache), a training-free caching policy that adaptively allocates the reuse budget on timesteps with lower downstream impact. Experiments on image and video synthesis models demonstrate that EpaCache consistently improves the latency--fidelity trade-off over existing caching methods. On FLUX.1-dev, EpaCache outperforms the prior state-of-the-art caching method in both latency and fidelity, reducing inference time from $11.7$ s to $11.3$ s while improving PSNR from $21.4$ to $22.8$ . On HunyuanVideo, EpaCache achieves a $2.63\times$ speedup over uncached inference and improves SSIM from $0.891$ to $0.905$ over the prior state-of-the-art method at matched latency.

Figures & tables

Figure 1: (a) PSNR–latency trade-off compared with Vanilla, TaylorSeer, TeaCache, and SeaCache, where the upper-left direction indicates better performance. (b) A representative case on FLUX.1-dev.

Figure 2: L1 deviation trajectories caused by cache-induced approximation error at selected denoising steps $s$ . Darker curves indicate larger terminal deviations.

Figure 3: Illustration of EpaCache’s error-propagation-aware threshold allocation. By getting the downstream impact of each cacheable timestep over the calibration prompts, the inverse-normalized impact profile is scaled by the user-specified mean threshold

\bar{\delta}

to obtain timestep-specific reuse thresholds. Therefore, high-impact timesteps receive smaller thresholds, whereas low-impact timesteps receive larger thresholds.

Figure 4: Overview of the EpaCache caching proxy and the online cache decision. The adopted timestep-modulated proxy estimates the local variation

\Delta_{k}

. The accumulated variation determines whether the cached residual

R_{a}

is reused or the full DiT computation is conducted.

FLUX.1-dev (image, $1024\times 1024$ )
Method	Latency (s) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
FLUX.1-dev ( $T=50$ )	43.0	–	–	–
FLUX.1-dev ( $T=25$ )	22.1	17.7	0.731	0.303
FLUX.1-dev ( $T=15$ )	13.3	15.3	0.648	0.420
TeaCache ( $\delta=0.3$ )	23.0	20.8	0.807	0.199
TaylorSeer ( $\mathcal{S}=3$ )	17.4	19.8	0.780	0.222

Table 1: Quantitative comparison on FLUX.1-dev under two efficiency regimes, corresponding to mean skip rates of approximately 50% and 70%. For TaylorSeer, we use the default approximation order

\mathcal{O}=2

Figure 5: Qualitative comparison of EpaCache and the baselines on FLUX.1-dev at skip rates of approximately 50% and 70%.

Figure 6: Qualitative comparisons on Wan2.1 and HunyuanVideo. All methods are compared at matched frame indices; for HunyuanVideo, we show two representative frames within a video.

Wan2.1-1.3B (81 frames, $832\times 480$ )
Method	Latency $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Wan2.1-1.3B ( $T=50$ )	409.55	–	–	–
Wan2.1-1.3B ( $T=35$ )	289.05	16.74	0.6198	0.3105
TeaCache ( $\delta=0.09$ )	214.65	22.36	0.7960	0.1388
MagCache ( $\delta=0.12$ )	205.09	28.04	0.9128	0.0510
SeaCache ( $\delta=0.20$ )	209.59	29.60	0.9222	0.0496

Table 2: Quantitative comparison on Wan2.1-1.3B.

\tau

and

I

denote the error threshold and the identity matrix, respectively, which are hyperparameters in NaviCache ( Lv et al. 2026 ) .

Method	Venue	Efficiency				Visual Retention			VBench $\uparrow$
Method	Venue	Latency (s) $\downarrow$	Speedup $\uparrow$	Memory (GB) $\downarrow$	FLOPs (P) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	VBench $\uparrow$
HunyuanVideo (65 frames, $960\times 544$ )
HunyuanVideo ( $T=50$ )	–	753.84	1.00 $\times$	45.60	24.904	–	–	–	81.11
HunyuanVideo ( $T=25$ )	–	392.44	1.92 $\times$	45.60	13.031	20.16	0.7370	0.2759	80.95
TeaCache ( $\delta=0.18$ )	CVPR’25	287.40	2.62 $\times$	46.20	9.580	21.59	0.7653	0.2378	81.14
MagCache ( $\delta=0.24$ )	NeurIPS’25	285.11	2.64 $\times$	45.80	9.505	29.44	0.8965	0.0877	80.19

Table 3: Efficiency, visual retention, and VBench comparison on HunyuanVideo. FLOPs are reported in peta floating-point operations (P). The best and second-best results for each metric are highlighted in bold and underline , respectively.

Figure 7: Refresh pattern across timesteps on FLUX.1-dev on DrawBench. Each strip proceeds from the first to the last of the 50 denoising steps. Darker colors indicate lower skip probabilities and more frequent recomputation.

Figure 8: Visualization of the per-step threshold-allocation schedules used in the ablation study. For Minimum 50% $L_{t}$ , the steps with the lowest 50% of $L_{t}$ values are assigned an effectively infinite threshold (denoted by $\infty$ ), forcing cache reuse, while the remaining steps are assigned a threshold of zero.

Line Shape	Latency $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Static( $\delta=0.08$ )	209.61	21.04	0.7454	0.1798
Minimum 50% $L_{t}$	232.45	23.87	0.7412	0.1865
Reverse $L_{t}$ ( $\bar{\delta}=0.08$ )	205.73	22.20	0.7799	0.1491
Slash ( $\bar{\delta}=0.08$ )	207.89	24.10	0.8310	0.1089
Ours ( $\bar{\delta}=0.08$ )	215.49	28.33	0.9058	0.0577

Table 4: Ablation study of threshold allocation on Wan2.1. We compare the proposed inverse-normalized Final-L1-weighted schedule with linear Slash schedules, a reverse

L_{t}

schedule, and hard Minimum

50\%

step-selection policies on 100 VBench prompts. For continuous schedules, the mean threshold is kept fixed when comparing curve shapes.

$N$	Latency (s) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1 ( $\bar{\delta}=0.45$ )	11.3	22.4	0.804	0.218
5 ( $\bar{\delta}=0.45$ )	9.6	20.0	0.761	0.276
10 ( $\bar{\delta}=0.45$ )	11.1	22.3	0.804	0.216
20 ( $\bar{\delta}=0.45$ )	11.3	22.8	0.812	0.207

Table 5: Ablation study on the number of calibrated prompts

N

using FLUX.1-dev.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Visualization of the error-propagation-aware calibration steps. For a cache injected at a timestep

s

, the remaining

T-s

denoising steps are executed to evaluate the effect of cache reuse.

FLUX.1-dev (image, $1024\times 1024$ )
Scheme	Per-Prompt Cost ( $\times$ ) $\downarrow$	Latency (s) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Standard inference	1.00	43.0	–	–	–
$\bar{\delta}=0.20$
Full calibration	24.52	18.0	27.6	0.905	0.088
Interpolation ( $1/2$ steps)	12.26	18.2	27.9	0.905	0.085
Interpolation ( $1/3$ steps)	8.17	18.8	28.0	0.910	0.083

Appendix

Table 6: Effect of sparse-step calibration on EpaCache for FLUX.1-dev. The reported cost is the per-prompt profiling cost relative to one standard inference; the total cost for

N

calibration prompts is

N

times the reported value.

Semantic Dimension $\uparrow$	Original (50 steps)	Original (25 steps)	EpaCache	TeaCache	MagCache	SeaCache
Object Class	79.27	79.35	79.11	77.69	75.95	78.09
Multiple Objects	72.79	66.08	73.09	69.28	71.88	74.16
Human Action	92.00	92.00	92.00	90.00	88.00	88.00
Color	90.96	93.92	90.70	91.81	91.40	91.05
Spatial Relationship	70.64	67.14	67.09	75.88	64.99	67.16
Scene	30.05	33.06	28.46	30.76	31.29	29.43

Appendix

Table 7: Dimension-level VBench Semantic Scores on the 540p, 65-frame evaluation setting. Bold and underlined entries denote the best and second-best results, respectively, among the four caching methods; the original-model columns are excluded from this comparison.

Figure 10: Qualitative comparison of video generated by the uncached reference, EpaCache, TeaCache, MagCache, and SeaCache with the prompt "a bear sniffing the air for scents of food" . The prompt measures the VBench metric Dynamic Degree . The figure also reports reference-fidelity metrics and the Dynamic Degree decision for each caching method.

Metric	TeaCache	EpaCache	Better Method
Dynamic Degree $\uparrow$	$43/72=0.5972$	$37/72=0.5139$	TeaCache
Dynamic agreement with Standard $\uparrow$	$64/72=88.9\%$	$68/72=94.4\%$	EpaCache
PSNR $\uparrow$	$21.31$	$29.49$	EpaCache
SSIM $\uparrow$	$0.7127$	$0.9021$	EpaCache
LPIPS $\downarrow$	$0.2357$	$0.0720$	EpaCache
Subject Consistency $\uparrow$	$0.964116$	$0.964839$	EpaCache

Appendix

Table 10: Comparison between Dynamic Degree and other quality metrics.

Figure 11: Supplementary qualitative comparison on FLUX.1-dev. Each row compares the uncached baseline with EpaCache (Ours), TeaCache, TaylorSeer, and SeaCache using the same configuration.

Diffusion Model Caching

Diffusion model caching reuses features, activations, or attention states across denoising steps to reduce inference cost while preserving output quality.

See all 112 papers

SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching
Feb 27, 2026Yasaman Haghighi, Alexandre AlahiVideo Diffusion Models Diffusion Model Caching
Ecole Polytechnique Fédérale de Lausanne (EPFL)
From Local Mismatch to Global Impact: Optimizing Cache Reuse Policy for Efficient Diffusion
Aug 13, 2026Xichen Ye, Yifan Wu, Zhikang Xie +3Video Diffusion Models Diffusion-Based Image Generation
College of Computer Science and Artificial Intelligence, Fudan University · Faculty of Engineering, The Chinese University of Hong Kong · School of Data Science, Fudan University
OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference
Jul 31, 2026Zhikang Xie, Xichen Ye, Yifan Wu +5Diffusion Model Caching Diffusion Model Inference Acceleration
College of Computer Science and Artificial Intelligence, Fudan University · School of Data Science, Fudan University · 3Shanghai Key Laboratory of Intelligent Information Processing

Feb 27, 2026J/K move · Enter open · S save

SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching

Yasaman Haghighi, Alexandre Alahi

Ecole Polytechnique Fédérale de Lausanne (EPFL)

Video Diffusion Models Diffusion Model Caching Diffusion Model Inference Acceleration

Open paper AskPDF arXiv

Diffusion models achieve state-of-the-art video generation quality, but their inference remains expensive due to the large number of sequential denoising steps. This has motivated a growing line of research on accelerating diffusion inference. Among training-free acceleration methods, caching reduces computation by reusing previously computed model outputs across timesteps. Existing caching methods rely on heuristic criteria to choose cache/reuse timesteps and require extensive tuning. We address this limitation with a principled sensitivity-aware caching framework. Specifically, we formalize the caching error through an analysis of the model output sensitivity to perturbations in the denoising inputs, i.e., the noisy latent and the timestep, and show that this sensitivity is a key predictor of caching error. Based on this analysis, we propose Sensitivity-Aware Caching (SenCache), a dynamic caching policy that adaptively selects caching timesteps on a per-sample basis. Our framework provides a theoretical basis for adaptive caching, explains why prior empirical heuristics can be partially effective, and extends them to a dynamic, sample-specific approach. Experiments on Wan 2.1, CogVideoX, and LTX-Video show that SenCache achieves better visual quality than existing caching methods under similar computational budgets.

Abstract

Diffusion Model Caching

Explore similar work

SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching

From Local Mismatch to Global Impact: Optimizing Cache Reuse Policy for Efficient Diffusion

OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference