MORCA: Offline-to-Online Reinforcement Learning for Adaptive Cache Reuse in Video Diffusion Acceleration

Yuxiang Xiong; Ruiyan Wang; Wenqiang Wang; Teng Hu; Songhang Shen; Bohao Feng; Hongqian Deng; Ran Yi

cs.CVOct 7, 2026

MORCA: Offline-to-Online Reinforcement Learning for Adaptive Cache Reuse in Video Diffusion Acceleration

Authors: Yuxiang Xiong, Ruiyan Wang, Wenqiang Wang, Teng Hu, Songhang Shen, Bohao Feng, Hongqian Deng, Ran Yi

Organizations: Shanghai Jiao Tong University · Alibaba Cloud Computing · Alibaba Token Hub, Alibaba Group

Abstract

Diffusion Transformers (DiTs) achieve remarkable performance in video synthesis, but their iterative denoising process suffers from high inference latency. To address this, caching has emerged as an effective acceleration strategy by capitalizing on inter-step redundancy during denoising. Existing dynamic caching methods typically estimate the error that cache reuse would introduce at each denoising step (step error) to guide cache decisions, whereas our concern is how much quality loss cache reuse would cause in the final generated video (terminal error). We show that step error does not directly correspond to terminal error and that latent information helps capture their relationship, thereby informing cache decisions. Moreover, existing threshold-based methods cannot provide precise speedup control, making it difficult to meet practical requirements for user-specified acceleration targets. To address these limitations, we introduce MORCA, a cache scheduling framework trained through offline-to-online reinforcement learning to make latent-aware reuse/recompute decisions under user-specified acceleration targets. Extensive experiments on different video generation models across multiple target acceleration ratios demonstrate that MORCA achieves better generation fidelity than state-of-the-art caching methods under comparable computational budgets. Code is available at https://github.com/x10ngyx/MORCA.

Figures & tables

Figure 1: Step error and terminal error. (a) Following SeaCache, we cache and reuse the residuals of the DiT blocks. For each step, we evaluate two cases: (i) reusing the residual cached at the previous step and (ii) recomputing the DiT blocks, with all other steps computed normally. We measure step error as the relative L1 difference between the cached and recomputed residuals, and terminal error using the PSNR of the resulting video with cache reuse relative to the no-cache reference. (b) Step error does not directly correspond to terminal error. Results are shown for Wan2.1.

Input	RMSE (dB) $\downarrow$	Spearman $\uparrow$
Step error	6.971	0.5288
Step error + step index	4.421	0.7528
Step error + latent	2.849	0.9357
Step error + latent + step index	2.832	0.9361

Table 1: Latent features improve terminal-error prediction. Terminal error is measured using PSNR. Results on Wan2.1.

Figure 2: Overview of MORCA. Bottom: A learned scheduler makes reuse/recompute decisions at each denoising step. Top left: The scheduler is budget-constrained and latent-aware. The target speedup is mapped to the required total number of reuse steps, which, together with step-error proxies and step statistics, latent features, and the current reuse count extracted from the denoising state, is fed into the scheduler for decision-making (Sec. 3.2 ). Top right: The scheduler is optimized through offline-to-online reinforcement learning using IQL (Sec. 3.3 ).

Figure 3: Calibration of reuse budget $K$ against (a) inference latency and (b) speedup on Wan2.1. The fitted functions take the forms (a) $K=\alpha-\beta T_{\mathrm{inf}}$ and (b) $K=\alpha-\beta/\rho$ .

Model	Target	Method	Latency (s) $\downarrow$	Speedup $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Wan2.2	—	Original	990.56	$1.00\times$	—	—	—
	$1.8\times$	TeaCache	562.63	$1.76\times$	17.04	0.5984	0.2882
		MagCache	544.78	$1.82\times$	16.41	0.5662	0.3041
		DiCache	551.29	$1.80\times$	18.51	0.6516	0.2232
		SeaCache	546.99	$1.81\times$	19.91	0.7016	0.1888
		Ours	549.19	$1.80\times$	22.13	0.7507	0.1681

Table 2: Quantitative comparison with baselines. The best result is highlighted in bold, while the second-best result is underlined.

Figure 4: Qualitative comparison with baselines. Differences are highlighted in the boxed regions.

Variant	$\mathcal{P}_{3}(x)$	$\mathcal{P}_{2}(\mu_{T}(x))$	$\mathcal{P}_{2}(\sigma_{T}^{2}(x))$	Latency (s) $\downarrow$	Speedup $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$A_{0}$	$\cdot$	$\cdot$	$\cdot$	85.46	$2.99\times$	19.93	0.6420	0.2240
$A_{1}$	$\checkmark$	$\cdot$	$\cdot$	86.10	$2.97\times$	19.92	0.6515	0.2273
$A_{2}$	$\cdot$	$\checkmark$	$\cdot$	86.89	$2.95\times$	19.76	0.6429	0.2314
$A_{3}$	$\cdot$	$\cdot$	$\checkmark$	85.39	$3.00\times$	19.27	0.6151	0.2441
$A_{4}$	$\checkmark$	$\checkmark$	$\cdot$	85.38	$3.00\times$	20.06	0.6598	0.2197
$A_{5}$	$\checkmark$	$\cdot$	$\checkmark$	85.29	$3.00\times$	19.83	0.6484	0.2320

Table 3: Ablation on Latent-Aware State Design. Offline training results on Wan2.1 at approximately

3.0\times

speedup.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Wan2.2	Wan2.1
Resolution ( $W\times H$ )	$832\times 480$	$832\times 480$
Frames	45	81
Frame rate	16 fps	16 fps
Sampler	DPM++	UniPC
Denoising steps	50	50
Flow shift	12	5

Appendix

Table 5: Inference configurations for Wan2.2 and Wan2.1.

Group	Setting	Offline IQL	Online fine-tuning
Training	Training duration	200 epochs	16 rounds
	Optimizer	AdamW	AdamW
	$\pi$ LR	$10^{-4}$	$4\times 10^{-5}$
	Q/V LR	$10^{-4}$	$10^{-4}$
	Batch size	256	256
	Adam betas	$(0.9,0.999)$	$(0.9,0.999)$

Appendix

Table 6: Training and IQL hyperparameters for the two stages, shared across Wan2.2 and Wan2.1.

Model	Method	$\bm{1.8\times}$	$\bm{2.4\times}$	$\bm{3.0\times}$
Wan2.2	TeaCache	$\delta=0.29$	$\delta=0.45$	$\delta=0.68$
	MagCache	$\delta=0.072$ , $K=2$ , $R=0.2$	$\delta=0.199$ , $K=4$ , $R=0.2$	$\delta=0.198$ , $K=5$ , $R=0.1$
	DiCache	$\delta=0.075$ , $R=0.2$	$\delta=0.172$ , $R=0.2$	$\delta=0.396$ , $R=0.2$
	SeaCache	$\delta=0.24$	$\delta=0.38$	$\delta=0.55$
Model	Method	$\bm{2.4\times}$	$\bm{3.0\times}$	$\bm{4.0\times}$
Wan2.1	TeaCache	$\delta=0.142$	$\delta=0.195$	$\delta=0.36$

Appendix

Table 7: Baseline configurations at three target acceleration ratios for each model.

\delta

denotes the cache threshold,

K

the maximum number of consecutive reuse steps for MagCache, and

R

the retention ratio.

Target	Latency (s)	Scheduler time (s)	Proportion (%)
$1.8\times$	549.19	3.95	0.72
$2.4\times$	421.15	3.94	0.94
$3.0\times$	326.93	3.93	1.20

Appendix

Table 8: Scheduler overhead on Wan2.2. The scheduler introduces acceptable overhead.

Figure 5: Visualization of schedules and corresponding generation results under the same reuse budget on Wan2.2.

Figure 6: Speed–quality trade-offs of MORCA and SeaCache on Wan2.2. MORCA consistently achieves better fidelity at comparable speedups.

Figure 7: Additional qualitative comparisons on Wan2.2. From top to bottom: target acceleration ratios of $1.8\times$ , $2.4\times$ , and $3.0\times$ .

Figure 8: Additional qualitative comparisons on Wan2.1. From top to bottom: target acceleration ratios of $2.4\times$ , $3.0\times$ , and $4.0\times$ .

Explore similar work

SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching
Feb 27, 2026Yasaman Haghighi, Alexandre AlahiDiffusion Models Cache
Ecole Polytechnique Fédérale de Lausanne (EPFL)
ACID: Adaptive Caching for vIDeo generation
Jul 14, 2026Om Agrawal, Saurabh Agarwal, Aditya AkellaVideo Diffusion Models Cache
UT Austin
OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference
Jul 31, 2026Zhikang Xie, Xichen Ye, Yifan Wu +5Iterative Denoising Methods Cache
College of Computer Science and Artificial Intelligence, Fudan University · School of Data Science, Fudan University · 3Shanghai Key Laboratory of Intelligent Information Processing

Feb 27, 2026, cs.CVJ/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)

Diffusion Models Cache Timestep

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

Explore similar work

SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching

ACID: Adaptive Caching for vIDeo generation

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