CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration

Qinhan Lyu; Jusheng Zhang; Zhixing Zhang; Hao Liu; Zijie Liang; Jin Jiang; Keze Wang

cs.LGOct 8, 2026

CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration

Authors: Qinhan Lyu, Jusheng Zhang, Zhixing Zhang, Hao Liu, Zijie Liang, Jin Jiang, Keze Wang

Organizations: Sun Yat-sen University

Abstract

Training-free feature reuse accelerates diffusion transformer inference by replacing full denoising computation at selected steps with cached approximations. Existing methods mainly improve the reconstruction of skipped features, but we find that once a stable first-order reconstructor is used, increasing reconstruction complexity brings limited gains, whereas the placement of exact-computation steps has a much larger effect on fidelity. More importantly, absolute shallow-probe variation is jointly affected by input-trajectory displacement and network response strength, and therefore does not directly reflect cache risk. We address this confounding by normalizing probe-output variation with the corresponding input displacement, yielding a directional gain per unit trajectory movement. Based on this signal, we propose CacheScope, which identifies high-risk reuse steps online and selectively falls back to full denoising computation while leaving the underlying cache reconstructor unchanged. CacheScope requires no additional training, gradients, auxiliary networks, or prompt-specific offline schedule optimization. Across FLUX.1-dev, Wan2.1-T2V-1.3B, and Open-Sora 1.2, CacheScope achieves up to 3.57x acceleration and reduces median LPIPS by up to 70.1% over the strongest probe-based baseline.

Figures & tables

Figure 1: Overview of the CacheScope insight. Under a fixed cache reconstructor and a fixed exact-compute budget, more flexible reconstruction rules do not consistently outperform a bounded first-order reconstructor, while the placement of exact evaluations dominates fidelity. Raw probe variation

\Delta y_{t}

conflates trajectory displacement with network response; CacheScope instead schedules exact computation using the trajectory-conditioned sensitivity

s_{t}=\Delta y_{t}/(\Delta x_{t}+\epsilon)

Figure 2: Overview of CacheScope. At every denoising step, the first transformer block is executed as a shallow probe. From its input and output we compute the relative trajectory movement $\Delta x_{t}$ and probe response $\Delta y_{t}$ , and accumulate the trajectory-conditioned sensitivity $s_{t}=\Delta y_{t}/(\Delta x_{t}+\epsilon)$ — the only change relative to raw probe scheduling. When the accumulator reaches the threshold $\tau$ , the remaining blocks are evaluated exactly and the cache is refreshed; otherwise the step is reconstructed by the unchanged bounded first-order reconstructor. Under classifier-free guidance, the two branches maintain independent accumulators and cache states.

Backbone	Regime	Method	Param.	$\rho$ (%)	Lat. (s) $\downarrow$	Speedup $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Wan2.1 [-1pt] 50 steps, 81 frames [-1pt] $832{\times}480$	Moderate	TeaCache ( Liu et al. 2025a )	$\delta{=}0.14$	60	102.41	2.40 $\times$	0.1153	23.83	0.8676
		MagCache ( Ma et al. 2025 )	$\delta{=}0.12$	58	107.10	2.29 $\times$	0.0663	27.26	0.9153
		EasyCache ( Zhou et al. 2025 )	$\tau{=}0.058$	64	92.59	2.65 $\times$	0.0909	24.89	0.8319
		DiCache ( Bu et al. 2025 )	$\delta{=}0.16$	64	98.47	2.49 $\times$	0.0835	25.78	0.8958
		CacheScope	$\tau{=}5.5$	64	98.39	2.49 $\times$	0.0414	28.89	0.9220
	Fast	TeaCache ( Liu et al. 2025a )	$\delta{=}0.20$	72	73.76	3.33 $\times$	0.2335	20.12	0.7358

Table 1: Main results.

\rho

denotes the fraction of skipped model evaluations, and Param. denotes each method’s original speed-control hyperparameter. Within each backbone and regime, the best and second-best fidelity results in the target speed range are bolded and underlined, respectively; “–” indicates that the target operating point is unavailable.

Backbone	Regime	Scheduler	Speedup $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Wan2.1	Moderate	D: $\Delta y$	2.492	0.0772	26.010	0.8961
		M: $\|\Delta y-\Delta x\|$	2.491	0.0470	28.82	0.9329
		S: Probe-normalized	2.493	0.0414	28.891	0.9220
	Fast	D: $\Delta y$	3.462	0.2120	21.013	0.7437
		M: $\|\Delta y-\Delta x\|$	3.463	0.2517	19.69	0.6131
		S: Probe-normalized	3.463	0.1607	21.802	0.7363

Table 2: Controlled scheduler comparison on

32

held-out prompts. The three signals (D, M, S) share the same DiCache bounded first-order

\Gamma

reconstructor, probe depth, and warm-up policy, and differ only in the accumulated scheduling signal. Values are medians over prompts. Statistical testing uses the strictly matched-skip subsets described in Appendix F.2 . Best values within each backbone and regime are in bold; second-best values are underlined.

Figure 3: Qualitative comparison on FLUX.1-dev in the Fast regime. Columns show the dense output and different training-free caching methods, while rows correspond to three prompts. Red boxes highlight visible deviations from the dense reference. CacheScope better preserves fine structures and object geometry at $3.57\times$ speedup.

Figure 4: Reconstruction ablation and multi-metric comparison. (a) $\Delta$ LPIPS relative to the bounded first-order $\Gamma$ reconstructor under identical replayed exact/skip schedules (positive $=$ worse). (b) VBench sub-dimension scores on Wan2.1 (Moderate), min–max normalized per metric across methods ; raw scores are given in Appendix F.4 .

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Scheduler	Signal/source	Prompt assignment	Total exact evals	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	Paired LPIPS $p$
Online-S	$s_{t}$ , online	Matched	515	0.0816	27.74	0.9219	—
Global-S	Mean $s_{t}$ , dev16	Global	512	0.0792	27.54	0.9202	1.00
Global-D	Mean $\Delta y_{t}$ , dev16	Global	512	0.1337	23.31	0.8734	$2.6{\times}10^{-6}$
Cross-prompt replay	Mismatched Online-S masks	Deranged	515	0.0794	27.35	0.9201	1.00

Appendix

Table 3: Offline aggregation and schedule-information analysis on FLUX.1-dev at the Moderate operating point. Global-S and Global-D are constructed from the same 16-prompt development set using the same aggregation procedure, threshold solver, and target compute budget; they differ only in the scheduling signal. Metric values are medians over 32 held-out prompts. The exact-evaluation column reports totals across all 32 prompts. All variants have a median skip count of 34. Paired two-sided sign tests compare each variant with Online-S .

Backbone	Regime	Method	Speedup $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Wan2.1-T2V-1.3B	Moderate	SenCache	2.398 $\times$	0.0775	25.88	0.8908
		NaviCache	2.519 $\times$	0.0752	25.72	0.8642
		D2Cache	2.394 $\times$	0.0933	23.88	0.8778
		CacheScope	2.493 $\times$	0.0414	28.89	0.9220
	Fast	NaviCache	3.336 $\times$	0.2334	20.81	0.6560
		D2Cache	3.322 $\times$	0.2075	19.50	0.7269

Appendix

Table 4: Comparison with additional open-source baselines. All metrics are computed against the same dense references. CacheScope results are reproduced for comparison. SenCache is omitted from the Wan2.1 Fast regime because its fastest attainable configuration reaches only

2.644\times

, below the target speedup of approximately

3.46\times

. Best results within each backbone and regime are in bold.

Backbone	Regime	Skip ratio	$\tau_{\mathrm{S}}$	$\tau_{\mathrm{D}}$	$\tau_{\mathrm{M}}$
Wan2.1	Moderate	$0.64$	$5.5$	$0.16$	$0.028$
Wan2.1	Fast	$0.76$	$20.0$	$0.50$	$0.110$
FLUX	Moderate	$0.68$	$4.6$	$0.15$	$0.042$
FLUX	Fast	$0.74$	$6.4$	$0.25$	$0.065$

Appendix

Table 5: Frozen scheduling thresholds. S, D, and M differ only in their accumulated scheduling signal.

Backbone	Regime	Seed 42	Seed 43	$p_{42}$	$p_{43}$	Ratio
Wan2.1	Moderate	$-0.03353$	$-0.03905$	$9.31{\times}10^{-10}$	$2.16{\times}10^{-7}$	$1.16$
Wan2.1	Fast	$-0.05408$	$-0.04648$	$1.12{\times}10^{-7}$	$1.86{\times}10^{-9}$	$0.86$
FLUX	Moderate	$-0.05092$	$-0.03514$	$8.05{\times}10^{-7}$	$1.10{\times}10^{-5}$	$0.69$
FLUX	Fast	$-0.04578$	$-0.04469$	$1.20{\times}10^{-4}$	$3.05{\times}10^{-5}$	$0.98$

Appendix

Table 6: Independent-seed replication. Values are median paired LPIPS differences,

L_{\mathrm{S}}-L_{\mathrm{D}}

, on strict matched-skip pairs.

Backbone	Level	Pair	Pairing	$n$	Median $\Delta$ [95% CI]	W–T–L	$p$
Wan2.1	Moderate	S vs. M	matched	30	$-0.0051$ $[-0.0072,-0.0016]$	26–0–4	$<0.001$
			all	32	$-0.0062$ $[-0.0079,-0.0021]$	28–0–4	$<0.001$
		S vs. D	matched	31	$-0.0335$ $[-0.0504,-0.0209]$	31–0–0	$<0.001$
			all	32	$-0.0332$ $[-0.0498,-0.0209]$	32–0–0	$<0.001$
		M vs. D	matched	29	$-0.0273$ $[-0.0399,-0.0169]$	28–0–1	$<0.001$
			all	32	$-0.0273$ $[-0.0399,-0.0169]$	30–0–2	$<0.001$

Appendix

Table 7: Full prompt-level paired comparison of LPIPS on the held-out set. “Matched” restricts each pair to prompts with identical skip counts (primary read); “all” pools all

32

prompts. Negative median differences favor the first scheduler. W–T–L counts wins, ties, and losses of the first scheduler.

				LPIPS $\downarrow$		PSNR $\uparrow$		SSIM $\uparrow$
Backbone	Level	Pair	$n$	Median $\Delta$	Verdict	Median $\Delta$	Verdict	Median $\Delta$	Verdict
Wan2.1	Moderate	S vs. M	30	$-0.0051$	first	$-0.0186$	n.s.	$+0.0122$	second
		S vs. D	31	$-0.0335$	first	$-2.7257$	first	$-0.0124$	first
		M vs. D	29	$-0.0273$	first	$-3.4833$	first	$-0.0275$	first
	Fast	S vs. M	32	$-0.0634$	first	$-1.6990$	first	$-0.1107$	first
		S vs. D	29	$-0.0541$	first	$-0.6389$	first	$+0.0233$	second

Appendix

Table 8: Metric agreement on the held-out set, matched pairing only. Median paired differences; negative favors the first scheduler for all three metrics. “n.s.” denotes a sign-test tie at the

0.05

level.

Method	Subj.	BG	Flick.	Smooth	Aes.	Imag.	Overall	Dyn.
Wan2.1 (dense)	0.9723	0.9791	0.9843	0.9905	0.6152	0.7155	0.1974	0.2500
TeaCache	0.9710	0.9783	0.9847	0.9898	0.6225	0.7051	0.1976	0.2812
MagCache	0.9721	0.9790	0.9843	0.9898	0.6081	0.7091	0.1981	0.2188
EasyCache	0.9705	0.9761	0.9841	0.9901	0.5903	0.7127	0.1980	0.1875
DiCache	0.9719	0.9771	0.9843	0.9901	0.6128	0.7117	0.1956	0.2500
CacheScope	0.9813	0.9970	0.9852	0.9902	0.6274	0.7133	0.1986	0.2500

Appendix

Table 9: Raw VBench sub-dimension scores on Wan2.1 (Moderate regime). Higher is better for all dimensions except dynamic degree (Dyn.), where the dense baseline’s value is the reference and deviations in either direction indicate drift; Dyn. is therefore excluded from the bolding. Best value per column among the caching methods is in bold. Subj.

=

subject consistency, BG

=

background consistency, Flick.

=

temporal flickering, Smooth

=

motion smoothness, Aes.

=

aesthetic quality, Imag.

=

imaging quality, Overall

=

overall consistency.

Cell	Metric	$n$	Median	$p$
Wan2.1, FLUX-matched, S vs. D	SSIM	30	$-0.03180$	$0.043$
FLUX, ret0.2-Moderate, M vs. D	PSNR	31	$-0.06527$	$0.071$
Wan2.1, FLUX-matched, S vs. M	LPIPS	21	$-0.01347$	$0.078$
Wan2.1, FLUX-matched-lo, S vs. M	PSNR	28	$-0.11398$	$0.087$
Wan2.1, FLUX-matched-lo, S vs. M	SSIM	28	$+0.00621$	$0.087$
FLUX, Aggressive, S vs. M	PSNR	19	$-0.16893$	$0.167$

Appendix

Table 10: The

16

exploratory cells not rejected under Benjamini–Hochberg (

q{=}0.05

). Only the first was nominally significant before correction.

Retention $0.2$ (Wan2.1 main-table setting):
Backbone	Level	$\rho$	S vs. D	S vs. M	M vs. D
Wan2.1	Conservative †	0.52	$-0.0154$ (28)	$n{=}5$	$n{=}1$
Wan2.1	Moderate	0.64	$-0.0335$ (31)	$-0.0051$ (30)	$-0.0273$ (29)
Wan2.1	Fast	0.76	$-0.0541$ (29)	$-0.0634$ (32)	$+0.0098$ (29)
Retention $0.1$ (matched to FLUX):
Wan2.1	FLUX-matched-lo †	0.68	$-0.0729$ (31)	$-0.0058$ (28)	$-0.0612$ (27)

Appendix

Table 11: Paired LPIPS medians on identical-skip prompts as a function of the skip ratio

\rho

; negative favors the first scheduler,

n

in parentheses. Cells with

n<8

are below the pre-registered minimum and reported without adjudication.

\dagger

marks post-hoc levels.

Level	Scheduler	$\bar{s}$	LPIPS $\downarrow$	PSNR $\uparrow$	Speedup	Gate	Paired LPIPS (matched)
Wan2.1 Conservative †	S	52	0.0169	32.24	1.95 $\times$	FAIL	S vs. D: $n{=}28$ , $-0.01539$ , 23–0–5, $p{<}0.001$
( $\rho{=}0.52$ , ret. $0.2$ )	D	52	0.0429	29.10	1.95 $\times$		S vs. M: $n{=}5$ , $-0.00795$ , 5–0–0, $p{=}0.062$ (n.a.)
	M	54	0.0284	30.56	2.02 $\times$		M vs. D: $n{=}1$ , $-0.03604$ , 1–0–0, $p{=}1.000$ (n.a.)
Wan2.1 FLUX-matched-lo †	S	68	0.1041	24.05	2.66 $\times$	PASS	S vs. D: $n{=}31$ , $-0.07290$ , 28–0–3, $p{<}0.001$
( $\rho{=}0.68$ , ret. $0.1$ )	D	68	0.1787	21.08	2.75 $\times$		S vs. M: $n{=}28$ , $-0.00576$ , 21–0–7, $p{=}0.013$
	M	68	0.0963	23.81	2.75 $\times$		M vs. D: $n{=}27$ , $-0.06117$ , 27–0–0, $p{<}0.001$

Appendix

Table 12: Post-hoc operating points (held-out set,

n{=}32

prompts, seed

42

\bar{s}

is the median skip count. The matched-compute gate passes when the median-skip spread across the three schedulers is zero. Paired LPIPS rows use the matched pairing; cells with

n<8

are reported without adjudication. For Wan2.1/Aggressive, the M arm was calibrated to

82

on the development set but reached

83.5

on the held-out set (calibration drift); for Wan2.1/Conservative, no threshold places M at skip

52

because its median skip jumps from

49

54

between adjacent threshold values on the development grid.

Retention	S	D	M
$r=0.1$ (8–10 free)	0.423	0.570	0.471
$r=0.2$ (4 free)	0.449	0.635	0.595

Appendix

Table 13: Where each scheduler spends its exact forward passes (FLUX,

\rho{=}0.68

). Top: centroids of the kept passes. Bottom: pairwise placement statistics;

\Delta

LPIPS repeats the held-out paired median for alignment. Mechanism columns are measured on the

16

development prompts, the only runs with per-step traces.

Variant	Signal or schedule source	Prompt-specific	Exact fwd. total	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	Median $\Delta L$ ; $p$
Online-S	$s_{t}$ , evaluated online	Yes	515	0.0816	27.74	0.9219	Reference
Global-S	Mean $s_{t}$ on development set	No	512	0.0792	27.54	0.9202	$+0.0000$ ; $p{=}1.000$
Global-D	Mean $\Delta y_{t}$ on development set	No	512	0.1337	23.31	0.8734	$+0.0437$ ; $p{=}2.6{\times}10^{-6}$
Cross-prompt replay	Online-S mask from another prompt	Mismatched	515	0.0794	27.35	0.9201	$+0.0000$ ; $p{=}1.000$

Appendix

Table 14: Offline signal aggregation and schedule-information analysis on FLUX.1-dev at the Moderate operating point (

\rho{=}0.68

). Global-S and Global-D are constructed from the same

16

-prompt development set using the same timestep aggregation, threshold solver, and skip budget; they differ only in the aggregated scheduling signal. Quality values are medians over

32

held-out prompts. The paired LPIPS difference is

\Delta L=L_{\mathrm{variant}}-L_{\mathrm{Online\text{-}S}}

; positive values favor Online-S. The final column reports an exact two-sided sign test against Online-S.

	Wan2.1				FLUX
Reconstructor	Mod.	W–L	Fast	W–L	Mod.	W–L	Fast	W–L
Direct reuse	0.0823	16–0	0.1314	15–1	0.1634	0–16	0.4020	8–8
$\gamma$ extrap. (used)	0.0357	16–0	0.0747	16–0	0.1170	16–0	0.3379	14–2
LS, $k{=}1$	0.0812	–	0.1312	–	–	–	–	–
LS, $k{=}2$	0.3128	(ref)	0.5273	(ref)	0.1295	(ref)	0.3770	(ref)
LS, $k{=}4$	0.6670	1–15	0.5894	6–10	0.1266	7–9	0.4537	1–15

Appendix

Table 15: Median LPIPS of five cached-residual reconstructors on the

16

-prompt development set, with W–L counts against the least squares

k{=}2

reference. All arms replay identical exact/skip schedules.

Scheduler	Level	$\gamma_{\max}$	$n_{\mathrm{eff}}$	Median $\Delta$ LPIPS	$p$
S	Fast	2.0	16	$-0.00453$	0.077
S	Fast	3.0	15	$+0.00392$	1.000
S	Fast	$\infty$	15	$+0.00519$	1.000
S	Aggressive	2.0	16	$+0.00612$	0.804
S	Aggressive	3.0	14	$+0.01785$	0.180
S	Aggressive	$\infty$	13	$+0.02660$	0.267

Appendix

Table 16: Ablation of

\gamma_{\max}

on the

16

-prompt development set, paired against the default

\gamma_{\max}{=}1.5

. No cell reaches significance (smallest of the

12

p

-values:

0.077

, before any correction).

Backbone	Blocks	Tok./fwd	$N$	Non-skip.	Exact fwd. FLOPs	Probe fwd. FLOPs
Wan2.1	30	32760	100	20	$2.83{\times}10^{14}$	$9.43{\times}10^{12}$ (3.3%)
FLUX	19+38	4608	50	6+1	$7.44{\times}10^{13}$	$1.30{\times}10^{12}$ (1.8%)

Appendix

Table 17: Per-forward cost from the architecture specification. “Non-skip.” counts forwards the schedule can never skip (warm-up

+

forced tail).

Backbone	Skip	$\rho$	$S_{\mathrm{theo}}$	$S_{\mathrm{meas}}$	$S_{\mathrm{pred}}$	$n$
Wan2.1	50	0.50	1.935	1.877	1.879	8
	52	0.52	2.011	1.947	1.947	31
	54	0.54	2.092	2.021	2.021	19
	56	0.56	2.180	2.098	2.100	21
	58	0.58	2.276	2.184	2.186	28
	60	0.60	2.381	2.279	2.279	39

Appendix

Table 18: Analytic versus measured speedup at every skip count with at least eight prompts, pooled across all three schedulers (the cost model depends only on the skip count).

S_{\mathrm{pred}}

is re-predicted from the single fitted constant

f

, not fitted per row.

Level	Pair	Metric ( $n$ )	Lossless $\tilde{\Delta}$	Lossy $\tilde{\Delta}$	Same verdict?
Moderate	M vs. D	LPIPS (29)	$-0.02773$	$-0.02733$	✓
Moderate	M vs. D	PSNR (29)	$-3.55151$	$-3.48327$	✓
Moderate	M vs. D	SSIM (29)	$-0.02829$	$-0.02752$	✓
Moderate	S vs. D	LPIPS (31)	$-0.03436$	$-0.03353$	✓
Moderate	S vs. D	PSNR (31)	$-2.84926$	$-2.72569$	✓
Moderate	S vs. D	SSIM (31)	$-0.01354$	$-0.01244$	✓

Appendix

Table 19: Codec-independence check on Wan2.1 (

n{=}32

held-out prompts): median paired differences recomputed from raw decoder tensors (lossless) versus the H.264 pipeline (lossy). All

18

cells keep the same adjudication.

Configuration	Signal	$\tau$	$\lfloor\mathrm{ret}\cdot N\rfloor$	$\bar{s}$	LPIPS $\downarrow$	Worst	Speedup
Official default (M)	$\|\Delta y{-}\Delta x\|$	0.08	0	41	0.2496	0.4736	4.94 $\times$
Official default (D)	$\Delta y$	0.4	10	35	0.1680	0.3401	3.13 $\times$
Ours Moderate	S	4.6	5	34	0.0816	0.2048	2.95 $\times$
Ours Moderate	D	0.15	5	34	0.1353	0.3182	2.95 $\times$
Ours Moderate	M	0.042	5	34	0.1014	0.2681	2.95 $\times$
Ours Fast	S	6.4	5	37	0.1246	0.2487	3.57 $\times$

Appendix

Table 20: Operating points of the two released default FLUX baseline configurations on the same held-out set (

n{=}32

), with our matched-compute levels for reference. “Worst” is the maximum per-prompt LPIPS.

Adaptive Computation for Diffusion Models

See all 12 papers

May 24, 2026

Trajectory-Consistent Calibration for Cache-Accelerated Diffusion Models

Diffusion Transformers require repeated denoiser evaluations during iterative sampling, making inference computationally expensive. Cache-based acceleration reduces this cost by reusing intermediate representations across denoising steps, but can introduce representation deviations and degrade generation quality. In this paper, we analyze these deviations and show that effective calibration should consider both the direct mismatch caused by reuse and the subsequent trajectory shift induced by earlier corrections. To address this challenge, we propose Trajectory-Consistent Calibration (TCC), a training-free method that calibrates cached representations toward their full-computation counterparts. Specifically, rather than estimating all calibration priors from a single uncorrected cache trajectory, TCC uses an offline iterative procedure so that each prior accounts for the trajectory shift induced by preceding calibrations. Experiments on PixArt-alpha and DiT-XL/2 show that TCC consistently improves FID across representative cache-based acceleration methods while preserving their underlying reuse policies. Notably, in a representative PixArt-alpha cache-acceleration setting based on FORA, TCC reduces FID from 29.83 to 27.35, slightly surpassing the full-computation baseline.

Mingyu Liang, Dingkun Xu, Jingwei Xu

Laboratory for Novel Software Technology, Nanjing University, China

Diffusion Transformer Diffusion-Based Image Generation Diffusion Model Caching Diffusion Model Inference Acceleration

Ask PDF

Feb 26, 2026

Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache

Diffusion models have demonstrated remarkable success in image and video generation, yet their practical deployment remains hindered by the substantial computational overhead of multi-step iterative sampling. Among acceleration strategies, caching-based methods offer a training-free and effective solution by reusing or predicting features across timesteps. However, existing approaches rely on fixed or locally adaptive schedules without considering the global structure of the denoising trajectory, often leading to error accumulation and visual artifacts. To overcome this limitation, we propose DPCache, a novel training-free acceleration framework that formulates diffusion sampling acceleration as a global path planning problem. DPCache constructs a Path-Aware Cost Tensor from a small calibration set to quantify the path-dependent error of skipping timesteps conditioned on the preceding key timestep. Leveraging this tensor, DPCache employs dynamic programming to select an optimal sequence of key timesteps that minimizes the total path cost while preserving trajectory fidelity. During inference, the model performs full computations only at these key timesteps, while intermediate outputs are efficiently predicted using cached features. Extensive experiments on DiT, FLUX, and HunyuanVideo demonstrate that DPCache achieves strong acceleration with minimal quality loss, outperforming prior acceleration methods by

+

0.031 ImageReward at 4.87

\times

speedup and even surpassing the full-step baseline by

+

0.028 ImageReward at 3.54

\times

speedup on FLUX, validating the effectiveness of our path-aware global scheduling framework. Code is available at https://github.com/argsss/DPCache.

Bowen Cui, Yuanbin Wang, Huajiang Xu +6

Alibaba Group

Diffusion Model Inference Acceleration Diffusion Model Caching Diffusion Timestep Scheduling

Ask PDF

Mar 5, 2026

Frequency-Aware Error-Bounded Caching for Accelerating Diffusion Transformers

Diffusion Transformers (DiTs) have emerged as the dominant architecture for high-quality image and video generation, yet their iterative denoising process incurs substantial computational cost during inference. Existing caching methods accelerate DiTs by reusing intermediate computations across timesteps, but they share a common limitation: treating the denoising process as uniform across time,depth, and feature dimensions. In this work, we identify three orthogonal axes of non-uniformity in DiT denoising: (1) temporal -- sensitivity to caching errors varies dramatically across the denoising trajectory; (2) depth -- consecutive caching decisions lead to cascading approximation errors; and (3) feature -- different components of the hidden state exhibit heterogeneous temporal dynamics. Based on these observations, we propose SpectralCache, a unified caching framework comprising Timestep-Aware Dynamic Scheduling (TADS), Cumulative Error Budgets (CEB), and Frequency-Decomposed Caching (FDC). On FLUX.1-schnell at 512x512 resolution, SpectralCache achieves 2.46x speedup with LPIPS 0.217 and SSIM 0.727, outperforming TeaCache (2.12x, LPIPS 0.215, SSIM 0.734) by 16% in speed while maintaining comparable quality (LPIPS difference < 1%). Our approach is training-free, plug-and-play, and compatible with existing DiT architectures.

Guandong Li

iFLYTEK

Diffusion Model Caching Diffusion Model Inference Acceleration Diffusion Transformer Diffusion Timestep Scheduling

Ask PDF

Abstract

Adaptive Computation for Diffusion Models

Explore similar work

Trajectory-Consistent Calibration for Cache-Accelerated Diffusion Models

Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache

Frequency-Aware Error-Bounded Caching for Accelerating Diffusion Transformers