The Golden Path Hypothesis: Reusable Schedules in Diffusion Caching

Dong Wang; Wenwu Tang; Francesco Corti; Yun Cheng; Lothar Thiele; Olga Saukh

cs.AISep 30, 2026

The Golden Path Hypothesis: Reusable Schedules in Diffusion Caching

Authors: Dong Wang, Wenwu Tang, Francesco Corti, Yun Cheng, Lothar Thiele, Olga Saukh

Organizations: Graz University of Technology, Austria · Swiss Data Science Center, Switzerland · ETH Zurich, Switzerland · Complexity Science Hub, Austria

Abstract

Diffusion caching accelerates generation by replacing transformer computation with cached or predicted features at selected denoising steps. We introduce the Golden Path Hypothesis (GPH): under fixed inference conditions, prompt-independent cache schedules can achieve final-output quality comparable to the best prompt-specific schedules across prompts. We investigate the GPH across ten caching methods, four image and video models, and three cache ratios. Prompt-adaptive methods repeatedly select a small number of schedules, and reusing their most frequent schedules on new prompts closely matches the quality of prompt-specific choices. Exhaustive evaluation of 1.4 million schedules on four examples further identifies prompt-independent schedules that remain competitive on unseen prompts. To explain this transfer, we analyze denoising trajectories and the accumulation of caching errors. Latent-state trajectories exhibit similar structures across datasets and seeds, while an exact error decomposition shows that accumulated effects of earlier errors predict final latent-state error better than local approximation errors. This motivates searching for end-to-end schedules using final-output quality. With only a small set of examples, the resulting golden paths transfer across prompts and datasets, and can be tuned to the desired quality objective, including reconstruction fidelity or perceptual similarity.

Figures & tables

	Fixed $-$ adaptive
Method	$\Delta$ PSNR, dB $\uparrow$	$\Delta$ SSIM $\uparrow$	$\Delta$ LPIPS $\downarrow$
SeaCache	$-0.03\;[-0.07,+0.02]$	$-0.0011\;[-0.0025,+0.0034]$	$+0.0009\;[-0.0019,+0.0049]$
TeaCache	$\boldsymbol{+0.11}\;[-0.81,+1.56]$	$\boldsymbol{+0.0011}\;[-0.0333,+0.0268]$	$\boldsymbol{-0.0026}\;[-0.0390,+0.0333]$
SenCache	$-0.02\;[-0.26,+0.30]$	$+0.0008\;[-0.0182,+0.0122]$	$-0.0014\;[-0.0153,+0.0203]$
DiCache	$-0.06\;[-0.16,+0.18]$	$0.0000\;[-0.0062,+0.0047]$	$-0.0005\;[-0.0075,+0.0133]$

Table 1: The most frequent schedules retain quality on new prompts. The evaluation uses the 1,088 prompts excluded from schedule selection. For each metric, we subtract the value obtained with the adaptive method from the value obtained with the fixed schedule on the same prompt and seed. We average these differences within each model–ratio combination, using only runs where the two schedules differ, with 34 to 2,901 runs per combination. Each entry gives the median [minimum, maximum] of the six means from two image models and three target cache ratios. Arrows show which direction favors the fixed schedule. Bold marks the largest median

\Delta

PSNR and

\Delta

SSIM and the smallest median

\Delta

LPIPS. Table 20 in Appendix B.2 gives each combination.

Prompt index	Seed	Best PSNR	Shared-schedule PSNR	Gap
5	47	22.334	21.086	1.249
8	50	28.621	28.590	0.030
9	51	20.915	20.813	0.102
15	57	20.765	20.496	0.269

Table 2: The shared schedule is within 0.269 dB of the best schedule for three of the four selection runs. Each run uses FLUX.1-dev at cache ratio 0.82 with residual reuse. The best PSNR is the maximum over all 1,370,754 schedules with steps 0, 1, 2, and 49 fixed as full computations. The shared schedule maximizes mean PSNR across the four selection runs. The gap is the best PSNR minus the shared schedule’s PSNR, computed before rounding. All quality values are in dB.

Approximation policy	Feature error $\left\lVert r_{k}\right\rVert$	Propagation gain $h_{k}^{\rm LOO}$	Feature error $\times$ gain $\left\lVert r_{k}\right\rVert h_{k}^{\rm LOO}$	Step only $-k$
Residual reuse	0.528	0.803	0.886	0.867
Hermite order 2	0.586	0.800	0.896	0.880
Taylor order 1	0.676	0.788	0.899	0.880

Table 3: Scores that include propagation rank output errors more accurately than local feature error alone. Entries give Spearman correlations between each score and the L2 distance between the final latent states with and without caching, over 4,800 isolated cached steps per approximation policy.

\left\lVert r_{k}\right\rVert

is the feature-error norm, and

h_{k}^{\mathrm{LOO}}

is the propagation gain estimated from the other 99 prompts at the same step and with the same approximation policy. The step-only score is the negative cached-step index

-k

. Bold marks the highest correlation in each row.

Model	Cache ratio	Mean $\Delta$ PSNR: searched $-$ comparison schedule, dB
		MeanCache				BudCache				Random search
		DB	GE	PP	DDB	DB	GE	PP	DDB	DB	GE	PP	DDB
FLUX	0.58	+0.70	+0.66	+0.28	+0.37	+5.49	+6.21	+4.35	+3.26	+2.83	+2.94	+2.36	+1.99
FLUX	0.74	+0.49	+0.50	+0.36	+0.32	+1.14	+1.17	+0.91	+0.74	+0.57	+0.58	+0.57	+0.40
FLUX	0.82	+0.41	+0.58	+0.42	+0.44	+0.47	+0.70	+0.31	+0.12	+0.60	+0.83	+0.61	+0.46
Qwen	0.58	-0.02	-0.05	+0.04	+0.09	+1.19	+0.84	+1.08	+1.15	+3.95	+3.71	+3.93	+3.88

Table 4: Searched schedules achieve higher or comparable mean PSNR relative to all three reference schedules. For each prompt and seed, we subtract the reference schedule’s output PSNR from the searched schedule’s output PSNR. Entries give the mean differences, with positive values favoring the searched schedule. All schedules use residual reuse. The random-search schedule is selected from random search alone. FLUX denotes FLUX.1-dev and Qwen denotes Qwen-Image. DB, GE, PP, and DDB denote DrawBench, GenEval-style, PartiPrompts, and DiffusionDB-clean10k, respectively. We average over 600, 1,659, 4,896, and 30,000 prompt–seed runs, respectively. Confidence intervals and four additional metrics appear in Table 38 .

Model	Cache ratio	LPIPS-selected $-$ PSNR-selected
		PSNR range (dB)	SSIM range	LPIPS range
FLUX.1-dev	0.58	$-0.002$ to $+0.269$	$+0.0023$ to $+0.0144$	$-0.0224$ to $-0.0043$
FLUX.1-dev	0.74	$-0.451$ to $+0.080$	$+0.0009$ to $+0.0241$	$-0.0493$ to $-0.0025$
FLUX.1-dev	0.82	$-0.630$ to $-0.004$	$+0.0010$ to $+0.0410$	$-0.0992$ to $-0.0044$
Qwen-Image	0.58	$-0.883$ to $+0.329$	$+0.0007$ to $+0.0210$	$-0.0261$ to $-0.0074$
Qwen-Image	0.74	$-1.262$ to $-0.365$	$+0.0059$ to $+0.0206$	$-0.0454$ to $-0.0124$

Table 5: Optimizing LPIPS selects different schedules and improves the targeted metric. Each entry gives the range across eight comparisons, using hill climbing and annealing on four datasets. For each procedure and dataset, we subtract the mean PSNR, SSIM, or LPIPS of the PSNR-selected schedule’s outputs from the corresponding mean for the LPIPS-selected schedule’s outputs. Positive PSNR and SSIM differences and negative LPIPS differences favor the LPIPS-selected schedule. Both means use the same 600 DrawBench, 1,659 GenEval-style, 4,896 PartiPrompts, or 30,000 DiffusionDB-clean10k prompt–seed runs, with three seeds per prompt.

Appendix figures & tables35 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Schedule choice	Selection rule or objective	Approximation policy
SeaCache ( Chung et al., 2026 )	Adaptive	Spectral feature change	Residual reuse
TeaCache ( Liu et al., 2025a )	Adaptive	Calibrated feature change	Residual reuse
SenCache ( Haghighi and Alahi, 2026 )	Adaptive	Latent and timestep sensitivity	Residual reuse
DiCache ( Bu et al., 2025 )	Adaptive	First-block output change	Two-anchor prediction
TaylorSeer ( Liu et al., 2025b )	Fixed	Uniform spacing	First-order Taylor prediction
HiCache ( Feng et al., 2026 )	Fixed	Uniform spacing	Second-order Hermite prediction

Appendix

Table 6: The methods differ in both schedule selection and approximation policy. The last row describes the search experiments in Section 6 .

		$\rho=.58$ ( $K=29$ )					$\rho=.74$ ( $K=37$ )					$\rho=.82$ ( $K=41$ )
Data	Method	P $\uparrow$	S $\uparrow$	L $\downarrow$	C $\uparrow$	IR $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	C $\uparrow$	IR $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	C $\uparrow$	IR $\uparrow$
DB	Sea	27.208	.9066	.0845	27.619	.9983	21.435	.8006	.2114	27.627	.9897	19.407	.7365	.3078	27.621	.9537
DB	Tea	18.497	.7463	.2759	27.604	1.0032	16.868	.6880	.3628	27.657	.9765	15.458	.6307	.4590	27.531	.8569
DB	Sen	27.769	.9083	.0834	27.568	.9943	22.437	.7711	.2740	27.723	.8880	19.826	.6927	.3876	27.700	.8271
DB	Di	28.200	.9172	.0721	27.590	.9964	21.397	.7157	.3861	27.652	.7174	19.070	.6981	.4010	27.579	.7091
DB	Tay O1	21.442	.8250	.1713	27.521	.9812	17.571	.7194	.3017	27.516	.9893	14.275	.5709	.5221	27.204	.7305

Appendix

Table 7: Complete FLUX.1-dev results, averaged over three base seeds. P is PSNR in dB, S is SSIM, and L is LPIPS. C is CLIP score and IR is ImageReward.

\rho=K/50

is the target cache ratio. Bold marks the best unrounded mean within each dataset, ratio, and metric.

		$\rho=.58$ ( $K=29$ )					$\rho=.74$ ( $K=37$ )					$\rho=.82$ ( $K=41$ )
Data	Method	P $\uparrow$	S $\uparrow$	L $\downarrow$	C $\uparrow$	IR $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	C $\uparrow$	IR $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	C $\uparrow$	IR $\uparrow$
DB	Sea	27.437	.9192	.0765	29.226	1.2162	21.385	.8267	.1873	29.113	1.1725	16.778	.7007	.3526	29.002	1.1046
DB	Tea	20.363	.8218	.1796	29.217	1.2135	17.022	.7146	.3163	29.073	1.1344	14.457	.6162	.4615	28.793	.9185
DB	Sen	28.182	.9252	.0707	29.181	1.2186	21.825	.7634	.2641	28.768	1.0396	18.904	.7092	.3339	28.883	.9792
DB	Di	30.210	.9463	.0458	29.242	1.2233	20.468	.7283	.3225	28.722	.8573	17.205	.6562	.4084	28.381	.6534
DB	Tay O1	20.834	.8368	.1613	29.142	1.2188	13.984	.6163	.4399	28.220	.9413	9.680	.4308	.6760	26.485	.1029

Appendix

Table 8: Complete Qwen-Image results, averaged over three base seeds. P is PSNR in dB, S is SSIM, and L is LPIPS. C is CLIP score and IR is ImageReward.

\rho=K/50

is the target cache ratio. Bold marks the best unrounded mean within each dataset, ratio, and metric.

		$\rho=.58$ ( $K=29$ )					$\rho=.74$ ( $K=37$ )					$\rho=.82$ ( $K=41$ )
Data	Method	P $\uparrow$	S $\uparrow$	L $\downarrow$	$\Delta_{t}$	VB $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	$\Delta_{t}$	VB $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	$\Delta_{t}$	VB $\uparrow$
P599	Sea	29.36	.8954	.0725	-.00141	–	23.26	.7756	.1891	-.00101	–	19.40	.6771	.3171	-.00495	–
P599	Tea	23.18	.7800	.1726	+.00190	–	19.18	.6713	.2952	+.00531	–	17.80	.6217	.3760	+.00732	–
P599	Sen	30.60	.9093	.0630	-.00208	–	24.91	.7908	.1919	-.00335	–	21.65	.7058	.2877	+.00027	–
P599	Di	31.82	.9241	.0471	-.00042	–	22.24	.6910	.3632	-.00650	–	20.19	.6631	.3947	-.00340	–
P599	Bud	30.92	.9123	.0658	-.00281	–	25.07	.7837	.2133	-.00396	–	22.13	.7052	.3253	-.00386	–

Appendix

Table 9: Complete HunyuanVideo results, averaged over three base seeds. P is PSNR in dB, S is SSIM, and L is LPIPS.

\Delta_{t}

is temporal LPIPS for the cached video minus that for the full-compute video. Values closer to zero are better. VB is the VBench score, reported only on VB944.

\rho=K/50

is the target cache ratio. Bold marks the best unrounded mean within each dataset, ratio, and metric.

		$\rho=.58$ ( $K=29$ )					$\rho=.74$ ( $K=37$ )					$\rho=.82$ ( $K=41$ )
Data	Method	P $\uparrow$	S $\uparrow$	L $\downarrow$	$\Delta_{t}$	VB $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	$\Delta_{t}$	VB $\uparrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	$\Delta_{t}$	VB $\uparrow$
P599	Sea	26.96	.8732	.0885	-.00245	–	21.72	.7452	.2041	-.00094	–	18.31	.6408	.3235	+.00401	–
P599	Tea	23.32	.8080	.1376	-.00082	–	21.35	.7480	.2046	-.00191	–	18.76	.6414	.3516	+.00318	–
P599	Sen	26.57	.8733	.0889	-.00156	–	22.42	.7302	.2262	+.00173	–	19.61	.6382	.3332	+.01492	–
P599	Di	26.83	.8824	.0744	-.00078	–	20.43	.7319	.2129	-.00057	–	18.44	.6381	.3612	+.00029	–
P599	Bud	24.23	.8282	.1216	-.00182	–	22.86	.7673	.1897	-.00202	–	20.36	.6742	.2956	+.00479	–

Appendix

Table 10: Complete Wan2.1 results, averaged over three base seeds. P is PSNR in dB, S is SSIM, and L is LPIPS.

\Delta_{t}

is temporal LPIPS for the cached video minus that for the full-compute video. Values closer to zero are better. VB is the VBench score, reported only on VB944.

\rho=K/50

is the target cache ratio. Bold marks the best unrounded mean within each dataset, ratio, and metric.

Dataset	Method	$K_{\rm tgt}=29$	$K_{\rm tgt}=37$	$K_{\rm tgt}=41$
DrawBench	SeaCache	28.998	37.000	41.000
DrawBench	TeaCache	29.013	36.917	40.993
DrawBench	SenCache	28.718	37.000	40.793
DrawBench	DiCache	28.668	37.000	41.000
Parti	SeaCache	28.991	37.000	41.000
Parti	TeaCache	29.021	36.864	40.989

Appendix

Table 11: FLUX.1-dev mean realized cache counts for the four adaptive methods.

Dataset	Method	$K_{\rm tgt}=29$	$K_{\rm tgt}=37$	$K_{\rm tgt}=41$
DrawBench	SeaCache	29.003	37.003	41.000
DrawBench	TeaCache	29.000	37.195	40.875
DrawBench	SenCache	28.915	37.000	41.000
DrawBench	DiCache	28.932	37.000	41.000
Parti	SeaCache	29.002	37.002	41.000
Parti	TeaCache	28.999	37.175	40.894

Appendix

Table 12: Qwen-Image mean realized cache counts for the four adaptive methods.

Dataset	Method	$K_{\rm target}=29$	$K_{\rm target}=37$	$K_{\rm target}=41$
Penguin599	SeaCache	29.042	37.000	41.000
Penguin599	TeaCache	28.053	36.534	41.019
Penguin599	SenCache	28.959	36.000	40.853
Penguin599	DiCache	28.710	37.000	41.000
VBench944	SeaCache	29.019	37.000	41.000
VBench944	TeaCache	28.055	36.687	41.019

Appendix

Table 13: HunyuanVideo mean realized cache counts for the four adaptive methods.

Dataset	Method	$K_{\rm target}=29$	$K_{\rm target}=37$	$K_{\rm target}=41$
Penguin599	SeaCache	29.973	37.018	40.000
Penguin599	TeaCache	29.000	36.000	41.000
Penguin599	SenCache	28.975	37.000	41.010
Penguin599	DiCache	28.954	36.900	41.000
VBench944	SeaCache	29.956	37.007	40.000
VBench944	TeaCache	29.000	36.000	41.000

Appendix

Table 14: Wan2.1 mean realized cache counts for the four adaptive methods.

Method	$K=29$	$K=37$	$K=41$
SeaCache	4.81 / 2.12 $\times$	3.26 / 3.13 $\times$	2.51 / 4.06 $\times$
TeaCache	4.73 / 2.15 $\times$	3.16 / 3.22 $\times$	2.45 / 4.17 $\times$
SenCache	4.74 / 2.15 $\times$	3.19 / 3.20 $\times$	2.45 / 4.15 $\times$
DiCache	4.89 / 2.08 $\times$	3.34 / 3.05 $\times$	2.61 / 3.91 $\times$
TaylorSeer O1	5.42 / 1.88 $\times$	4.00 / 2.55 $\times$	3.31 / 3.08 $\times$
HiCache O2	5.86 / 1.74 $\times$	4.55 / 2.24 $\times$	3.88 / 2.63 $\times$

Appendix

Table 15: Generation is faster with caching on FLUX.1-dev. Columns specify the target cache count

K

. Each entry gives the mean time in seconds per image, followed by the speedup relative to full-compute generation. Full-compute generation takes 10.19 s per image. We average 12 H100 measurements after excluding four initial samples.

Method	$K=29$	$K=37$	$K=41$
SeaCache	14.79 / 2.26 $\times$	9.48 / 3.53 $\times$	6.84 / 4.89 $\times$
TeaCache	14.67 / 2.28 $\times$	9.30 / 3.60 $\times$	6.68 / 5.00 $\times$
SenCache	14.58 / 2.29 $\times$	9.36 / 3.57 $\times$	6.66 / 5.02 $\times$
DiCache	14.85 / 2.25 $\times$	9.73 / 3.44 $\times$	7.11 / 4.71 $\times$
TaylorSeer O1	17.62 / 1.90 $\times$	12.95 / 2.58 $\times$	10.59 / 3.16 $\times$
HiCache O2	19.76 / 1.69 $\times$	15.40 / 2.17 $\times$	13.13 / 2.55 $\times$

Appendix

Table 16: Generation is faster with caching on Qwen-Image. Columns specify the target cache count

K

. Each entry gives the mean time in seconds per image, followed by the speedup relative to full-compute generation. Full-compute generation takes 33.45 s per image. We average 12 H100 measurements after excluding four initial samples.

Method	$K=29$	$K=37$	$K=41$
SeaCache	47.64 / 2.17 $\times$	32.24 / 3.21 $\times$	24.50 / 4.22 $\times$
TeaCache	49.26 / 2.10 $\times$	32.99 / 3.14 $\times$	24.20 / 4.28 $\times$
SenCache	47.36 / 2.19 $\times$	33.78 / 3.06 $\times$	24.32 / 4.25 $\times$
DiCache	48.97 / 2.11 $\times$	33.17 / 3.12 $\times$	25.72 / 4.02 $\times$
TaylorSeer O1	49.68 / 2.08 $\times$	34.53 / 3.00 $\times$	26.98 / 3.84 $\times$
HiCache O2	50.83 / 2.04 $\times$	35.72 / 2.90 $\times$	28.14 / 3.68 $\times$

Appendix

Table 17: Generation is faster with caching on HunyuanVideo. Columns specify the target cache count

K

. Each entry gives the mean time in seconds per video, followed by the speedup relative to full-compute generation. Full-compute generation takes 103.49 s per video. H100 measurements include writing the video file. We exclude each worker’s first video.

Method	$K=29$	$K=37$	$K=41$
SeaCache	32.66 / 2.30 $\times$	22.64 / 3.31 $\times$	18.42 / 4.07 $\times$
TeaCache	33.74 / 2.22 $\times$	23.77 / 3.15 $\times$	16.75 / 4.48 $\times$
SenCache	33.76 / 2.22 $\times$	22.34 / 3.36 $\times$	16.66 / 4.50 $\times$
DiCache	35.38 / 2.12 $\times$	24.48 / 3.06 $\times$	18.91 / 3.96 $\times$
TaylorSeer O1	37.21 / 2.02 $\times$	26.29 / 2.85 $\times$	20.91 / 3.58 $\times$
HiCache O2	38.81 / 1.93 $\times$	27.93 / 2.68 $\times$	22.44 / 3.34 $\times$

Appendix

Table 18: Generation is faster with caching on Wan2.1. Columns specify the target cache count

K

. Each entry gives the mean time in seconds per video, followed by the speedup relative to full-compute generation. Full-compute generation takes 74.98 s per video. H100 measurements include writing the video file. We exclude each worker’s first video.

Method	Runs	Unique	Effective	Top-1	Top-3	Top-10
SeaCache	222,930	8.75 [2,18]	2.23 [1.04,3.77]	0.746	0.960	0.9996
TeaCache	222,930	18.25 [3,62]	5.56 [1.60,13.75]	0.456	0.778	0.9752
SenCache	222,930	10.46 [3,25]	3.99 [1.23,8.64]	0.561	0.857	0.9918
DiCache	222,930	16.46 [5,51]	5.55 [2.28,12.89]	0.426	0.777	0.9811

Appendix

Table 19: Schedules chosen by the four adaptive image methods. Within each model, dataset, and target cache ratio, we combine the three seeds and compute schedule frequencies. Top-

r

is the fraction of runs using one of the

r

most frequent schedules within that combination. The table averages these fractions equally over 24 combinations, each with 600 to 30,000 prompt–seed runs. Unique and effective counts are averaged over the same combinations, with their minimum and maximum in brackets. Runs gives the total across combinations.

Method	$K$	$m$ (%)	$n_{\rm off}$	$\overline{\Delta P}$	$n_{\rm off}^{\rm held}$	$\overline{\Delta P}^{\rm held}$	95% interval	Med. $\|\Delta P\|^{\rm held}$
FLUX.1-dev
SeaCache	29	85.7	699	-0.07	461	-0.07	$[-0.104,\,-0.046]$	0.06
TeaCache	29	38.5	3,009	+0.26	1,992	+0.23	$[0.142,\,0.329]$	0.02
SenCache	29	11.9	4,313	+0.05	2,901	+0.05	$[0.031,\,0.067]$	0.13
DiCache	29	18.7	3,979	-0.12	2,639	-0.12	$[-0.146,\,-0.088]$	0.13
SeaCache	37	68.7	1,534	+0.01	1,057	+0.01	$[0.009,\,0.019]$	0.03

Appendix

Table 20: Replacing adaptive schedules with one fixed schedule changes mean PSNR by

-0.81

+1.56

dB on new prompts where the schedules differ. We select the fixed schedule on 544 prompts and reuse it with the adaptive method’s approximation policy.

m

is the fraction of prompt–seed runs on which the adaptive method already used the fixed schedule.

n_{\rm off}

counts runs on which the adaptive method follows another schedule. For each run, we obtain

\Delta P

by subtracting the adaptive method’s output PSNR from the fixed schedule’s output PSNR on the same prompt and seed.

\overline{\Delta P}

is the mean of these differences over the

n_{\rm off}

runs. Each model–method–ratio combination uses 1,632 prompts under three base seeds, or 4,896 runs. The held-out columns use the 1,088 prompts not used for schedule selection. All PSNR differences are in dB. Positive differences favor the fixed schedule. The 95% interval belongs to

\overline{\Delta P}^{\rm held}

and is given to three decimals. It comes from resampling prompts and keeping each prompt’s runs together, using only runs where the schedules differ. Prompt means are weighted by these run counts. The final column is the median

|\Delta P|

over the

n_{\rm off}^{\rm held}

runs. Table 19 in Appendix B.1 reports a larger fraction for the same methods because it averages over four image datasets and counts the most frequent schedule whatever its cache count, while

m

counts one schedule that caches exactly

K

steps on PartiPrompts. The lower block replaces the fixed schedule with random schedules of the same

K

. For each random schedule, we retain its available held-out runs where the adaptive method’s schedule differs from the selected fixed schedule. Each retained run contributes one paired evaluation for that random schedule. We subtract the adaptive method’s output PSNR from each random schedule’s output PSNR. The median and minimum-to-maximum range summarize these differences across random schedules, models, and cache ratios.

Model	Method	$K$	Base seed A	Base seed B	Base seed C
FLUX	SeaCache	29	157/ $-0.08$ /0.06	157/ $-0.10$ /0.06	147/ $-0.04$ /0.06
FLUX	TeaCache	29	675/ $+0.26$ /0.02	654/ $+0.19$ /0.02	663/ $+0.25$ /0.02
FLUX	SenCache	29	967/ $+0.05$ /0.13	975/ $+0.04$ /0.13	959/ $+0.06$ /0.13
FLUX	DiCache	29	868/ $-0.11$ /0.12	886/ $-0.13$ /0.14	885/ $-0.11$ /0.13
FLUX	SeaCache	37	353/ $+0.01$ /0.03	356/ $+0.02$ /0.03	348/ $+0.01$ /0.03
FLUX	TeaCache	37	873/ $+0.91$ /0.07	884/ $+0.93$ /0.08	887/ $+0.93$ /0.06

Appendix

Table 21: Replacing adaptive schedules with one fixed schedule changes mean PSNR by

-0.83

+1.60

dB across base seeds on new prompts. We use only runs where the adaptive method chooses a different schedule. For each run, we subtract the adaptive method’s output PSNR from the fixed schedule’s output PSNR. Each entry gives the number of these runs, the mean PSNR difference, and the median absolute PSNR difference, in that order. Both differences are in dB. Each base-seed assignment uses 1,088 PartiPrompts. Base seeds A, B, and C are 41, 42, and 43 for FLUX.1-dev and 42, 100042, and 200042 for Qwen-Image.

Model	Method	Target $K$	Actual $K$	$m$ (%)	$n_{\rm off}$	$\overline{\Delta P}_{\rm off}$	Mean loss $\leq 0.25$ dB
HunyuanVideo	SeaCache	29	29	55	134	+0.03	yes
HunyuanVideo	TeaCache	29	29	2	294	+0.03	yes
HunyuanVideo	SenCache	29	29	26	223	-0.54	no
HunyuanVideo	DiCache	29	29	36	192	+0.05	yes
HunyuanVideo	SeaCache	37	37	76	71	+0.02	yes
HunyuanVideo	TeaCache	37	37	51	146	-0.03	yes

Appendix

Table 22: Reusing a fixed schedule loses at most 0.25 dB in mean PSNR relative to adaptive scheduling in 17 of 19 comparisons. In these comparisons, the fixed schedule has the requested cache count and differs from the adaptive schedule on at least one prompt. Each row uses 300 prompts, split evenly between Penguin599 and VBench944.

m

is the fraction of prompts on which the adaptive method already used the fixed schedule.

n_{\rm off}

counts prompts where it chooses a different schedule. For each such prompt, we subtract the adaptive method’s output PSNR from the fixed schedule’s output PSNR, using the same seed.

\overline{\Delta P}_{\rm off}

is the mean of these differences in dB. Positive values favor the fixed schedule. Target

K

is the requested cached-step count. Actual

K

is the count in the fixed schedule. The lower block contains schedules whose counts differ from the target. Mean differences are not applicable when

n_{\rm off}=0

Model	Method	Target $K$	Actual $K$	Penguin599	VBench944
HunyuanVideo	SeaCache	29	29	64/ $+0.02$ /0.04	70/ $+0.04$ /0.03
HunyuanVideo	TeaCache	29	29	146/ $0.00$ /0.01	148/ $+0.06$ /0.01
HunyuanVideo	SenCache	29	29	113/ $-0.53$ /0.51	110/ $-0.55$ /0.49
HunyuanVideo	DiCache	29	29	89/ $+0.05$ /0.20	103/ $+0.05$ /0.16
HunyuanVideo	SeaCache	37	37	37/ $+0.02$ /0.02	34/ $+0.02$ /0.02
HunyuanVideo	TeaCache	37	37	83/ $-0.03$ /0.02	63/ $-0.03$ /0.02

Appendix

Table 23: Replacing adaptive schedules with a fixed schedule causes the largest mean PSNR losses for SenCache on HunyuanVideo in both video datasets. We use only prompts where the schedules differ. For each such prompt, we subtract the adaptive method’s output PSNR from the fixed schedule’s output PSNR, using the same seed. Each entry gives the number of these prompts, the mean PSNR difference, and the median absolute PSNR difference, in that order. Both differences are in dB. Each dataset uses 150 prompts. Penguin599 uses base seed 54 and VBench944 uses base seed 42. Target

K

is the requested cached-step count. Actual

K

is the count in the fixed schedule. The lower block contains schedules whose counts differ from the target. A zero count means that the adaptive method uses the fixed schedule on every prompt, so the mean and median are not applicable.

Model	Cache ratio	Schedules	Fixed schedule	0.25 dB	0.5 dB	1.0 dB
FLUX.1-dev	0.58	32	MeanCache	.616 [.571]	.737 [.696]	.865 [.832]
FLUX.1-dev	0.74	31	MeanCache	.500 [.455]	.670 [.627]	.817 [.780]
FLUX.1-dev	0.82	15	BudCache	.556 [.515]	.671 [.631]	.806 [.772]
Qwen-Image	0.58	29	MeanCache	.449 [.405]	.559 [.514]	.703 [.661]
Qwen-Image	0.74	30	MeanCache	.438 [.393]	.549 [.504]	.727 [.686]
Qwen-Image	0.82	13	MeanCache	.589 [.549]	.667 [.628]	.800 [.765]

Appendix

Table 24: One fixed schedule stays within 0.5 dB of the per-prompt best on most prompts in every enlarged image candidate set. Each candidate set uses the same 1,088 prompts at base seed 42 and residual reuse for every schedule. “Schedules” gives the number of candidates. Within each candidate set, we select the schedule with the highest coverage at 0.25 dB under this common seed and report it at all three margins. Entries give the fraction of prompts within each PSNR margin of the per-prompt best. Brackets give the Bonferroni-adjusted one-sided 95% lower bound.

Model	$K$	Candidates	Fixed schedule	0.25 dB	0.5 dB	1.0 dB
FLUX.1-dev	29	17	MeanCache	0.694 [0.670]	0.776 [0.754]	0.869 [0.851]
FLUX.1-dev	37	17	MeanCache variant	0.443 [0.418]	0.571 [0.546]	0.715 [0.692]
FLUX.1-dev	41	10	BudCache	0.653 [0.629]	0.752 [0.729]	0.882 [0.865]
Qwen-Image	29	17	MeanCache	0.482 [0.456]	0.583 [0.558]	0.767 [0.744]
Qwen-Image	37	17	MeanCache	0.530 [0.505]	0.617 [0.592]	0.763 [0.741]
Qwen-Image	41	10	MeanCache	0.756 [0.734]	0.849 [0.830]	0.949 [0.937]

Appendix

Table 25: One shared schedule in each smaller candidate set stays within 0.5 dB of the per-prompt best on most prompts. These sets contain 9–17 schedules. For every model and cache ratio, the per-prompt best is the highest seed-averaged PSNR reached by any schedule in the candidate set, all of which use residual reuse. Each entry gives the share of prompts inside the margin, with the unadjusted one-sided 95% binomial lower bound in brackets. The reported counts of combinations whose lower bounds exceed 0.5 use Bonferroni-adjusted bounds. We report the candidate set member with the largest share at 0.25 dB. The reported MeanCache variant exchanges one cached step with one full step. Image rows use the 1,088 PartiPrompts prompts excluded from the 544-prompt recurring-schedule selection split. Video rows use 300 prompts split between Penguin599 and VBench944.

Model	$K$	Candidates	Fixed schedule	Tight	Middle	Wide
SSIM, margins 0.0075, 0.015, and 0.03
FLUX.1-dev	29	17	MeanCache	0.825 [0.805]	0.917 [0.902]	0.968 [0.958]
FLUX.1-dev	37	17	MeanCache variant	0.484 [0.459]	0.602 [0.577]	0.773 [0.751]
FLUX.1-dev	41	10	SeaCache most frequent	0.608 [0.583]	0.723 [0.700]	0.866 [0.848]
Qwen-Image	29	17	MeanCache	0.619 [0.595]	0.835 [0.816]	0.964 [0.953]
Qwen-Image	37	17	MeanCache	0.481 [0.455]	0.626 [0.601]	0.821 [0.801]

Appendix

Table 26: Shared schedules also approach prompt-specific choices under SSIM and LPIPS. Candidate sets and evaluation prompts are those of Table 25 . For each metric, we select the schedule with the highest coverage at the tightest margin and report it at all three margins. Cells show coverage and the unadjusted one-sided 95% lower bound in brackets. Each prompt’s reference is its best seed-averaged score among the candidates. All schedules use residual reuse. The MeanCache variant exchanges one cached step with one full step. LPIPS is negated for the test, while margins are reported as positive LPIPS differences.

Selection rule	Full-compute steps	Mean PSNR	Minimum PSNR
Highest mean	0, 1, 2, 4, 6, 11, 24, 41, 49	22.746	20.496
Highest minimum	0, 1, 2, 4, 6, 13, 23, 40, 49	22.672	20.739
MeanCache schedule	0, 1, 2, 3, 4, 9, 19, 34, 49	21.619	18.041
BudCache schedule	0, 1, 2, 4, 7, 13, 20, 39, 49	22.235	19.190

Appendix

Table 27: Schedules selected from four calibration examples. The four boundary steps are required for every schedule. PSNR values are in dB.

Dataset	Prompts	Runs	Schedules evaluated
DrawBench	200	600	337 candidates and method comparisons
GenEval-style	553	1,659	337 candidates and method comparisons
PartiPrompts	1,632	4,892	337 candidates and method comparisons
DiffusionDB-clean10k	10,000	30,000	selected schedules and methods
Total		37,151

Appendix

Table 28: Evaluation data for the high-ratio search. Each dataset uses three base seeds.

Scope	Method or schedule	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Three datasets	selected by four-run mean	20.944	0.7638	0.2880
Three datasets	selected by four-run minimum	20.782	0.7608	0.2890
Three datasets	MeanCache schedule	20.766
Three datasets	BudCache schedule	20.742	0.7583	0.2922
Three datasets	SeaCache	19.916	0.7551	0.2914
DiffusionDB	selected by four-run mean	20.485	0.6981	0.3963

Appendix

Table 29: The schedule selected by mean PSNR transfers beyond its four scoring examples. All fixed schedules use residual reuse. SeaCache uses its original adaptive schedule and residual-reuse rule. The three-dataset results exclude the four selection runs. MeanCache’s SSIM and LPIPS are reported on the full 7,155-run scope in Appendix B.5.4 . Its DiffusionDB row uses all 30,000 runs. PSNR is in dB. Bold marks the best value within each scope.

Scope	Approximation policy	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Three datasets	residual reuse	20.945	0.7638	0.2880
Three datasets	two-anchor	21.193	0.7865	0.2392
DiffusionDB	residual reuse	20.485	0.6981	0.3963
DiffusionDB	two-anchor	20.929	0.7316	0.3192

Appendix

Table 30: Two-anchor prediction improves all three quality metrics on the schedule selected by four-run mean. Three-dataset values equally weight the nine dataset–seed means. DiffusionDB values average its three base-seed means. The four selection runs are included. PSNR is in dB. Bold marks the better approximation policy within each scope.

Approximation policy	Step only $-k$	Feature error $\left\lVert r_{k}\right\rVert$	Feature error $\times$ step size $\left\lVert r_{k}\right\rVert\|\Delta\tau_{k}\|$	Next-state error estimate	Propagation gain $h_{k}^{\rm LOO}$	Feature error $\times$ gain $\left\lVert r_{k}\right\rVert h_{k}^{\rm LOO}$
Residual reuse	0.867	0.528	-0.382	-0.215	0.803	0.886
Hermite order 2	0.880	0.586	-0.391	-0.180	0.800	0.896
Taylor order 1	0.880	0.676	-0.350	-0.115	0.788	0.899

Appendix

Table 31: Combining feature error with propagation gain gives the highest correlation with output error magnitude. We compute Spearman correlations over 100 prompts and steps 1–48, giving 4,800 isolated cached steps per approximation policy. The step-only score is the negative cached-step index

-k

. The propagation gain

h_{k}^{\mathrm{LOO}}

uses the other 99 prompts at the same step and with the same approximation policy. The next-state error estimate is

\left\lVert r_{k}\right\rVert|\Delta\tau_{k}|

multiplied by the first mean ratio in Equation 17 .

Model	$K$	Schedule	Approximation policy	Interaction	Mean PSNR (dB)
FLUX.1-dev	29	.950	.037	.013	25.20
FLUX.1-dev	37	.613	.117	.270	20.95
FLUX.1-dev	41	.493	.269	.238	18.53
Qwen-Image	29	.983	.012	.005	26.78
Qwen-Image	37	.657	.181	.162	20.26
Qwen-Image	41	.368	.499	.133	15.79

Appendix

Table 32: Interactions between schedules and approximation policies account for 13–27% of the variation in mean PSNR at cache ratios 0.74 and 0.82. For each model and cache ratio, we fit an additive model to the 20 mean PSNR values from four schedules and five approximation policies, with equal weight for every combination. The schedule and approximation policy columns give the fractions of the total sum of squares explained by their separate effects. The interaction column gives the remaining fraction. The last column is the average of the 20 mean PSNR values, in dB.

Model	$K$	PSNR range across approximation policies (dB)	Schedule difference (dB)
FLUX.1-dev	29	1.457	-3.238
FLUX.1-dev	37	1.502	+1.002
FLUX.1-dev	41	2.897	+0.708
Qwen-Image	29	1.683	+1.870
Qwen-Image	37	5.319	+2.727
Qwen-Image	41	6.225	+2.431

Appendix

Table 33: BudCache and SeaCache differ by up to 3.24 dB on images and 3.01 dB on video under residual reuse. “PSNR range across approximation policies” is the difference between the highest and lowest mean PSNR of TaylorSeer, HiCache, and L2P, which share a schedule. “Schedule difference” subtracts adaptive SeaCache’s mean PSNR from the mean PSNR obtained with the fixed BudCache schedule. Both use residual reuse at the same target ratio, as in Figures 12 and 13 . We give equal weight to the mean PSNR of each dataset, using four image datasets or two video datasets. Positive schedule differences favor BudCache.

Prompt category	Source	Base seed
person	COCO 2014 val	50042
animal	COCO 2014 val	50042
indoor object	COCO 2014 val	50042
outdoor scene	COCO 2014 val	50042
stylized A	DiffusionDB calibration prompt set	60042
stylized B	DiffusionDB calibration prompt set	60042

Appendix

Table 34: The eight scoring examples. Base seeds 50042 and 60042 differ from every evaluation base seed.

Model	$K$	PSNR range	PSNR sd	Median swap delta	Median SE	$t_{\max}$	$t_{\min}$
FLUX.1-dev	29	15.45–22.37	1.430	0.0236	0.625	0.24	0.00024
FLUX.1-dev	37	14.79–18.92	1.103	0.0433	0.589	0.43	0.00043
FLUX.1-dev	41	14.64–18.38	0.921	0.0536	0.537	0.54	0.00054
Qwen-Image	29	14.07–21.30	1.603	0.0174	0.693	0.17	0.00017
Qwen-Image	37	12.93–19.48	1.552	0.0347	0.542	0.35	0.00035
Qwen-Image	41	11.73–17.23	1.427	0.0766	0.585	0.77	0.00077

Appendix

Table 35: Initial score differences determine the annealing temperatures. The PSNR range and standard deviation are over the 50 initial schedules. The swap delta is the absolute change in mean PSNR on the eight scoring runs between schedules that differ by one swap. Median SE is the median standard error across the 50 schedules, with each standard error computed from that schedule’s eight scores. All values are in dB.

Model	$K$	Validation PSNR (dB)	Full steps
FLUX.1-dev	29	30.705	0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 24, 30, 38, 46, 49
FLUX.1-dev	37	24.832	0, 1, 2, 3, 4, 6, 8, 10, 14, 19, 30, 43, 49
FLUX.1-dev	41	21.552	0, 1, 2, 3, 5, 8, 13, 27, 49
Qwen-Image	29	30.381	0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16, 19, 22, 26, 33, 40, 46, 49
Qwen-Image	37	24.499	0, 1, 2, 3, 4, 5, 6, 8, 11, 14, 22, 35, 49
Qwen-Image	41	20.766	0, 1, 2, 3, 5, 9, 17, 31, 49

Appendix

Table 36: The selected schedule for each model–ratio combination. Full steps are the positions that run the full model. Every other step of the 50 is cached. Validation PSNR is the mean over the 50 validation prompts.

Model, cache ratio	Procedure	Selected PSNR	Replaced PSNR	$\Delta$ PSNR (dB)	$\Delta$ SSIM	$\Delta$ LPIPS
FLUX.1-dev, ratio 0.58	hill	30.705	30.418	+0.072 [+0.052, +0.091]	+0.0010	-0.0005
FLUX.1-dev, ratio 0.74	anneal	24.663	24.654	-0.002 [-0.011, +0.006]	+0.0022	-0.0008
FLUX.1-dev, ratio 0.74	greedy	24.499	24.476	+0.221 [+0.182, +0.255]	+0.0011	+0.0019
FLUX.1-dev, ratio 0.74	hill	24.832	24.825	-0.012 [-0.020, -0.006]	+0.0006	-0.0008
Qwen-Image, ratio 0.58	anneal	30.381	30.201	+0.149 [+0.128, +0.170]	+0.0011	-0.0015
Qwen-Image, ratio 0.82	anneal	20.739	20.662	+0.001 [-0.014, +0.019]	-0.0002	+0.0005

Appendix

Table 37: Validation changes the selected candidate in eight search runs. The replaced candidate has the highest mean PSNR on the eight scoring runs. The selected candidate has the highest mean PSNR on 50 validation prompts. The first two numeric columns give the candidates’ mean validation PSNR. For each metric, we subtract the value for the replaced candidate’s output from the value for the selected candidate’s output on the same prompt and seed. The last three columns report the means of these differences over 4,896 PartiPrompts prompt–seed runs. Positive PSNR and SSIM differences and negative LPIPS differences favor the selected candidate. All PSNR values are in dB. Brackets give a 95% bootstrap interval for the PSNR difference. Hill, anneal, and greedy denote hill climbing, annealing, and greedy coordinate ascent.

Model, cache ratio	Dataset	Reference	Runs	$\Delta$ PSNR (dB)	$\Delta$ SSIM	$\Delta$ LPIPS	$\Delta$ IR	$\Delta$ CLIP
FLUX.1-dev, ratio 0.58	DrawBench	MeanCache	600	+0.697 [0.55, 0.85]	-0.0062	+0.0114	-0.012	-0.013
	DrawBench	BudCache	600	+5.489 [5.18, 5.80]	+0.0558	-0.0539	-0.017	-0.058
FLUX.1-dev, ratio 0.58	GenEval-style	MeanCache	1,659	+0.661 [0.56, 0.77]	-0.0049	+0.0084	-0.011	+0.108
	GenEval-style	BudCache	1,659	+6.214 [6.03, 6.40]	+0.0517	-0.0535	-0.014	+0.121
FLUX.1-dev, ratio 0.58	PartiPrompts	MeanCache	4,896	+0.277 [0.22, 0.33]	-0.0124	+0.0191	-0.016	+0.062
	PartiPrompts	BudCache	4,896	+4.351 [4.25, 4.45]	+0.0375	-0.0339	-0.015	+0.075

Appendix

Table 38 : Searched schedules improve PSNR over BudCache in every reported model, cache ratio, and dataset combination. Each row compares the selected schedule with one reference schedule under residual reuse. For each metric, we subtract the value for the reference schedule’s output from the value for the searched schedule’s output, using the same prompt and seed. Entries are means of these differences over the listed prompt–seed runs. Positive PSNR, SSIM, ImageReward, and CLIP differences and negative LPIPS differences favor the searched schedule. IR denotes ImageReward. PSNR differences are in dB. Brackets give 95% bootstrap intervals obtained by resampling prompt–seed runs individually.

Model	Cache ratio	Reuse	Taylor O1	Hermite O2	Interval-average	Two-anchor
FLUX.1-dev	0.58	29.104	29.931	30.257	30.342	30.775
FLUX.1-dev	0.74	24.001	24.013	24.513	24.542	24.990
FLUX.1-dev	0.82	21.060	19.459	20.334	20.325	21.131
Qwen-Image	0.58	31.017	31.812	31.981	32.205	32.450
Qwen-Image	0.74	25.183	24.612	25.294	24.549	25.637
Qwen-Image	0.82	21.471	20.473	21.414	20.346	21.628

Appendix

Table 39: Two-anchor prediction gives the highest mean PSNR on all six searched schedules. Each row holds the PSNR-selected schedule fixed. Mean PSNR is in dB and gives equal weight to the four evaluation datasets, totaling 37,155 prompt–seed runs per row and approximation policy. Bold marks the highest mean in each row. Interval-average denotes MeanCache’s interval-average velocity prediction. Two-anchor uses the first transformer block to extrapolate from two full-step residuals.

Search algorithm	50	200	1,000	5,000	Exact by 1,000	To 0.05 dB
Random sampling	1.19	0.74	0.38	0.18	0/50	–
Hill climb, first improvement	0.75	0.18	0.010	0.000	21/50	324
Hill climb, steepest ascent	1.82	1.49	0.055	0.0003	9/50	802
Hill climb, seeded	0.57	0.16	0.005	0.000	25/50	308
Annealing	1.02	0.08	0.000	0.000	50/50	171
Annealing, seeded	0.65	0.05	0.000	0.000	50/50	172

Appendix

Table 40: Ten search algorithms evaluated using the stored scores of all 1,370,754 FLUX.1-dev schedules that cache 41 of 50 steps. The columns 50 to 5,000 give the number of candidate evaluations. Gaps are PSNR differences in dB below the best exhaustive score. Each gap is averaged over 50 searches. 0.000 means the algorithm found that score exactly. “Exact by 1,000” counts how many of the 50 searches found the best exhaustive score within 1,000 candidate evaluations. The last column is the median evaluation count among repetitions that reached a gap of 0.05 dB. A dash indicates that fewer than 25 of the 50 repetitions reached this gap within 5,000 evaluations. The annealing implementation follows BudCache ( Lei et al., 2026 ) , with temperatures scaled to the observed score differences.

Explore similar work

ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE
Jun 4, 2026Mishan Aliev, Eva Neudachina, Ilya Bykov +4Diffusion Models Iterative Denoising Methods
1HSE University, Russia · 2Yandex Research, Russia
OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models
Jun 30, 2026Huanlin Gao, Fang Zhao, Qiang Hui +8Diffusion Models Diffusion Sampling
Data Science & Artificial Intelligence Research Institute, China Unicom · Unicom Data Intelligence, China Unicom · National Key Laboratory for Novel Software Technology, Nanjing University
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

Jun 4, 2026, cs.CVJ/K move · Enter open · S save

ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE

Mishan Aliev, Eva Neudachina, Ilya Bykov, Aleksandr Oganov+3

1HSE University, Russia · 2Yandex Research, Russia

Diffusion Models Iterative Denoising Methods Recast

Open paper AskPDF arXiv

Modern diffusion models generate high-quality images and videos, but their iterative denoising process makes inference expensive. Feature caching accelerates sampling by reusing or predicting intermediate activations across neighboring denoising steps, exploiting the redundancy of computations along the reverse trajectory. In this work, we focus on the caching schedule: selecting which denoising steps should be fully recomputed. Existing schedules are either fixed (e.g. uniform) or chosen adaptively from per-step error heuristics; in both cases, the actual compute cost is a side-effect of hand-tuned thresholds rather than a quantity the user can specify. We propose ReCache, which inverts this: given a target budget k, it learns the recomputation schedule that maximizes generation quality, turning compute into a directly controllable input. ReCache trains via policy gradients, sidestepping backpropagation through full diffusion inference, and uses no labelled data. Generations from uncached inference serve as matching targets, paired with a reward for generation quality. ReCache is compatible with any caching mechanism, including feature reuse and feature forecasting; for each mechanism, a single trained policy adapts across computational budgets at inference time. ReCache consistently outperforms scheduling baselines: under a

\times5.04

FLOPs reduction on FLUX, it reduces LPIPS by 31% (from 0.456 to 0.316) compared to DiCache; on Wan 2.1 at a

\sim \times2.6

speedup, it drops LPIPS by 65% (from 0.480 to 0.169) and boosts the VBench score by 7% (5.6 points, from 70.4 to 76.0) over uniform HiCache. Code is available at https://github.com/thecrazymage/ReCache.

Abstract

Explore similar work

ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE

OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models

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