Diffusion Transformers generate high-quality images, but iterative sampling remains computationally expensive. Feature caching reduces the number of full network evaluations, yet its approximation error can become more severe under short sampling schedules, where fewer full-compute steps and a coarser sampling grid can make history-based extrapolation less reliable. In this paper, we propose CalibCache, a lightweight offline calibration method that corrects systematic discrepancies between cached and full-compute feature dynamics. Specifically, CalibCache calibrates feature acceleration, defined as the second-order finite difference across sampling steps. Using only a single pair of full-compute and naive-cache trajectories at each setting, CalibCache fits scalar acceleration maps offline by centered least squares. These maps are then applied during sampling without additional full-network evaluations or online optimization. Across FLUX.1-dev and Qwen-Image on DrawBench and DiT-XL/2 on ImageNet-256, CalibCache consistently improves fidelity or generation quality over TaylorSeer under 20-step sampling at cache intervals N=3,4,5. At N=5, it reduces FLUX.1-dev LPIPS from 0.563 to 0.429 while achieving a 2.86× denoising speedup over full-compute sampling, and lowers DiT-XL/2 FID from 16.39 to 10.37 relative to TaylorSeer.
Figures & tables
Figure 1: Five-column qualitative comparison on FLUX.1-dev with 20-step DrawBench sampling. Within each row, all methods use the same prompt and initial noise. The columns show Native, FORA ( N=3 , 2.59× ), TaylorSeer ( N=5 , 2.69× ), HiCache ( N=5 , 2.43× ), and CalibCache ( N=5 , 2.86× ); speedups refer to denoising latency relative to Native. FORA uses N=3 to show a broadly comparable speedup. The enlarged regions highlight differences in the ear and clock-hand details; in these examples, CalibCache more closely preserves the Native appearance.
Figure 2: Overview of CalibCache . (a) One full-compute/naive-cache trajectory pair with the same condition and initial noise provides acceleration observations for centered scalar fitting at the first cache step of each interval. (b) On the online corrected-cache rollout, the frozen scalars estimate full-compute acceleration; the difference from the online reference acceleration corrects the current cached residual without additional full-network evaluations.
Method
Efficiency
Alignment / preference
Full-compute fidelity
Latency (s) ↓
FLOPs (T) ↓
CLIP ↑
IR ↑
PSNR ↑
SSIM ↑
LPIPS ↓
FLUX.1-dev (20 steps)
7.0495 (1.00 × )
1487.836 (1.00 × )
31.3273
0.95212
∞
1
0
FORA ( N=3 )
2.7235 (2.59 × )
521.051 (2.86 × )
31.2644
0.7261
15.9921
0.64406
0.47070
TaylorSeer ( N=3,O=2 )
3.2887 (2.14 × )
595.719 (2.50 × )
31.5913
0.93172
16.7465
0.66782
0.41637
HiCache ( N=3,O=2 )
3.2052 (2.20 × )
595.539 (2.50 × )
31.5534
0.93172
17.0422
0.66129
0.41414
CalibCache ( N=3 )
3.2094 (2.20 × )
621.914 (2.39 × )
31.5376
0.95549
18.6202
0.72185
0.34093
Table 1: FLUX.1-dev on DrawBench-200 at 10242 with 20 denoising steps.
Method
Efficiency
Alignment / preference
Full-compute fidelity
Latency (s) ↓
FLOPs (T) ↓
CLIP ↑
IR ↑
PSNR ↑
SSIM ↑
LPIPS ↓
Qwen-Image (20 steps)
12.7757 (1.00 × )
2728.591 (1.00 × )
32.8989
1.17516
∞
1
0
FORA ( N=3 )
6.1979 (2.06 × )
955.599 (2.86 × )
31.0854
0.3168
13.9155
0.54600
0.53277
TaylorSeer ( N=3,O=2 )
7.4616 (1.71 × )
1092.537 (2.50 × )
32.1725
0.98366
13.4594
0.56454
0.49408
HiCache ( N=3,O=2 )
8.0308 (1.59 × )
1092.170 (2.50 × )
32.0616
0.98763
13.3377
0.56193
0.48487
CalibCache ( N=3 )
6.7595 (1.89 × )
1091.973 (2.50 × )
32.9019
1.10332
18.1881
0.70563
0.31579
Table 2: Qwen-Image on DrawBench-200 at 10242 with 20 denoising steps.
Method
Efficiency
Distributional evaluation
Latency (s) ↓
FLOPs (T) ↓
FID ↓
sFID ↓
IS ↑
Precision ↑
Recall ↑
DiT-XL/2 (DDIM-20)
0.2449 (1.00 × )
9.1535 (1.00 × )
3.4931
4.9673
223.1583
0.78447
0.57610
FORA ( N=3 )
0.1212 (2.02 × )
3.2167 (2.85 × )
14.2871
17.0764
144.4846
0.65272
0.47860
TaylorSeer ( N=3,O=2 )
0.1761 (1.39 × )
3.6804 (2.49 × )
5.3006
7.2121
196.4148
0.75718
0.55490
HiCache ( N=3,O=2 )
0.2220 (1.10 × )
3.6759 (2.49 × )
7.8976
10.8324
176.9213
0.72976
0.51910
CalibCache ( N=3 )
0.1662 (1.47 × )
3.6766 (2.49 × )
4.1941
6.1416
207.6349
0.75742
0.59240
Table 3: DiT-XL/2 on ImageNet-256 with DDIM-20 and 50K-image quality evaluation.
Supervision
FID ↓
sFID ↓
IS ↑
Precision ↑
Recall ↑
Residual supervision
41.80262
45.89258
64.36098
0.42992
0.42250
Velocity ( v )
37.27242
40.45567
72.52494
0.45670
0.43760
Acceleration ( a )
10.36915
6.40612
151.75644
0.66828
0.57380
Table 4: Correction-target ablation on DiT-XL/2, ImageNet-256, DDIM-20, N=5 , warmup-3. All variants share a two-anchor predictor, one calibration trajectory pair, centered scalar fitting, first-cache correction, corrected writeback, isolated history, and 50K-image evaluation. Only the target used to fit the scalar coefficients differs: residual, velocity, or acceleration.
FLUX.1-dev
Latency (s) ↓
FLOPs (T) ↓
CLIP ↑
IR ↑
PSNR ↑
SSIM ↑
LPIPS ↓
Uncorrected reuse
2.2425
446.683
31.5936
0.84588
18.7906
0.60969
0.45937
CalibCache
2.4644
466.550
31.9461
0.89891
19.1014
0.67348
0.42925
Qwen-Image
Latency (s) ↓
FLOPs (T) ↓
CLIP ↑
IR ↑
PSNR ↑
SSIM ↑
LPIPS ↓
Uncorrected reuse
5.4920
819.098
32.5109
0.91775
18.1857
0.65986
0.30701
CalibCache
5.6295
819.158
32.4704
1.00084
19.2460
0.71478
0.30412
DiT-XL/2
Latency (s) ↓
FLOPs (T) ↓
FID ↓
sFID ↓
IS ↑
Precision ↑
Recall ↑
Table 5: Matched-schedule comparison with uncorrected reuse at N=5 , warmup-3, and 20 steps. FLUX.1-dev and Qwen-Image use DrawBench-200 at 10242 ; DiT-XL/2 uses ImageNet-256, DDIM-20, and 50K images per row. Full/cache positions are matched; denoising latency and FLOPs reflect each implementation.
Application policy
FID ↓
sFID ↓
IS ↑
Precision ↑
Recall ↑
Correction without writeback
18.63283
13.06961
124.41991
0.59702
0.53530
Ordinary writeback
14.63578
9.01988
133.63596
0.62458
0.58280
Writeback + isolated history ( CalibCache )
10.36915
6.40612
151.75644
0.66828
0.57380
Table 6: Application-policy ablation on DiT-XL/2, ImageNet-256, DDIM-20, N=5 . All rows use the same scalar calibration pack, warmup-3 schedule, and 50K-image evaluation protocol.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Methods
N=3
N=4
N=5
FLUX / Qwen-Image
FORA
0
0
0
FLUX / Qwen-Image
TaylorSeer, HiCache, CalibCache
2
3
3
DiT-XL/2
FORA
0
0
0
DiT-XL/2
TaylorSeer, HiCache, CalibCache
2
3
3
Appendix
Table 7: Sampling warmup: the initial full-compute steps before periodic caching. Zero denotes no additional warmup phase. The ordinary cache schedule still supplies full-compute evaluations.
Model
N=3
N=4
N=5
FLUX.1-dev
36 (144 B)
24 (96 B)
24 (96 B)
Qwen-Image
24 (96 B)
16 (64 B)
16 (64 B)
DiT-XL/2
12 (48 B)
8 (32 B)
8 (32 B)
Appendix
Table 8: Active calibration coefficients per deployment configuration. Each entry gives the scalar count and its FP32 payload in bytes.
Supervision
Fit input x
Fit target y
Online αg
Residual
uτC
rτF−rτC
αgr
Velocity
uτC
rτF−rτ−1F
sgv
Acceleration
aτC=−uτC
aτF
1−sg
Appendix
Table 9: Supervision definitions for the common online update in Equation ( 6 ). The final column gives the coefficient multiplying f−p . Residual supervision targets the correction needed to match the full-trajectory residual. Velocity and acceleration supervision target the first and second residual differences along the full-compute trajectory.
Fitting form
FID ↓
sFID ↓
IS ↑
Precision ↑
Recall ↑
Centered s -only ( b=0 )
10.36915
6.40612
151.75644
0.66828
0.57380
Through-origin slope-only
14.18379
8.65522
136.35312
0.62440
0.58140
s+b
10.35668
6.33262
151.42812
0.66792
0.56760
Appendix
Table 10: Fitting-form ablation for DiT-XL/2 on ImageNet-256. Lower FID and sFID are better; higher IS, precision, and recall are better.
Coefficient setting
Fitted scalars
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
Fixed s=0
0
17.3736
0.59297
0.57690
31.8232
0.79039
Shared s=0.26131
1
18.4725
0.67029
0.43009
31.8310
0.87135
Grouped sg ( CalibCache )
24
19.1014
0.67348
0.42925
31.9461
0.89891
Appendix
Table 11: Coefficient-choice ablation on FLUX.1-dev, DrawBench-200, with 20 steps, N=5 , and warmup-3. The scalar count denotes independently fitted parameters; bold marks the best reported mean.
Calib. prompts
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
1
19.1014
0.67348
0.42925
31.9461
0.89891
4
19.2467
0.68151
0.41684
31.8786
0.90214
16
19.1494
0.67215
0.43108
31.8759
0.88410
64
19.1270
0.66995
0.43479
31.8941
0.88120
Appendix
Table 12: Calibration-set-size sensitivity for FLUX.1-dev with 20 steps, N=5 , warmup-3, and step-by-branch sharing on DrawBench-200. Each calibration prompt supplies one trajectory pair.
Metric
Mean ± std
Range
CLIP ↑
31.9113±0.0255
[31.8755,31.9534]
IR ↑
0.89312±0.01033
[0.87758,0.90675]
PSNR ↑
19.1151±0.0848
[18.9346,19.2567]
SSIM ↑
0.67070±0.00673
[0.65619,0.68096]
LPIPS ↓
0.43368±0.00956
[0.41778,0.45368]
Appendix
Table 13: Random single-prompt calibration on FLUX.1-dev, DrawBench-200, N=5 , warmup-3. Each of ten independently fitted step-by-branch packs uses the same calibration noise seed; evaluation prompts and seeds are shared. Mean, sample standard deviation, and range are across the ten pack-level aggregates.
Method
Traj.
Coeff.
Latency (s) ↓
FLOPs (T) ↓
FID ↓
sFID ↓
IS ↑
Precision ↑
Recall ↑
L 2 P (adapted)
64
151
0.0829
2.7461
39.9860
28.7022
79.2690
0.47538
0.54490
CalibCache
1
8
0.1354
2.7622
10.3691
6.4061
151.7564
0.66828
0.57380
Appendix
Table 14: Comparison with our L 2 P adaptation on DiT-XL/2, ImageNet-256, using DDIM-20, CFG 1.5 , N=5 , warmup-3, and 50 K generated images. Traj. counts calibration conditioning/noise realizations: L 2 P uses 64 full-compute trajectories, whereas CalibCache uses one full-compute/naive-cache trajectory pair. Coeff. counts active deployed coefficients.
Method
Latency (s) ↓
FLOPs (T) ↓
CLIP ↑
IR ↑
PSNR ↑
SSIM ↑
LPIPS ↓
TC-Padé ( Nth=0.7 )
5.8379
1116.023
31.6134
0.89883
11.7447
0.52002
0.62702
TC-Padé ( Nth=1.0 )
3.4730
669.747
31.5535
0.82747
11.8787
0.53759
0.65077
TC-Padé ( Nth=1.3 )
3.0231
520.987
31.4045
0.69882
11.9939
0.54075
0.66214
CalibCache ( N=3 )
3.2094
621.914
31.5376
0.95549
18.6202
0.72185
0.34093
CalibCache ( N=4 )
2.8173
544.223
31.6116
0.96672
19.8328
0.74463
0.28787
CalibCache ( N=5 )
2.4644
466.550
31.9461
0.89891
19.1014
0.67348
0.42925
Appendix
Table 15: Efficiency and quality comparison with our reproduction of the public TC-Padé implementation on FLUX.1-dev, DrawBench-200, with 20 denoising steps at 10242 . TC-Padé’s Nth is a curvature threshold, whereas CalibCache ’s N is a fixed cache interval.
Figure 3: FLUX.1-dev comparisons on DrawBench at 10242 with 20 denoising steps. Six prompts are shown in rows, with the exact prompt below each row. As in the introductory figure, FORA uses N=3 without extra warmup; TaylorSeer, HiCache, and CalibCache use N=5 and warmup-3. TaylorSeer and HiCache use O=2 , and HiCache uses scale 0.6 . These examples highlight reduced texture artifacts and closer preservation of the full-compute appearance and layout.
Figure 4: Qwen-Image comparisons on DrawBench at 10242 with 20 denoising steps. FORA uses N=3 without extra warmup; TaylorSeer, HiCache, and CalibCache use N=5 and warmup-3. TaylorSeer and HiCache use O=2 , and HiCache uses scale 0.6 . These examples illustrate clearer subjects and backgrounds and closer preservation of the full-compute scene and material appearance.
Figure 5: DiT-XL/2 comparisons on ImageNet-256 with DDIM-20 and CFG 1.5. Six examples are selected from 100 paired class–noise inputs, with the class and noise seed below each row. All methods start from the identical saved latent and class label for each example. FORA uses N=3 without extra warmup; TaylorSeer, HiCache, and CalibCache use N=5 and warmup-3. TaylorSeer and HiCache use O=2 , and HiCache uses scale 0.6 . These examples illustrate differences in texture and detail; aggregate quality is reported in Table 3 .
Method
Latency ↓
Speedup ↑
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
FLUX.1-dev (30 steps)
10.225
1.00 ×
∞
1.0000
0.0000
31.236
0.9937
FORA
2.347
4.36 ×
14.42
0.5801
0.5577
31.433
0.6679
TaylorSeer
3.530
2.90 ×
17.07
0.6864
0.3637
31.300
0.9526
HiCache
3.435
2.98 ×
17.50
0.7056
0.3365
31.314
0.9495
CalibCache (Ours)
3.154
3.24 ×
18.16
0.7060
0.3192
31.379
0.9689
Appendix
Table 16: FLUX.1-dev with 30 sampling steps on DrawBench-200. IR denotes ImageReward; latency is denoising seconds per image.
Method
Latency ↓
Speedup ↑
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
FLUX.1-dev (40 steps)
13.803
1.00 ×
∞
1.0000
0.0000
31.263
0.9844
FORA
3.162
4.36 ×
14.24
0.5993
0.5179
31.562
0.8264
TaylorSeer
4.564
3.02 ×
16.81
0.6935
0.3607
31.370
0.9896
HiCache
4.446
3.10 ×
17.15
0.7040
0.3371
31.361
0.9913
CalibCache (Ours)
4.029
3.43 ×
17.78
0.7068
0.3264
31.511
1.0034
Appendix
Table 17: FLUX.1-dev with 40 sampling steps on DrawBench-200. IR denotes ImageReward; latency is denoising seconds per image.
Method
Latency ↓
Speedup ↑
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
FLUX.1-dev (50 steps)
16.913
1.00 ×
∞
1.0000
0.0000
31.215
0.9799
FORA
3.886
4.35 ×
14.10
0.6048
0.4993
31.311
0.9097
TaylorSeer
5.490
3.08 ×
16.52
0.6908
0.3620
31.324
0.9902
HiCache
5.350
3.16 ×
16.86
0.7046
0.3392
31.277
1.0003
CalibCache (Ours)
4.781
3.54 ×
17.46
0.7047
0.3334
31.535
1.0029
Appendix
Table 18: FLUX.1-dev with 50 sampling steps on DrawBench-200. IR denotes ImageReward; latency is denoising seconds per image.
Method
Latency ↓
Speedup ↑
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
Qwen-Image (30 steps)
19.382
1.00 ×
∞
1.0000
0.0000
33.001
1.2189
FORA
6.960
2.78 ×
12.57
0.4858
0.5882
30.953
0.1195
TaylorSeer
9.196
2.11 ×
13.58
0.5817
0.4942
31.353
0.8149
HiCache
10.274
1.89 ×
13.64
0.5779
0.4734
31.648
0.8857
CalibCache (Ours)
8.837
2.19 ×
18.21
0.7122
0.3128
32.558
1.1231
Appendix
Table 19: Qwen-Image with 30 sampling steps on DrawBench-200. IR denotes ImageReward; latency is denoising seconds per image.
Method
Latency ↓
Speedup ↑
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
Qwen-Image (40 steps)
25.596
1.00 ×
∞
1.0000
0.0000
32.970
1.2359
FORA
9.333
2.74 ×
12.51
0.4906
0.5624
31.584
0.4950
TaylorSeer
11.951
2.14 ×
14.48
0.6356
0.4155
32.355
1.0860
HiCache
13.438
1.90 ×
14.53
0.6283
0.3977
32.407
1.0993
CalibCache (Ours)
11.556
2.21 ×
18.32
0.7370
0.2595
32.982
1.2094
Appendix
Table 20: Qwen-Image with 40 sampling steps on DrawBench-200. IR denotes ImageReward; latency is denoising seconds per image.
Method
Latency ↓
Speedup ↑
PSNR ↑
SSIM ↑
LPIPS ↓
CLIP ↑
IR ↑
Qwen-Image (50 steps)
31.723
1.00 ×
∞
1.0000
0.0000
32.928
1.2289
FORA
11.621
2.73 ×
12.75
0.5093
0.5307
32.188
0.7593
TaylorSeer
14.651
2.17 ×
14.65
0.6487
0.3973
32.505
1.1449
HiCache
16.373
1.94 ×
14.93
0.6495
0.3771
32.596
1.1638
CalibCache (Ours)
13.997
2.27 ×
18.18
0.7410
0.2663
32.895
1.1927
Appendix
Table 21: Qwen-Image with 50 sampling steps on DrawBench-200. IR denotes ImageReward; latency is denoising seconds per image.
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 Transformers (DiTs) have driven substantial progress in image and video generation but suffer from prohibitive computational costs. Feature caching accelerates inference by reusing intermediate representations. Existing methods rely on historical features for implementation simplicity, yet suffer from severe error accumulation at high acceleration ratios. To address this limitation, we investigate the nature of the requisite feature correction. We demonstrate that the optimal calibration update is characterized by a shared low-rank subspace across diverse prompts. Guided by this structural insight, we propose LearniBridge, a learnable calibration mechanism for feature caching that bridges multiple timesteps through lightweight LoRA updates. This mechanism enables effective calibration requiring only 3-5 training samples. Extensive experiments on image and video generation show that LearniBridge achieves up to 5.87×, 5.75×, and 4.10× acceleration on FLUX, HunyuanVideo, and WAN2.1, respectively. On WAN2.1, it improves VBench by 1.28% over the previous SOTA at 4.10× acceleration. Our code is available at https://github.com/Iiiiiiirene/LearniBridge.
Xuyue Huang, Zhe Chen, Wang Shen +1
Shenzhen Ubiquitous Data Enabling Key Laboratory, Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Central Media Technology Institute, Huawei Technologies Co., Ltd., Shenzhen, China
Diffusion Transformers (DiTs) can generate high-quality images and videos, but generating each sample requires multiple costly DiT forward passes. Two common ways to accelerate DiT sampling are step distillation, which reduces the number of sampling steps, and caching, which skips some DiT evaluations by reusing a tensor computed at an earlier step. Most caching methods decide in advance which tensor to reuse. After distillation, adjacent sampling steps are farther apart. Reusing a tensor across this larger gap introduces more error, so choosing what to cache becomes especially important. We therefore introduce AutoTarget, a method that chooses the cached tensor for a given model, solver, and reuse schedule. AutoTarget uses a small set of runs without cache reuse to measure the error caused by reusing each candidate tensor, then selects the candidate with the lowest error. We also analyze how an error at one reuse step affects the final sample. For Euler sampling, we identify cache targets that produce the same trajectory and show why a stored solver update may not. Experiments on distilled image and video DiTs show that the best cache target changes with the model, image resolution, and solver. AutoTarget reduces DiT evaluations and retained cache storage. Generation quality remains close to the corresponding uncached run. On the tested PixArt-LCM and FLUX.1-schnell settings, its calibration ranking matches the ranking from held-out cached runs. To help others reproduce the method, we provide its core implementation on GitHub at https://github.com/wali1024-offical/AutoTarget.
Shuo Yang, Lihao Fang, Yi Zhang +5
Beijing University of Chemical Technology · ROBOT-XZC · Tsinghua University