Feature caching approaches accelerate diffusion transformers (DiTs) by storing the output features of computationally expensive modules at certain timesteps, and exploiting them for subsequent steps to reduce redundant computations. Recent forecasting-based caching approaches employ temporal extrapolation techniques to approximate the output features with cached ones. Although effective, relying exclusively on temporal extrapolation still suffers from significant prediction errors, leading to performance degradation. Through a detailed analysis, we find that 1) these errors stem from the irregular magnitude of changes in the output features, and 2) an input feature of a module is strongly correlated with the corresponding output. Based on this, we propose relational feature caching (RFC), a novel framework that leverages the input-output relationship to enhance the accuracy of the feature prediction. Specifically, we introduce relational feature estimation (RFE) to estimate the magnitude of changes in the output features from the inputs, enabling more accurate feature predictions. We also present relational cache scheduling (RCS), which estimates the prediction errors using the input features and performs full computations only when the errors are expected to be substantial. Extensive experiments across various DiT models demonstrate that RFC consistently outperforms prior approaches significantly. Project page is available at https://cvlab.yonsei.ac.kr/projects/RFC
Figures & tables
Figure 1: Feature analysis and comparison between existing approaches (FORA ( Selvaraju et al., 2024 ) , TaylorSeer ( Liu et al., 2025b ) ), and our method (RFC) using DiT-XL/2 ( Peebles and Xie, 2023 ) . (a-b) Min-max normalized L2 distances of output and input features, measured between consecutive timesteps. While the variations of feature changes are irregular, those of input and output remain closely aligned with each other. (c) The prediction errors across different modules. We measure the relative L1 error between output features with and without applying caching methods and average the values over the timesteps. (d) Quantitative results on ImageNet ( Deng et al., 2009 ) evaluated in terms of FLOPs and sFID ( Nash et al., 2021 ) .
Figure 2: Empirical analyses in DiT-XL/2 ( Peebles and Xie, 2023 ) . (a) RSD of sk(t−k) with varying t . (b) Relative L1 errors of the output and input features, i.e ., EO(t) and EI(t) , respectively. Please see the text for details.
Methods
NFC
FLOPs (T) ↓
Latency (s) ↓
FID ↓
sFID ↓
FID2FC ↓
sFID2FC ↓
IS ↑
Full-Compute
50
23.74
7.61
2.32
4.32
-
-
241.25
FORA ( N=4 ) ( Selvaraju et al., 2024 )
14
6.66
2.37
4.33
7.38
1.90
6.32
215.33
ToCa ( N=4 ) ( Zou et al., 2025 )
17
8.73
3.64
3.03
5.02
0.86
3.56
229.01
DuCa ( N=4 ) ( Zou et al., 2024 )
17
7.66
2.94
3.39
4.95
1.19
4.22
224.03
TaylorSeer ( N=4,m=2 ) ( Liu et al., 2025b )
14
6.66
2.84
2.55
5.30
0.44
2.17
232.26
RFC ( m=2 )
14.01
6.67
2.86
2.52
4.60
0.30
1.33
231.00
Table 1: Quantitative comparisons on class-conditional image generation for DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . Numbers in bold indicate the best performance, while underscored ones are the second best among similar computational costs. N is the interval between full computations, and m refers to the order of the Taylor expansion.
Methods
NFC
FLOPs (T) ↓
Latency (s) ↓
PSNR ↑
SSIM ↑
LPIPS ↓
IR ↑
CLIP ↑
Full-Compute
50
2813.50
26.210
-
-
-
0.9655
17.0720
FORA ( N=3 ) ( Selvaraju et al., 2024 )
17
957.33
10.489
17.4967
0.7236
0.4209
0.9111
16.9515
ToCa ( N=5 ) ( Zou et al., 2025 )
14
1056.26
18.591
18.0333
0.7269
0.3490
0.9457
17.0063
TaylorSeer ( N=4,m=1 ) ( Liu et al., 2025b )
14
788.59
8.863
19.5722
0.7695
0.3226
0.9204
17.0659
TaylorSeer ( N=4,m=2 )
14
788.59
8.985
19.7733
0.7706
0.3175
0.9407
17.0527
RFC ( m=1 )
14.02
789.82
9.148
20.2707
0.7905
0.2934
0.9448
17.0288
Table 2: Quantitative comparisons on text-to-image generation for FLUX.1 dev ( Batifol et al., 2025 ) on DrawBench ( Saharia et al., 2022 ) . Numbers in bold indicate the best performance, while underscored ones are the second best among similar computational costs. N is the interval between full computations, and m refers to the order of the Taylor expansion.
Methods
NFC
FLOPs (T) ↓
Latency (s) ↓
PSNR ↑
SSIM ↑
LPIPS ↓
VBench ↑
Full-Compute
50
7520.00
263.480
-
-
-
81.40
FORA ( N=6 ) ( Selvaraju et al., 2024 )
9
1359.19
55.339
15.5221
0.4349
0.2743
78.51
TaylorSeer ( N=5,m=1 ) ( Liu et al., 2025b )
10
1510.21
60.051
16.3606
0.5126
0.1985
80.77
TaylorSeer ( N=6,m=1 )
9
1359.19
55.872
15.5321
0.4606
0.2448
79.52
RFC ( m=1 )
8.96
1354.65
57.786
18.5408
0.6352
0.1329
80.83
FORA ( N=8 ) ( Selvaraju et al., 2024 )
7
1058.45
48.340
14.8878
0.4114
0.2959
77.50
Table 3: Quantitative comparisons on text-to-video generation for HunyuanVideo ( Kong et al., 2024 ) on VBench ( Huang et al., 2024 ) . Numbers in bold indicate the best performance among similar computational costs. N is the interval between full computations, and m refers to the order of the Taylor expansion.
Figure 3: Qualitative comparisons of (left) class-conditional image generation for DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) , and (right) text-to-image generation for FLUX.1 dev ( Batifol et al., 2025 ) on DrawBench ( Saharia et al., 2022 ) .
Figure 4: Qualitative comparisons of text-to-video generation for HunyuanVideo ( Kong et al., 2024 ) on VBench ( Huang et al., 2024 ) . Please see the supplementary material for the actual video clips.
Methods
NFC
FID ↓
sFID ↓
FID2FC ↓
sFID2FC ↓
TaylorSeer ( Liu et al., 2025b )
14
2.65
5.60
0.57
2.77
RFE
14
2.52
5.18
0.43
2.02
RCS
14
2.52
4.76
0.36
1.88
RFC (RFE + RCS)
14
2.51
4.66
0.31
1.41
TaylorSeer ( Liu et al., 2025b )
11
2.87
5.85
0.73
3.53
RFE
11
2.69
5.22
0.55
2.55
Table 4: Quantitative comparison of different components for DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . All results are obtained using the first-order approximation ( m=1 ).
Methods
NFC
FID2FC ↓
sFID2FC ↓
Latency (s) ↓
Linear
14
0.73
3.40
2.85
w(t)=0.8
14
0.73
3.36
2.84
w(t)=1.0
14
0.57
2.77
2.84
w(t)=1.2
14
0.52
2.51
2.84
RFE
14
0.43
2.02
2.86
Linear
11
0.93
4.08
2.52
Table 5: Quantitative comparison of different strategies for estimating the change in output features for DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . All results are obtained using the first-order approximation ( m=1 ).
Methods
NFC
FID2FC ↓
sFID2FC ↓
Latency (s) ↓
Input distance
14.03
0.50
2.53
2.62
All modules
14.02
0.37
1.85
2.79
RCS
14.00
0.36
1.88
2.63
Input distance
11.15
0.72
3.47
2.23
All modules
11.00
0.61
3.03
2.41
RCS
11.01
0.62
3.09
2.26
Table 6: Quantitative comparison of different strategies for the cache scheduling using DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . All results are obtained using the first-order approximation ( m=1 ).
Figure 5: Feature analyses using DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . (a) A linearity analysis of diffusion features. The solid lines present the linearity between the input and output features, while the dashed lines show the linearity between the timesteps and output features. (b-c) The directional consistency of the input and output features.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Analyses using FLUX.1 dev ( Batifol et al., 2025 ) on DrawBench ( Saharia et al., 2022 ) (upper row) and HunyuanVideo ( Kong et al., 2024 ) on VBench ( Huang et al., 2024 ) (lower row). (a-b) Min-max normalized L2 distances of output and input features, measured between consecutive timesteps. (c) RSD of sk(t−k) in Eq. ( 5 ) with varying t . (d) Relative L1 errors of the output and input features, i.e ., EO(t) and EI(t) in Eq. ( 21 ), respectively.
Methods
NFC
FID ↓
sFID ↓
FID2FC ↓
sFID2FC ↓
Full-Compute
50
2.32
4.32
-
-
TaylorSeer ( Liu et al., 2025b )
6
13.57
13.19
10.68
15.19
RFC
6
4.31
5.39
1.83
4.80
TaylorSeer ( Liu et al., 2025b )
5
45.81
27.64
42.29
31.46
RFC
5
5.43
6.29
2.81
6.13
TaylorSeer ( Liu et al., 2025b )
4
160.85
135.55
157.53
137.65
Appendix
Table 7: Quantitative comparison under extremely high acceleration ratios for DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . All results are obtained using the second-order approximation ( m=2 ).
Methods
NFC
PSNR ↑
SSIM ↑
LPIPS ↓
Full-Compute
6
-
1.0000
0.0000
Step reduction
4
26.5377
0.8855
0.1164
TaylorSeer ( Liu et al., 2025b )
4
29.3435
0.9242
0.0736
RFC
4
32.7436
0.9275
0.0635
Step reduction
3
25.0214
0.8580
0.1511
TaylorSeer ( Liu et al., 2025b )
3
22.2134
0.7491
0.3687
Appendix
Table 8: Quantitative comparison on distilled model, FLUX.1 schnell ( Batifol et al., 2025 ) with 6 denoising steps, on DrawBench ( Saharia et al., 2022 ) . All results are obtained using the first-order approximation ( m=1 ).
Figure 7: Qualitative comparisons of class-conditional image generation for DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) .
Figure 8: Qualitative comparisons of text-to-image generation for FLUX.1 dev ( Batifol et al., 2025 ) on DrawBench ( Saharia et al., 2022 ) .
Figure 9: Qualitative comparisons of text-to-video generation for HunyuanVideo ( Kong et al., 2024 ) on VBench ( Huang et al., 2024 ) .
Model
Order ( m )
τ
NFC
DiT-XL/2 ( Peebles and Xie, 2023 )
2
0.19
14.01
0.38
11.01
0.48
10.04
0.88
8.02
1.20
7.04
FLUX.1 dev ( Batifol et al., 2025 )
1
0.30
14.02
Appendix
Table 9: Settings of τ for various models.
Figure 10: Trade-off between generation quality and computational cost on ImageNet ( Deng et al., 2009 ) using DiT-XL/2 ( Peebles and Xie, 2023 ) under varying values of τ . The plot shows the change in performance ( i.e ., sFID2FC) as τ increases, with the corresponding FLOPs indicated by the size and color of each circle.
Figure 11: Clip scores of text-to-image generation for FLUX.1 dev ( Batifol et al., 2025 ) on DrawBench ( Saharia et al., 2022 ) .
Methods
NFC
FLOPs (T) ↓
FID ↓
sFID ↓
FID2FC ↓
sFID2FC ↓
IS ↑
Full-Compute
50
23.74
2.32
4.32
-
-
241.25
Shallow Skip
16
6.66
3.09
4.65
0.75
2.61
220.00
Deep Skip
16
6.66
3.47
5.15
1.10
2.82
211.54
RFC ( m=2 )
14
6.66
2.52
4.60
0.30
1.33
231.00
Shallow Skip
8
3.35
11.70
9.13
9.10
11.71
147.88
Deep Skip
8
3.35
10.20
6.22
6.95
6.59
149.10
Appendix
Table 10: Quantitative results for layer-wise caching strategies on ImageNet ( Deng et al., 2009 ) using DiT-XL/2 ( Peebles and Xie, 2023 ) . We reduce the frequency of full computations by half for the first (shallow skip) or last (deep skip) 7 blocks out of 28. We adjust τ to use more full computations for these variants to match FLOPs. All results are obtained using the second-order approximation ( m=2 ).
Methods
NFC
FID2FC ↓
sFID2FC ↓
FORA ( Selvaraju et al., 2024 ) (N=5)
11
78.27
69.84
TaylorSeer ( Liu et al., 2025b ) (N=5)
11
18.83
11.83
RFC
11
8.55
7.59
FORA ( Selvaraju et al., 2024 ) (N=6)
10
103.48
89.17
TaylorSeer ( Liu et al., 2025b ) (N=6)
10
37.81
21.35
RFC
10
11.64
8.25
Appendix
Table 11: Quantitative comparison of FORA ( Selvaraju et al., 2024 ) , TaylorSeer ( Liu et al., 2025b ) , and RFC using the DDIM ( Song et al., 2020 ) model trained on LSUN-bedroom ( Yu et al., 2015 ) with 50 timesteps. We use the first-order approximation ( m=1 ).
Figure 12: Analysis of non-uniform intervals scheduled by RCS using DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) . We visualize the average interval between full-compute steps for (a) 14 and (b) 11 NFC settings. i→i+1 at the x-axis denotes the i -th and (i+1) -th full computation, and the y-axis shows the average interval. The results are averaged across 1,000 samples.
Methods
NFC
Order
Peak Memory (MB)
FID2FC ↓
sFID2FC ↓
FORA ( Selvaraju et al., 2024 )
10
–
3121
5.32
12.87
TaylorSeer ( Liu et al., 2025b )
10
1
3207
1.42
6.17
RFC
10.04
1
3335
0.58
2.79
TaylorSeer ( Liu et al., 2025b )
10
2
3333
1.05
4.86
RFC
10.02
2
3463
0.54
2.43
Appendix
Table 12: Quantitative results of DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) , in terms of NFC, order in the Taylor expansion, GPU peak memory, FID2FC, and sFID2FC.
Methods
Time (Overhead %)
Baseline (TaylorSeer ( Liu et al., 2025b ) )
2.840s
Input feature computation
+ 0.019s (0.67%)
Input feature prediction
+ 0.002s (0.07%)
Appendix
Table 13: Time overhead of RFC components compared to the baseline (TaylorSeer ( Liu et al., 2025b ) ) using DiT-XL/2 ( Peebles and Xie, 2023 ) on ImageNet ( Deng et al., 2009 ) with 14 NFCs. Percentages in parentheses indicate the time increase relative to the baseline.
Diffusion model caching reuses features, activations, or attention states across denoising steps to reduce inference cost while preserving output quality.
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.
Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To alleviate their massive computational overhead, temporal feature caching has been proposed to bypass redundant computations. However, existing cache-then-forecast methods driven by derivative-based polynomials often cause severe quality degradation under high acceleration due to unstable long-step predictions. To address this bottleneck, we propose Barycentric Rational Forecasting with Chebyshev Enhancement (BRACE). Motivated by the observation that DiT feature trajectories are globally smooth yet frequently exhibit sharp irregularities and local non-smoothness, BRACE shifts the paradigm from derivative-driven polynomial extrapolation to feature-driven rational forecasting. Specifically, it maintains a local sliding window to cache sparse historical features and leverages adapted Chebyshev weights to formulate a barycentric rational function, directly aggregating these raw features to ensure numerical stability. Extensive experiments demonstrate that BRACE achieves state-of-the-art quality-efficiency trade-offs across various DiT architectures with negligible computational overhead.
Jinlong Yang, Jinke Wu, Lizilin +1
Sichuan University Chengdu, China · School of Artificial Intelligence, Sichuan University Chengdu, China
To address the high sampling cost of Diffusion Transformers (DiTs), feature caching offers a training-free acceleration method. However, existing methods rely on hand-crafted forecasting formulas that fail under aggressive skipping. We propose L2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces fixed coefficients with learnable per-timestep weights. Rapidly trained in ~20 seconds on a single GPU, L2P accurately reconstructs current features from past trajectories. L2P significantly outperforms existing baselines: it achieves a 4.55x FLOPs reduction and 4.15x latency speedup on FLUX.1-dev, and maintains high visual fidelity under up to 7.18x acceleration on Qwen-Image models, where prior methods show noticeable quality degradation. Our results show learning linear predictors is highly effective for efficient DiT inference. Code is available at https://github.com/Aredstone/L2P-Cache.
Zhirong Shen, Rui Huang, Jiacheng Liu +6
Shanghai Jiao Tong University · University of Electronic Science and Technology of China · Shandong University +2