Diffusion Transformers (DiTs) have established themselves as the preeminent paradigm for high-fidelity generative modeling across various modalities, yet their practical deployment is hindered by the severe computational overhead of iterative sampling. While recent feature-caching methods mitigate this latency through feature reuse or forecasting based on integer-order operators, they remain limited in adapting to the non-stationary and highly curved dynamics of DiT latent trajectories. In this work, we reveal that DiT feature evolution exhibits anomalous diffusion characteristics and a certain degree of spatiotemporal heterogeneity, rendering static, integer-order forecasting operators inadequate and prone to error accumulation. To address this, we propose Fractional Anomalous Dynamics Extrapolation (FADE), a unified framework for training-free diffusion transformers acceleration. FADE models DiT feature evolution via anomalous diffusion and employs a Mittag--Leffler fractional operator to better adapt to high-curvature and non-linear latent trajectories of DiTs. Leveraging the empirical observation that these complex dynamics maintain cross-sample stability, we further construct a Look-Up Table with multiple dimensions through offline profiling. This enables negligible LUT-retrieval overhead, spatiotemporally adaptive parameter retrieval during online inference. Extensive evaluations across diverse architectures and modalities demonstrate that FADE achieves state-of-the-art acceleration, preserving exceptional generation fidelity under aggressive skip intervals with negligible latency overhead. Our project page is available at https://youngkinlon.github.io/fade-project-page/.
Figures & tables
Figure 1: Empirical evidence of anomalous and structured dynamics in DiT. (a) The normalized mean squared displacement across layers follows an anomalous diffusion scaling law. (b) Cross-sample trajectory consistency across different layers and denoising steps.
Figure 2: Overall framework of FADE. FADE consists of an offline profiling stage and an online inference stage. The offline stage calibrates and stores the optimal fractional orders in a static LUT, while the online stage alternates between full-compute and prediction steps using LUT-retrieved orders for feature forecasting. The upper-right panel illustrates the geometric contrast between the FADE trajectory and the integer-order trajectory in the latent space.
Figure 3: Empirical validation of fractional dynamics on DiT-XL/2. The optimal fractional orders deviate markedly from the static integer-order assumption and exhibit a pronounced non-stationary temporal pattern across denoising timesteps.
Method
Lat. (s) ↓
FLOPs (T) ↓
Spd. ↑
FID ↓
sFID ↓
IS ↑
DDIM-50 steps Song et al. (2022)
1.18
23.74
1.00 ×
2.29
4.29
237.31
DDIM-25 steps Song et al. (2022)
0.62
11.87
2.00 ×
2.91
4.57
232.22
DDIM-15 steps Song et al. (2022)
0.42
7.12
3.33 ×
5.09
6.11
205.54
FORA( N =4) Selvaraju et al. (2024)
0.45
6.66
3.56 ×
4.73
8.45
214.20
Taylor ( N =4, m =4) Liu et al. (2025b)
0.61
6.66
3.56 ×
2.59
5.18
232.90
HiCache ( N =4, m =4) Feng et al. (2026)
0.57
6.66
3.56 ×
2.51
5.20
232.72
Table 1: Quantitative comparison of class-conditional image generation with DiT-XL/2
Method
Latency(s) ↓
FLOPs(T) ↓
Speed ↑
Img. Reward ↑
CLIP Score ↑
Cycle Reward ↑
PSNR ↑
SSIM ↑
LPIPS ↓
FLUX.1 [dev] - 50 steps †
17.18
3719.50
1.00 ×
0.9898
27.4611
0.8800
∞
1.0000
0.0000
FLUX.1 [dev] - 25 steps
8.80
1859.75
2.00 ×
0.9375
27.3388
0.8466
19.2599
0.7749
0.3583
FLUX.1 [dev] - 20 steps
7.12
1487.80
2.62 ×
0.9387
27.2505
0.8466
18.1371
0.7484
0.4006
FORA ( N =5) Selvaraju et al. (2024)
5.01
893.54
4.16 ×
0.8253
27.2377
0.8353
15.9899
0.6677
0.5206
TaylorSeer ( N =5, m=1) Liu et al. (2025b)
5.14
893.54
4.16 ×
0.9919
27.5143
0.8540
18.4497
0.7424
0.4146
HiCache ( N =5, m=1) Feng et al. (2026)
5.12
893.54
4.16 ×
0.9350
27.4034
0.8346
19.2595
0.7503
0.3911
Table 2: Quantitative comparison of text-to-image generation on FLUX
Figure 4: Visualization examples for different acceleration methods on Flux.
Method
Lat. (s) ↓
FLOPs (T) ↓
Spd. ↑
VBench (%) ↑
Original (50 steps)
190.2
29773
1.00 ×
80.68
DDIM (22% steps) Song et al. (2022)
42.7
6550
4.55 ×
78.73
FORA (N=6) Selvaraju et al. (2024)
37.1
5359
5.56 ×
78.86
TaylorSeer (N=6, m=1) Liu et al. (2025b)
37.8
5359
5.56 ×
79.87
HiCache (N=6, m=1) Feng et al. (2026)
37.8
5359
5.56 ×
79.93
FADE (N=6, m=1)
37.9
5359
5.56 ×
80.12
Table 3: Quantitative comparison of text-to-video generation on HunyuanVideo.
Figure 5: Qualitative comparison on text-to-video generation.
Table 9
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Network Phase
Layer Indices
Attention Module (Attn)
Mean γ∗
Std ( σ )
MAD
Early Layers
Layers 0–9
0.64
0.144
0.128
Middle Layers
Layers 10–18
0.96
0.129
0.119
Deep Layers
Layers 19–27
0.46
0.128
0.109
Global Summary
All Layers
0.68
0.120
0.123
Appendix
Table 6: Cross-sample consistency of the optimal fractional order γ∗ in DiT attention layers. We highlight the statistics at the most stable denoising phase (Timestep t=40 ). The consistent behavior in deep layers and its robustness to early-stage variance formally validate our offline calibration strategy.
Figure 6: Spatiotemporal visualization of the optimal fractional orders ( γ∗ ) on DiT-XL/2. (a) Global KDE distribution of γ∗ . (b) Variation across denoising timesteps and transformer layers, showing a broadly shared U-shaped temporal pattern.
Figure 7: Spatiotemporal decoupling of latent dynamics across FLUX.1 architectures. Calibrated orders in (a) DoubleStream and (b) SingleStream blocks. SingleStream blocks exhibit a greater proportion of higher-order states. Across both streams, higher orders are concentrated in the middle stages of T0→T1 .
Figure 8: Spatiotemporal decoupling of latent dynamics across HunyuanVideo architectures. Both (a) DoubleStream and (b) SingleStream blocks show a broad shift toward higher-order regimes ( γ>1.1 ), while near-integer states ( γ≈1 ) are comparatively uncommon.
Method
Speed ↑
IR ↑
CLIP ↑
FADE (fixed γ=1.2 )
4.99 ×
0.9883
27.5061
FADE (fixed γ=1.6 )
4.99 ×
1.0016
27.6207
FADE (fixed γ=1.8 )
4.99 ×
0.9978
27.6239
FADE (fixed γ=2.0 )
4.99 ×
0.9936
27.6617
FADE (dynamic fit γ(t) )
4.99 ×
0.9948
27.7077
Appendix
Table 7: Additional ablation on fractional-order scheduling strategies for FLUX.1.
Method
FID ↓
sFID ↓
DDIM-50
2.29
4.29
HiCache
2.51
5.20
Taylor
2.590
5.18
FADE (fixed γ=0.85 )
2.49
4.81
FADE (Bernoulli-fit γ(t) )
2.51
5.14
Appendix
Table 8: Additional ablation on fractional-order scheduling strategies for DiT-XL/2. “–” indicates that the metric was not recorded.
Model
Cache interval N
Refinement factor λ
DiT-XL/2
4
0.6
DiT-XL/2
5
0.6
DiT-XL/2
6
0.7
FLUX.1-dev
5
1.3
FLUX.1-dev
6
1.0
FLUX.1-dev
7
1.1
Appendix
Table 9: Online inference settings of FADE. The refinement factor λ is fixed after offline calibration and introduces no online search overhead.
Diffusion Transformers (DiT) have emerged as a widely adopted backbone for high-fidelity image and video generation, yet their iterative denoising process incurs high computational costs. Existing training-free acceleration methods rely on feature caching and reuse under the assumption of temporal stability. However, reusing features for multiple steps may lead to latent drift and visual degradation. We observe that model outputs evolve smoothly along much of the diffusion trajectory, enabling principled predictions rather than naive reuse. Based on this insight, we propose PrediT, a training-free acceleration framework that formulates feature prediction as a linear multistep problem. We employ classical linear multistep methods to forecast future model outputs from historical information, combined with a corrector that activates in high-dynamics regions to prevent error accumulation. A dynamic step modulation mechanism adaptively adjusts the prediction horizon by monitoring the feature change rate. Together, these components enable substantial acceleration while preserving generation fidelity. Extensive experiments validate that our method achieves up to 5.54× latency reduction across various DiT-based image and video generation models, while incurring negligible quality degradation.
Hanshuai Cui, Zhiqing Tang, Qianli Ma +3
School of Artificial Intelligence, Beijing Normal University, Beijing, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China · Shenzhen University of Advanced Technology, Shenzhen, China
Diffusion Transformers achieve high-fidelity image and video generation, but their iterative sampling remains expensive, for each denoising step requires large matrix operations. Existing cache-based acceleration reduces redundant computation yet increases the VRAM footprint by storing intermediate states, which can directly constrain inference batch size. In this work, we propose a training-free acceleration method that performs stepwise forecasting for DiT sampling using a Barycentric Extrapolator. By leveraging barycentric extrapolation, our predictor is numerically stable and alleviates oscillatory artifacts analogous to the Runge phenomenon during forward forecasting. Across extensive experiments on both image and video generation, our approach provides a favorable trade-off between memory usage and perceptual quality, while delivering up to 3.30x end-to-end sampling speedup compared with baseline DiT inference.
Chengjie Lu, Tianchi Deng, Zhengqi He +3
College of Electronics and Information Engineering, Shenzhen University · School of Artificial Intelligence, Shenzhen University · College of Computer Science and Software Engineering, Shenzhen University
Diffusion Transformers (DiTs) have shown strong performance in high-fidelity image generation, but their sampling process remains computationally intensive due to full model execution at every timestep. While cache-based acceleration has been explored to mitigate inference cost, naive reuse schemes suffer from low accuracy over long intervals, and Taylor-series-based extrapolation methods often face instability caused by Runge oscillations. In this paper, we propose ChebBooster, a training-free extrapolation framework based on Chebyshev polynomial theory that achieves stable and efficient acceleration for DiTs. Specifically, we adopt the Barycentric formulation to evaluate Chebyshev approximants with high numerical stability and minimal overhead, and further decouple the extrapolation into an offline weight precomputation phase and a lightweight online application stage. Extensive experiments across three representative DiT-based models, including DiT-XL/2, PixArt-Σ, and FLUX.1-dev, demonstrate that ChebBooster achieves consistent improvements in visual quality and inference efficiency, reaching up to 3.68× latency speedup and 5.12× FLOPs reduction, outperforming existing training-free baselines under diverse generation tasks and resolutions.
Chengjie Lu, Tianchi Deng, Zhengqi He +2
College of Electronics and Information Engineering, Shenzhen University · School of Artificial Intelligence, Shenzhen University