Despite achieving state-of-the-art generation quality, diffusion models are hindered by the substantial computational burden of their iterative sampling process. While feature caching techniques achieve effective acceleration at higher step counts (e.g., 50 steps), they exhibit critical limitations in the practical low-step regime of 20-30 steps. As the interval between steps increases, polynomial-based extrapolators like TaylorSeer suffer from error accumulation and trajectory drift. Meanwhile, conventional caching strategies often overlook the distinct dynamical properties of different denoising phases. To address these challenges, we propose Trajectory-Consistent Padé approximation, a feature prediction framework grounded in Padé approximation. By modeling feature evolution through rational functions, our approach captures asymptotic and transitional behaviors more accurately than Taylor-based methods. To enable stable and trajectory-consistent sampling under reduced step counts, TC-Padé incorporates (1) adaptive coefficient modulation that leverages historical cached residuals to detect subtle trajectory transitions, and (2) step-aware prediction strategies tailored to the distinct dynamics of early, mid, and late sampling stages. Extensive experiments on DiT-XL/2, FLUX.1-dev, and Wan2.1 across both image and video generation demonstrate the effectiveness of TC-Padé. For instance, TC-Padé achieves 2.88x acceleration on FLUX.1-dev and 1.72x on Wan2.1 while maintaining high quality across FID, CLIP, Aesthetic, and VBench-2.0 metrics, substantially outperforming existing feature caching methods.
Figures & tables
Figure 1: Visual quality comparison on FLUX.1-dev with 50 steps (left) and 20 steps (right). Numbers indicate speedup ratios relative to the original sampling. Previous cache-based methods exhibit notable quality degradation with altered textures and colors under low-step settings. In contrast, TC-Padé preserves visual fidelity while achieving higher acceleration. Additional qualitative comparisons with other methods are provided in the Appendix.
Figure 2: PCA visualization of final layer outputs from various caching-based methods under 20 steps sampling regime. ∥V(t)∥ represents the PCA results of the model’s output velocity field. Data is collected from FLUX.1-dev on the DrawBench [ 40 ] dataset.
Figure 3: Overview of TC-Padé within a cache interval N . Here, N=4 . {xt+3,xt+2,xt+1,xt} and {yt+3,yt+2,yt+1,yt} denote the input and output of each timestep, respectively. {Rt+3,Rt+2,Rt+1,Rt} are the cached residuals. θ is a predefined threshold. In each cache interval, only the initial timestep performs full computation, while subsequent timesteps adaptively determine their computation mode via the Trajectory Stableness Indicator (TSI).
Figure 4: (a) Residual and raw feature similarity between our method and the original sampling schedule, showing that residuals have consistently higher similarity. (b) Raw feature similarity between the original sampling schedule and TaylorSeer under different settings.
Figure 5
Figure 6: Qualitative results of FLUX.1-dev across multiple resolutions. Each image pair shows the original 20-step sampling result and TC-Padé accelerated sampling result. Numbers indicate latency (top-left) and speedup ratio (top-right).
DrawBench
Visual Quality
ImgRwd ↑
Aesthetic ↑
Flux.1[dev], 20steps
1.01
5.83
Δ -DiT (N = 3)
0.52
5.42
ToCa(N=5)
0.82
5.24
TeaCache( δ=0.25 )
0.85
5.73
TeaCache( δ=0.4 )
0.86
5.71
Table 7: Quantitative results for text-to-image generation on DrawBench . Higher is better for quality metrics, and lower is better for efficiency metrics. The best results are in bold , and the second best are underlined .
Figure 7: VBench-2.0 metrics and acceleration ratio of TC-Padé and other methods.
Figure 8: User preference comparison for text-to-image generation on FLUX.1-dev across different caching methods.
Figure 9: Qualitative examples for TC-Padé and other cache methods on FLUX.1-dev using 20 sampling steps. The text prompts are sampled from Parti prompts.
Figure 10: Qualitative examples for TC-Padé and other cache methods on FLUX.1-dev using 20 sampling steps. The text prompts are sampled from Parti prompts.
Figure 11: Qualitative examples for TC-Padé and other cache methods on Wan2.1-1.3B using 20 sampling steps.
Figure 12: Qualitative examples for TC-Padé and other cache methods on Wan2.1-1.3B using 20 sampling steps.
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× speedup and even surpassing the full-step baseline by +0.028 ImageReward at 3.54× speedup on FLUX, validating the effectiveness of our path-aware global scheduling framework. Code is available at https://github.com/argsss/DPCache.
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 (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