TC-Padé: Trajectory-Consistent Padé Approximation for Diffusion Acceleration

Abstract

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 (TC-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.

Method

Figure 2. Overview of TC-Padé within a cache interval N. 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).

(1) PCA visualization

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 dataset.

(2) Residual and raw feature similarity analysis

(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.

Visualizations

Text-guided image examples for TC-Padé and other cache methods on FLUX.1-dev using 20 sampling steps. The text prompts are sampled from Parti prompts.

Results

Table 1. Quantitative results for text-to-image generation on COCO 2017 dataset. The best results are in bold, and the second best are underlined. Values marked with † indicate severe degradation in output image quality, with results falling outside the acceptable range for meaningful comparison.

Table 2. Quantitative comparison for text-to-video generation for Wan2.1-1.3B on VBench-2.0.

Table 3. Quantitative comparison for class-conditional image generation on ImageNet with DiT-XL/2.

BibTeX

@misc{cui2026tcpadetrajectoryconsistentpadeapproximation,
      title={TC-Pad\'e: Trajectory-Consistent Pad\'e Approximation for Diffusion Acceleration}, 
      author={Benlei Cui and Shaoxuan He and Bukun Huang and Zhizeng Ye and Yunyun Sun and Longtao Huang and Hui Xue and Yang Yang and Jingqun Tang and Zhou Zhao and Haiwen Hong},
      year={2026},
      eprint={2603.02943},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.02943}, 
    }

Trajectory-Consistent Padé Approximation for Diffusion Acceleration

Figure 1. Visual quality comparison on FLUX.1-dev. Previous methods exhibit notable quality degradation under low-step settings. In contrast, TC-Padé preserves visual fidelity while achieving higher acceleration.