Search papers, labs, and topics across Lattice.
The paper introduces Trajectory-Consistent Padé approximation (TC-Padé), a novel feature prediction framework for accelerating diffusion model sampling, particularly in the low-step regime. TC-Padé uses Padé approximants to model feature evolution, capturing asymptotic and transitional behaviors more effectively than Taylor expansions, thus mitigating error accumulation and trajectory drift. By incorporating adaptive coefficient modulation and step-aware prediction strategies, TC-Padé achieves significant speedups (e.g., 2.88x on FLUX.1-dev) while maintaining high generation quality across various benchmarks.
Diffusion models get a 2.88x speed boost in low-step regimes thanks to a new Padé approximation technique that predicts feature evolution more accurately than Taylor-based methods.
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.
