RBF-Solver: A Multistep Sampler for Diffusion Probabilistic Models via Radial Basis Functions

Soochul Park, Yeon Ju Lee, SeongJin Yoon, Jiyub Shin, Juhee Lee, Seongwoon JoarXiv 2026 · arXiv

TL;DR

RBF-Solver introduces a novel multistep sampler for diffusion probabilistic models by replacing rigid polynomial approximations with flexible Gaussian Radial Basis Functions (RBFs). By optimizing learnable shape parameters, it explicitly adapts to optimal sampling trajectories, delivering superior image fidelity at low function evaluations and maintaining numerical stability at high orders where traditional samplers fail.

What Problem Does This Paper Address?

Diffusion Probabilistic Models (DPMs) excel at generating high-fidelity data but suffer from high computational costs due to their iterative denoising process. To accelerate inference, recent state-of-the-art methods employ high-order multistep ODE solvers (such as DPM-Solver++ and UniPC). These samplers reuse previous model evaluations to calculate the next step, relying on algebraic polynomial approximations like Taylor expansions or Lagrange interpolations.

However, these polynomial approaches are rigid and follow predefined schemes that leave no room for trajectory optimization. More critically, algebraic polynomials suffer from the Runge phenomenon—severe numerical instability and wild oscillations at higher orders ($order \ge 4$)—which causes image generation quality to sharply deteriorate when trying to use longer history steps.

Methods

The authors introduce RBF-Solver, a multistep diffusion sampler grounded in Radial Basis Function interpolation rather than algebraic polynomials.

Key components include:

  • Gaussian RBF Interpolation: Instead of fitting a polynomial to past model evaluations, RBF-Solver approximates the diffusion trajectory using a linear combination of Gaussian RBFs and a constant basis term (essential for maintaining mathematical consistency with standard solvers).
  • Learnable Shape Parameters: Each Gaussian basis function features a shape parameter ($\gamma$) that controls its width. This parameter dictates the local influence of each past evaluation point.
  • Offline Optimization: The shape parameters are optimized via a grid search on a small target dataset (typically 128 image-noise pairs) to minimize local discretization errors, effectively forcing the solver to follow an optimized sampling trajectory.
  • Mathematical Generalization: RBF-Solver unifies several classic numerical methods. At first order, it reduces to the standard Euler method (DDIM). At higher orders, as the shape parameters approach infinity ($\gamma \to \infty$), the RBF interpolant converges exactly to the polynomial-based Adams method.
Figure 1: (a) During sampling, the three model evaluations at time points t3, t2, and t1 are interpolated using both radial basis functions (RBFs) and a constant basis, each basis being multiplied by its weight. Only the values of the central green pixels are plotted; the points used for interpolation are shown in black, while unused future and past points are shown in gray. (b) For second- and third- order schemes, the curves show how the interpolation coefficients vary with the shape parameter γ. As log γ →−∞, all coefficients converge to the same value, whereas as log γ →∞, they approach the coefficients of the Adams method.

Datasets

The method was evaluated using standard generative benchmarks:

  • CIFAR-10 (evaluated on Score-SDE and EDM backbones)
  • ImageNet at $64\times64$, $128\times128$, and $256\times256$ resolutions (using Improved Diffusion and Guided Diffusion models)
  • LAION-5B (evaluated via Stable Diffusion v1.4)

No new datasets or benchmarks were introduced in this paper.

Key Insights and Ideas

  • Conquering the Runge Phenomenon: Because Gaussian RBFs are localized functions (their influence decays rapidly away from their center), they inherently avoid the edge-oscillations that plague high-degree polynomials. This allows RBF-Solver to remain highly stable at fourth-order accuracy and beyond.
  • Data-Driven Trajectory Tuning: The shape parameter $\gamma$ acts as a continuous dial. Instead of guessing the best polynomial degree, engineers can optimize $\gamma$ to smoothly adapt the sampler's coefficients to the specific dynamics of a pre-trained diffusion model.
  • Target vs. Model Trajectories: During optimization, matching the solver to a "target trajectory" (constructed via forward diffusion from real images) yields substantially better performance than matching it to a "model trajectory" (the path carved out by a high-step teacher sampler).

Key Results

  • High-Order Stability: In head-to-head comparisons on ImageNet $64\times64$, polynomial samplers like UniPC experience catastrophic performance drops at orders $p \ge 5$. RBF-Solver remains completely stable, achieving its optimal Fréchet Inception Distance (FID) at orders 5 and 6.
  • Unconditional Generation Performance: On CIFAR-10 with the Score-SDE model, RBF-Solver consistently outperforms DPM-Solver++ and UniPC when the Number of Function Evaluations (NFE) is $\ge 15$, securing an FID of 2.87 at 15 NFE and 2.48 at 40 NFE.
Figure 2: Unconditional sampling results. FID↓on 50k samples.
  • Conditional Generation Gains: For conditional ImageNet $256\times256$ generation (Guidance Scale = 8.0), RBF-Solver achieves a 16.12% to 33.73% reduction in FID in the ultra-low NFE regime (5–10 steps) compared to polynomial baselines.

Limitations and Future Work

  • Kernel Restrictions: This study focuses exclusively on the Gaussian RBF kernel. Other popular RBF choices, such as multiquadrics or Wendland's compactly supported functions, were not explored.
  • Offline Optimization Overhead: The shape parameters must be optimized beforehand on a small dataset. The authors note that developing an adaptive, on-the-fly runtime scheme to adjust these parameters dynamically without an offline tuning phase remains an open challenge for future work.

Summary updated

Abstract

Diffusion probabilistic models (DPMs) are widely adopted for their outstanding generative fidelity, yet their sampling is computationally demanding. Polynomial-based multistep samplers mitigate this cost by accelerating inference; however, despite their theoretical accuracy guarantees, they generate the sampling trajectory according to a predefined scheme, providing no flexibility for further optimization. To address this limitation, we propose RBF-Solver, a multistep diffusion sampler that interpolates model evaluations with Gaussian radial basis functions (RBFs). By leveraging learnable shape parameters in Gaussian RBFs, RBF-Solver explicitly follows optimal sampling trajectories. At first order, it reduces to the Euler method (DDIM). At second order or higher, as the shape parameters approach infinity, RBF-Solver converges to the Adams method, ensuring its compatibility with existing samplers. Owing to the locality of Gaussian RBFs, RBF-Solver maintains high image fidelity even at fourth order or higher, where previous samplers deteriorate. For unconditional generation, RBF-Solver consistently outperforms polynomial-based samplers in the high-NFE regime (NFE >= 15). On CIFAR-10 with the Score-SDE model, it achieves an FID of 2.87 with 15 function evaluations and further improves to 2.48 with 40 function evaluations. For conditional ImageNet 256 x 256 generation with the Guided Diffusion model at a guidance scale 8.0, substantial gains are achieved in the low-NFE range (5-10), yielding a 16.12-33.73% reduction in FID relative to polynomial-based samplers.
BibTeX
@misc{park2026rbf,
  author = {Soochul Park and Yeon Ju Lee and SeongJin Yoon and Jiyub Shin and Juhee Lee and Seongwoon Jo},
  title = {{RBF-Solver: A Multistep Sampler for Diffusion Probabilistic Models via Radial Basis Functions}},
  year = {2026},
  eprint = {2603.13330},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.13330},
}
Figure