Adaptive Step Samplers: Zero-Cost Truncation Error Estimation and Exposed Stochastic Control for Diffusion and Flow-Matching Models

Adaptive Step Samplers is an open-source numerical analysis and inference acceleration library for generative diffusion and flow-matching models, integrated natively into ComfyUI. The research and codebase address two fundamental problems in modern multistep ODE/SDE sampling: Zero-cost local truncation error estimation: In classical numerical analysis, predictor–corrector algorithms (e.g. Adams–Bashforth predictor coupled with an Adams–Moulton corrector) require an extra function evaluation per step to evaluate the vector field at the predicted state. In multistep diffusion sampling, the evaluation produced by the corrector serves as the exact historical derivative needed by the subsequent step. Hence, the local truncation error estimate |xₜ - x̂ₜ| is obtained at zero additional neural network function evaluations (0 extra NFE). Decoupled stochastic control in SA-Solver: SA-Solver (Xue et al., NeurIPS 2023) decouples deterministic drift integration from Brownian diffusion, maintaining 1 NFE per step. In existing frameworks—including ComfyUI’s core—the stochastic intensity is immutably pinned to η = 1.0 within a fixed $[0.2, 0.8]$ noise window. This library exposes the parameter space (η, response curves, and window boundaries) directly to the execution graph. Empirical Benchmarks & Key Findings: The Geometry–Texture Tradeoff: Benchmarked on Anima (flow-matching DiT, 768×1024, CFG 5.0, normal scheduler, 10 NFE budget), drift correction allows the stochastic solver at η = 1.0 to outperform standard deterministic Euler in macro-geometry (RMS deviation from 100-step Euler ground truth: 0.3516 vs. 0.3677) while preserving 64% more high-frequency micro-relief (0.2244 vs. 0.1368). Optimal Stochastic Strength (η ∈ [1.6, 1.8]): Scaling η to 1.8 closes 59% of the texture deficit relative to a converged 30-step SDE reference. Beyond η = 1.8, the marginal rate of substitution degrades drastically ($0.28$ vs $1.06$), causing fine structures to disintegrate into noise. Negative Empirical Results: Adaptive step-size redistribution based on local error equidistribution decisively outperforms uniform grids, but achieves only 0.81–0.99× the accuracy of ComfyUI's heuristic normal scheduler. High-order ODE advantage decay: While 4th-order multistep solvers beat Euler by 1.30–1.52× at CFG 1.0, this advantage drops to 1.01–1.04× at CFG 5.0 and vanishes at CFG 8.0, confirming that guidance extrapolation noise dominates truncation error at production scales. Acknowledgments Compute infrastructure, hardware allocation, and editorial support provided by Yehor Yuzvynsky. Suggested Citation (BibTeX) @software{savelev_yuzvynsky_2026_adaptivesamplers, author = {Iakov Savelev and Yehor Yuzvynsky}, title = {Adaptive Step Samplers: Zero-Cost Truncation Error Estimation and Exposed Stochastic Control for ComfyUI}, year = {2026}, publisher = {Zenodo}, doi = {10.5281/zenodo.23529356}, url = {https://github.com/YakovSava/adaptive-step-samplers} }