Bi-Anchor Interpolation Solver for Accelerating Generative Modeling

TL;DR

The paper tackles the latency problem of Flow Matching-based generative modeling by introducing BA-solver, which keeps the backbone frozen and augments it with a lightweight SideNet to enable bidirectional temporal perception. It leverages Bi-Anchor Velocity Integration to interpolate intermediate velocities using two anchors, achieving high-order, batched integration with a strict Exact-$N$ NFE budget. Chain-based training for the SideNet ensures robustness with minimal backbone evaluations, yielding state-of-the-art generation quality at very low NFEs (e.g., 5–10) on ImageNet resolutions, while significantly reducing training cost and parameter count. This approach offers a practical, plug-and-play acceleration for real-world generative pipelines, with strong implications for efficient and scalable image synthesis and editing.

Abstract

Flow Matching (FM) models have emerged as a leading paradigm for high-fidelity synthesis. However, their reliance on iterative Ordinary Differential Equation (ODE) solving creates a significant latency bottleneck. Existing solutions face a dichotomy: training-free solvers suffer from significant performance degradation at low Neural Function Evaluations (NFEs), while training-based one- or few-steps generation methods incur prohibitive training costs and lack plug-and-play versatility. To bridge this gap, we propose the Bi-Anchor Interpolation Solver (BA-solver). BA-solver retains the versatility of standard training-free solvers while achieving significant acceleration by introducing a lightweight SideNet (1-2% backbone size) alongside the frozen backbone. Specifically, our method is founded on two synergistic components: \textbf{1) Bidirectional Temporal Perception}, where the SideNet learns to approximate both future and historical velocities without retraining the heavy backbone; and 2) Bi-Anchor Velocity Integration, which utilizes the SideNet with two anchor velocities to efficiently approximate intermediate velocities for batched high-order integration. By utilizing the backbone to establish high-precision ``anchors'' and the SideNet to densify the trajectory, BA-solver enables large interval sizes with minimized error. Empirical results on ImageNet-256^2 demonstrate that BA-solver achieves generation quality comparable to 100+ NFEs Euler solver in just 10 NFEs and maintains high fidelity in as few as 5 NFEs, incurring negligible training costs. Furthermore, BA-solver ensures seamless integration with existing generative pipelines, facilitating downstream tasks such as image editing.

Figures (8)

• Figure 1: (a). Extrapolation Solvers only need a single NFE to calculate current velocity within each interval, but it has large extrapolation error. (b). Interpolation Solvers are high-accuracy methods, but they need multiple sequential NFEs within each interval to calculate current and intermediate velocities, which is not efficient for sampling. (c). BA-solver utilizes lightweight SideNet to predict intermediate velocities based on bidirectional anchors ($\bm{v}_t$ and $\bm{v}_{t-h}$), efficiently providing a high-accuracy approximation for $\bm{x}_{t-h}$.
• Figure 2: FID, training iteration, and NFEs across methods on ImageNet-$256^2$. BA-solver is located at the sweet spot.
• Figure 3: (a) Forward Probe using Single-Anchor Interpolation Solver to acquire $\bm{x}^{pred}_{t-h}$. (b) Backward Refinement for intermediate velocities utilizing SideNet's lookback ability anchored on velocity $\bm{v}_{t-h}$. (c) Integration & State Reuse. By high-order integration for two anchor velocities and multiple velocities, we can acquire a more accurate $\bm{x}^{pred}_{t-h}$. Anchor $\bm{v}_{t-h}$ is cached for reuse in the next interval.
• Figure 4: Qualitative comparison of generated samples across different solvers and NFEs. We visualize random samples generated by our BA-solver and baseline methods (Flow-DPM, Euler, and Flow-UniPC solvers) at 3, 5, 7, and 10 NFEs.
• Figure 5: Performance comparison on ImageNet.(a)-(b) FID scores varying with NFE on ImageNet-$256^2$ and ImageNet-$512^2$. Our BA-Solver achieves superior performance compared to baselines. (c) Visual samples generated by BA-Solver on class-conditional ImageNet-$512^2$ with only 7 NFEs.