Flow Matching (FM) models have become a leading approach for high-fidelity generative synthesis, but their reliance on iterative Ordinary Differential Equation (ODE) solving creates a significant latency bottleneck, according to a new paper on arXiv. Existing solutions face a trade-off: training-free solvers degrade at low Neural Function Evaluations (NFEs), while training-based one- or few-step methods incur high training costs and lack plug-and-play versatility. The researchers — Chen, Hongxu, Li, Hongxiang, Wang, Zhen, and Long — propose the Bi-Anchor Interpolation Solver (BA-solver) to bridge this gap.
How BA-Solver Works
BA-solver retains the versatility of standard training-free solvers while achieving significant acceleration by introducing a lightweight SideNet (just 1-2% of the backbone size) alongside the frozen backbone. The method rests on two synergistic components:
- Bidirectional Temporal Perception: The SideNet learns to approximate both future and historical velocities without retraining the heavy backbone.
- Bi-Anchor Velocity Integration: The sideNet uses two anchor velocities from the backbone to efficiently approximate intermediate velocities for batched high-order integration.
By using 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-256²
The paper reports that BA-solver achieves generation quality comparable to 100+ NFE Euler solver in just 10 NFEs and maintains high fidelity in as few as 5 NFEs, all with negligible training costs. This represents an order-of-magnitude speedup for generative models like Flow Matching.
| Solver Type | Steps (NFEs) | Quality vs. 100-step Euler | Training Cost | Plug-and-Play |
|---|---|---|---|---|
| Training-free solvers | 10 | Significant degradation | None | Yes |
| Training-based one-step | 1-2 | Comparable | Prohibitive | No |
| BA-solver (proposed) | 10 | Comparable | Negligible | Yes |
Advantages for Enterprise AI
According to the paper, BA-solver ensures seamless integration with existing generative pipelines, facilitating downstream tasks such as image editing. This plug-and-play nature means organizations can accelerate their current generative AI workflows without retraining large models. The minimal training overhead (only a 1-2% SideNet) makes it cost-effective for deployment.
The researchers note that BA-solver retains the versatility of training-free solvers while overcoming their performance degradation at low NFEs, making it a practical solution for latency-sensitive applications. As generative modeling becomes more central to enterprise operations—from product design to synthetic data generation—such acceleration techniques could directly reduce inference costs and time.
Broader Implications
The BA-solver approach addresses a fundamental bottleneck in generative AI: the speed-quality trade-off. By combining a frozen backbone with a tiny learnable side network, it achieves high performance without the expense of full model retraining. This pattern may inspire similar hybrid approaches in other iterative generative paradigms, such as diffusion models.
