FlowSteer: Guiding Few-Step Image Synthesis with Authentic Trajectories

TL;DR

FlowSteer addresses the sampling efficiency bottleneck in flow-based image synthesis by diagnosing training-inference misalignments in PeRFlow and introducing Online Trajectory Alignment (OTA) to ensure the teacher and student share authentic trajectories during distillation. It further strengthens guidance with adversarial distillation on the ODE trajectory and fixes a flaw in the FlowMatchEulerDiscreteScheduler, substantially boosting few-step performance. The approach demonstrates clear gains on Stable Diffusion 3 (SD3) backbones, outperforming competing methods at 4-step inference and beating several baselines in both quantitative metrics and qualitative image quality. By aligning training with on-policy trajectories and providing perceptual feedback, FlowSteer unlocks strong ReFlow-based distillation performance with practical efficiency benefits.

Abstract

With the success of flow matching in visual generation, sampling efficiency remains a critical bottleneck for its practical application. Among flow models' accelerating methods, ReFlow has been somehow overlooked although it has theoretical consistency with flow matching. This is primarily due to its suboptimal performance in practical scenarios compared to consistency distillation and score distillation. In this work, we investigate this issue within the ReFlow framework and propose FlowSteer, a method unlocks the potential of ReFlow-based distillation by guiding the student along teacher's authentic generation trajectories. We first identify that Piecewised ReFlow's performance is hampered by a critical distribution mismatch during the training and propose Online Trajectory Alignment(OTA) to resolve it. Then, we introduce a adversarial distillation objective applied directly on the ODE trajectory, improving the student's adherence to the teacher's generation trajectory. Furthermore, we find and fix a previously undiscovered flaw in the widely-used FlowMatchEulerDiscreteScheduler that largely degrades few-step inference quality. Our experiment result on SD3 demonstrates our method's efficacy.

Figures (5)

• Figure 1: Two key mismatches caused by off-trajectory training. (a) Teacher Trajectory Mismatch: The teacher's trajectory deviates from its own generation process. (b) Inter-Stage Distribution Mismatch: Linear interpolation introduces distribution mismatch between consecutive stages, leading to error accumulation.
• Figure 2: Trajectory Divergence between Authentic and Piecewise Paths. (Top) States from the teacher's authentic, continuous 32-step trajectory at key timesteps ($t=8, 16, 24, 32$). (Bottom) States from a piecewise trajectory, where each 8-step stage is re-initialized via linear interpolation. A visible divergence emerges at $t=16$ and grows.
• Figure 3: An overview of our proposed FlowSteer. (1) OTA online generates on-trajectory starting points for each stage by simulating the teacher's own process. (2) A discriminator combined by DiT backbone and discriminator head force the student's trajectory to mimic teacher's. (3) We build sub-trajectory basicly on our improved scheduler, for both teacher and student.
• Figure 4: Qualitative comparison with other methods at 4 steps(8 NFE for PCM and 4 NFE for others). Our model generates images with higher quality and better text-image consistency.
• Figure 5: Diverse high-quality images generated by our method using four NFE. These examples demonstrate our proposed method can produce aesthetically appealing and diverse results.