Curvature-Adaptive Consistency Flow Matching: Autonomous Trajectory Optimization via Reinforcement Learning

Modern large-scale diffusion models, such as FLUX and SDXL, rely on computationally expensive iterative sampling. While consistency distillation aims to accelerate this process, existing methods typically use static sampling schedules (e.g., uniform or Logit-Normal) that assume a fixed distribution of optimization difficulty. This paper investigates whether these static approaches are optimal and proposes a dynamic, geometry-aware alternative.

The authors analyze the consistency error along the Probability Flow ODE trajectory and discover a U-shaped difficulty profile, where the most significant optimization bottlenecks occur at the boundary stages (initialization and final refinement) rather than the middle. To address this, they introduce Curvature-Adaptive Consistency Flow Matching (CACFM). This framework treats distillation as a dynamic decision process, utilizing a lightweight Reinforcement Learning (RL) agent to probe the trajectory and prioritize sub-trajectories with higher optimization returns. The RL agent operates on a compact state space of loss rankings, enabling an emergent, coarse-to-fine curriculum without manual scheduling. This is paired with a hybrid objective function incorporating Distribution Matching Distillation (DMD) and adversarial consistency to ensure high-fidelity output.

CACFM demonstrates that replacing static heuristics with an adaptive RL scheduler significantly improves the quality of few-step generation. By intelligently allocating the computational budget to high-curvature regions, the model effectively mitigates structural deformities and preserves high-frequency details. The authors report new state-of-the-art results on large-scale models, showing that the RL-based approach successfully learns to focus on critical trajectory segments, thereby achieving superior visual fidelity compared to traditional uniform or heuristic-based distillation strategies.