Fixing flow models when multiple constraints fight each other

Flow models can be steered at inference time to satisfy multiple constraints—like "realistic image" and "matches this description"—without retraining. The problem: when constraints contradict or push in different directions, the model generates unrealistic outputs by drifting off the learned data manifold. The authors identify that gradient misalignment causes this drift and propose Conflict-Aware Additive Guidance, which dynamically detects conflicting constraint signals and reconciles them during generation. Tests across image editing, planning, and control tasks show it maintains fidelity while satisfying multiple goals simultaneously.