FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing

Though Rectified Flows (ReFlows) with distillation offers a promising way forfast sampling, its fast inversion transforms images back to structured noisefor recovery and following editing remains unsolved. This paper introducesFireFlow, a simple yet effective zero-shot approach that inherits the startlingcapacity of ReFlow-based models (such as FLUX) in generation while extendingits capabilities to accurate inversion and editing in 8 steps. We firstdemonstrate that a carefully designed numerical solver is pivotal for ReFlowinversion, enabling accurate inversion and reconstruction with the precision ofa second-order solver while maintaining the practical efficiency of afirst-order Euler method. This solver achieves a 3times runtime speedupcompared to state-of-the-art ReFlow inversion and editing techniques, whiledelivering smaller reconstruction errors and superior editing results in atraining-free mode. The code is available athttps://github.com/HolmesShuan/FireFlow{this URL}.