FlowBypass: Rectified Flow Trajectory Bypass for Training-Free Image Editing
This addresses the problem of efficient and generalizable image editing for users needing training-free methods, though it appears incremental by building on existing trajectory-based approaches.
The paper tackles the trade-off between error accumulation and prompt alignment in training-free image editing by proposing FlowBypass, a framework that constructs a bypass between inversion and reconstruction trajectories using Rectified Flow, resulting in improved prompt alignment and high-fidelity details as demonstrated in experiments.
Training-free image editing has attracted increasing attention for its efficiency and independence from training data. However, existing approaches predominantly rely on inversion-reconstruction trajectories, which impose an inherent trade-off: longer trajectories accumulate errors and compromise fidelity, while shorter ones fail to ensure sufficient alignment with the edit prompt. Previous attempts to address this issue typically employ backbone-specific feature manipulations, limiting general applicability. To address these challenges, we propose FlowBypass, a novel and analytical framework grounded in Rectified Flow that constructs a bypass directly connecting inversion and reconstruction trajectories, thereby mitigating error accumulation without relying on feature manipulations. We provide a formal derivation of two trajectories, from which we obtain an approximate bypass formulation and its numerical solution, enabling seamless trajectory transitions. Extensive experiments demonstrate that FlowBypass consistently outperforms state-of-the-art image editing methods, achieving stronger prompt alignment while preserving high-fidelity details in irrelevant regions.