FlowBypass: Rectified Flow Trajectory Bypass for Training-Free Image Editing

TL;DR

FlowBypass tackles the persistent fidelity–alignment trade-off in training-free image editing by introducing a bypass that directly connects inversion and reconstruction trajectories within a Rectified Flow framework. It derives a principled, closed-form bypass term $b_t$ using a first-order Taylor expansion, enabling a practical Euler-discretized implementation that starts reconstruction at an intermediate state ${y}_{t_B}$ rather than from noisy inversion endpoints. The method is backbone-agnostic and eliminates the need for backbone-specific feature manipulations, achieving state-of-the-art results on EditEvalv2 with strong prompt alignment and preserved details in irrelevant regions. This approach significantly improves the practicality of training-free editing by reducing error accumulation while maintaining high semantic control, making it readily generalizable across diverse diffusion backbones and prompts.

Abstract

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.

Figures (16)

• Figure 1: Framework of FlowBypass. FlowBypass consists of three steps: (1) Inverse the input image ${x}_{t_0}$ with Euler discretization to obtain the inversion trajectory; (2) Calculate the bypass ${b}_{t_B}$ using inversion trajectory according to Equ. \ref{['equ:discret_form']}; (3) Reconstruct from intermediate state ${y}_{t_B} = {x}_{t_B} + {b}_{t_B}$ to obtain edited image ${y}_{t_0}$. We show the differences between the origin prompt and the edit prompt by marking removed words with strikethrough and added words with a gray background.
• Figure 2: Qualitative comparison with other image editing methods. Zoom in for a better view.
• Figure 3: Scatter figure of quantitative comparison. Green points denote the performance of FlowBypass, whereas blue points denote the performance of other methods. FluxSpace is omitted because its extremely poor performance would compromise the readability of the scatter plot.
• Figure 4: Visualization of bypass. Yellow regions indicate higher L1-norm values, while blue regions indicate lower values, reflecting the spatial distribution of bypass magnitude. Zoom in for a better view.
• Figure 5: Performance of different prompt choices. In each notation, the segment before the slash denotes inversion prompts and the segment after the slash denotes reconstruction prompts. Within each segment, the first character indicates the positive prompt and the second indicates the negative prompt, where "x" denotes $C_x$, "y" denotes $C_y$, and "e" denotes $\varnothing$.