SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing

TL;DR

SplitFlow addresses inversion and gradient-entanglement challenges in text-guided image editing by introducing an inversion-free framework that decomposes a target prompt into sub-prompts, learns independent editing flows, and then composes them via Latent Trajectory Projection and Velocity Field Aggregation. The method reduces gradient conflicts while preserving diversity, achieving better semantic fidelity and background consistency on PIE-Bench compared to state-of-the-art zero-shot editors. A theoretical justification shows that VFA yields improved alignment and stability over simple averaging. Empirically, SplitFlow demonstrates stronger target alignment (CLIP) and background preservation (PSNR/SSIM) than prior methods, with a fidelity-enhanced variant trading some diversity for higher reconstruction quality.

Abstract

Rectified flow models have become a de facto standard in image generation due to their stable sampling trajectories and high-fidelity outputs. Despite their strong generative capabilities, they face critical limitations in image editing tasks: inaccurate inversion processes for mapping real images back into the latent space, and gradient entanglement issues during editing often result in outputs that do not faithfully reflect the target prompt. Recent efforts have attempted to directly map source and target distributions via ODE-based approaches without inversion; however,these methods still yield suboptimal editing quality. In this work, we propose a flow decomposition-and-aggregation framework built upon an inversion-free formulation to address these limitations. Specifically, we semantically decompose the target prompt into multiple sub-prompts, compute an independent flow for each, and aggregate them to form a unified editing trajectory. While we empirically observe that decomposing the original flow enhances diversity in the target space, generating semantically aligned outputs still requires consistent guidance toward the full target prompt. To this end, we design a projection and soft-aggregation mechanism for flow, inspired by gradient conflict resolution in multi-task learning. This approach adaptively weights the sub-target velocity fields, suppressing semantic redundancy while emphasizing distinct directions, thereby preserving both diversity and consistency in the final edited output. Experimental results demonstrate that our method outperforms existing zero-shot editing approaches in terms of semantic fidelity and attribute disentanglement. The code is available at https://github.com/Harvard-AI-and-Robotics-Lab/SplitFlow.

Figures (7)

• Figure 1: Simplified visual illustration of the proposed SplitFlow. Purple, Blue, Orange line indicates independent sub-target flow. The aggregation is done on a certain timestep. After the aggregation, we use a single, unified flow. The notation in this figure follows the paper.
• Figure 2: Qualitative comparison of prompt-based image editing methods. Each row corresponds to a specific editing instruction, where the source prompt is modified into a target prompt. From top to bottom, the tasks are: (a) change “plant” to “flower”, (b) change “dumpling” to “sushi”, (c) change "red and white" to "blue and green", (d) add "with hat", (e) delete "a single pink lotus flower is growing in the middle of". The columns show the input image and the results generated by different models, including Directinv+P2P, RFsolver, FireFlow, FlowEdit-Flux, FlowEdit-SD3, and SplitFlow.
• Figure 3: Qualitative comparison results with more complex prompts.
• Figure S1: Detailed pipeline of SplitFlow. Given sub-target prompts, we define independent, decomposed flows. When the decomposition phase ends at timestep $\eta_{dec}$, we apply Latent Trajectory Projection (LTP) to obtain the projected sub-target latents. The combined velocity field $\bar{v}_t^{\Delta}$, computed via Velocity Field Aggregation (VFA), is used to update the projected target latent. The remaining flow is computed using only the target prompt.
• Figure S2: Qualitative comparison of prompt-based image editing methods. Each row corresponds to a specific editing instruction, where the source prompt is modified into a target prompt. From top to bottom, the tasks are: (a) change “goat” to “horse”, (b) change “spaceship” to “eagle”, (c) add "dog bone", (d) chage "teacup" to "cake", (e) change "boat" to "blanket", (f) chage "pink background" to "pink hearts background", (g) change "mountain" to "garden". The columns show the input image and the results generated by different models, including Directinv+P2P, RFsolver, FireFlow, FlowEdit-Flux, FlowEdit-SD3, and SplitFlow$^\dagger$.