Target-aware Image Editing via Cycle-consistent Constraints

TL;DR

FlowCycle tackles the limitation of target-agnostic intermediate states in text-based image editing by introducing a target-aware, inversion-free framework built on flow matching. It learns two noises via cycle-consistent optimization to produce an intermediate state that selectively destroys editing-relevant content while preserving editing-irrelevant parts, enabling faithful edits and strong source consistency. The method achieves superior performance on PIE-Bench and DIV2K against multiple baselines, with robust qualitative and quantitative gains in semantic alignment and structural fidelity. This approach offers a practical pathway toward more controllable and reliable image editing using pre-trained flow-based models.

Abstract

Recent advances in pre-trained text-to-image flow models have enabled remarkable progress in text-based image editing. Mainstream approaches always adopt a corruption-then-restoration paradigm, where the source image is first corrupted into an ``intermediate state'' and then restored to the target image under the prompt guidance. However, current methods construct this intermediate state in a target-agnostic manner, i.e., they primarily focus on realizing source image reconstruction while neglecting the semantic gaps towards the specific editing target. This design inherently results in limited editability or inconsistency when the desired modifications substantially deviate from the source. In this paper, we argue that the intermediate state should be target-aware, i.e., selectively corrupting editing-relevant contents while preserving editing-irrelevant ones. To this end, we propose FlowCycle, a novel inversion-free and flow-based editing framework that parameterizes corruption with learnable noises and optimizes them through a cycle-consistent process. By iteratively editing the source to the target and recovering back to the source with dual consistency constraints, FlowCycle learns to produce a target-aware intermediate state, enabling faithful modifications while preserving source consistency. Extensive ablations have demonstrated that FlowCycle achieves superior editing quality and consistency over state-of-the-art methods.

Figures (16)

• Figure 1: Comparisons between target-agnostic and target-aware manner. (a) Target-agnostic corruption equally destroys global content, making it difficult to restore the destroyed editing-irrelevant content well. (b) Target-aware corruption emphatically destroys editing-relevant content, which better maintains editing-irrelevant content.
• Figure 2: Corruption-then-restoration paradigm for existing flow-based image editing methods. Existing methods obtain target-agnostic intermediate states for editing. In contrast, FlowCycle obtains target-aware intermediate states via cycle-consistency optimization.
• Figure 3: Framework of FlowCycle. We rectify the corruption process by optimizing two learnable noises (i.e., $\epsilon_{src}$ and $\epsilon_{tar}$) under the cycle-consistent constraints of whole editing. After optimization, a target-aware intermediate state $x_t^{src}$ can be obtained by adding $\epsilon_{src}$ to the source image $x_0^{src}$. Starting the restoration from this optimized $x_t^{src}$, the quality of the editing result $x_0^{tar}$ can be significantly improved.
• Figure 4: Qualitative comparisons on $512^2$ resolution PIE-Bench (left half) and $1024^2$ resolution DIV2K (right half). Compared with various baselines across different editing types (e.g., object or style change), FlowCycle can significantly maintain the editing-irrelevant parts of the source image while adhering to the editing target. Meanwhile, these two datasets cover different resolutions and editing complexity (e.g., DIV2K contains cases that involve changing multiple objects), indicating our robust effectiveness.
• Figure 5: Quantitative comparisons. The closer to the bottom right corner, the better the balance between Structure Distance (source consistency) and CLIP Score (semantic alignment). Ours can achieve the best balance across datasets and diffusion models.