SNR-Edit: Structure-Aware Noise Rectification for Inversion-Free Flow-Based Editing

TL;DR

SNR-Edit tackles the problem of preserving structural fidelity during inversion-free flow-based editing by addressing the Structural--Stochastic Mismatch that arises when using a fixed Gaussian proxy. It introduces a structure-aware prior constructed from semantic region decomposition, RoPE-based geometric encoding, and a fixed randomized projection to produce a latent prior $\Phi_{\mathcal{Z}}$, then rectifies the latent trajectory with $\tilde{\epsilon} = \lambda_{\text{struct}} \Phi_{\mathcal{Z}} + \lambda_{\text{stoch}} \xi$ and a re-anchored velocity evaluation. The framework provides a theoretical Lipschitz-based bound on trajectory deviation and demonstrates, across SD3 and FLUX backbones on PIE-Bench and SNR-Bench, that SNR-Edit yields superior structural fidelity and semantic alignment with negligible overhead (approximately 1s per image) while remaining training-free and model-agnostic. These results highlight the importance of trajectory engineering via structure-aware initialization for robust, editable flow-based generation in real-world scenarios.

Abstract

Inversion-free image editing using flow-based generative models challenges the prevailing inversion-based pipelines. However, existing approaches rely on fixed Gaussian noise to construct the source trajectory, leading to biased trajectory dynamics and causing structural degradation or quality loss. To address this, we introduce SNR-Edit, a training-free framework achieving faithful Latent Trajectory Correction via adaptive noise control. Mechanistically, SNR-Edit uses structure-aware noise rectification to inject segmentation constraints into the initial noise, anchoring the stochastic component of the source trajectory to the real image's implicit inversion position and reducing trajectory drift during source--target transport. This lightweight modification yields smoother latent trajectories and ensures high-fidelity structural preservation without requiring model tuning or inversion. Across SD3 and FLUX, evaluations on PIE-Bench and SNR-Bench show that SNR-Edit delivers performance on pixel-level metrics and VLM-based scoring, while adding only about 1s overhead per image.

Figures (9)

• Figure 1: Qualitative visualization of SNR-Edit results. Our method demonstrates robust capabilities across a diverse range of tasks, including Add Object, Delete Object, Change Object, Change Style, Change Attribute, Change Background, and Change Pattern. Compared to the inversion-free method FlowEdit, SNR-Edit achieves superior structural fidelity, preserving the non-edited layout more faithfully than the baseline while precisely executing the text instructions.
• Figure 2: Schematic comparison of latent editing dynamics.(a) Inversion-based methods rely on a bidirectional mapping to recover latent noise, yet face an inherent fidelity-editability trade-off and high sensitivity to perturbations, making it difficult to maintain structural consistency. (b) FlowEdit circumvents inversion but is hindered by a fundamental Structural--Stochastic Mismatch. By initiating the flow from content-agnostic Gaussian noise $\xi \sim \mathcal{N}(0, I)$, the source proxy is evaluated off the source latent manifold, causing the differential editing trajectory to drift and leading to structural distortion. (c) Ours. We introduce structure-aware noise rectification by modulating Gaussian noise with a structural prior $\Phi_{\mathcal{Z}}$ (prepared by resizing, min--max normalization to $[-1,1]$, and channel-wise broadcasting), forming a rectified noise $\tilde{\epsilon} = \lambda_{\text{struct}} \Phi_{\mathcal{Z}} + \lambda_{\text{stoch}} \xi$. This yields a corrected latent source state $\tilde{Z}^{\text{src}}_t = (1 - t) Z_{\text{src}} + t \tilde{\epsilon}$, which anchors flow integration and enables a rectified velocity evaluation that preserves structural fidelity while reflecting target semantics.
• Figure 3: SNR-Edit Pipeline (cf. Alg. \ref{['alg:nlc']}).Phase 1 (Top) constructs a structural prior by extracting semantic masks $\mathcal{M}$, computing geometric signatures $\mathbf{s}_k$, and forming a single-channel structural map $\Phi_{\text{map}}$ via a fixed randomized projection $\psi$. This map is then resized to the latent resolution, normalized to $[-1,1]$, and broadcast across latent channels to obtain $\Phi_{\mathcal{Z}}$. Phase 2 (Bottom) executes rectified flow integration in latent space. Starting from $Z_{t_{\max}}^{\text{FE}} = Z_{\text{src}}=\mathcal{E}(X_{\text{src}})$, the dynamics iteratively update $Z^{\text{FE}}$ using a rectified velocity field driven by $\tilde{\epsilon}=\lambda_{\text{struct}}\Phi_{\mathcal{Z}}+\lambda_{\text{stoch}}\xi$, which mixes $\Phi_{\mathcal{Z}}$ and Gaussian noise $\xi$. This mechanism encourages the output $X_{\text{tar}}=\mathcal{D}(Z^{\text{FE}}_0)$ to preserve the source layout while realizing the target semantics.
• Figure 4: Visualization of image editing results on PIE-Bench. Our method demonstrates superior performance across both FLUX and SD backbones, producing images that better preserve structural details, maintain accurate text-image correspondence, and achieve higher overall visual quality compared to existing approaches.
• Figure 5: Sensitivity analysis of the stochastic scale $\lambda_{\text{stoch}}$. We observe a trade-off between structural consistency (SSIM) and text alignment (CLIP), identifying 0.9 as the optimal balance for text-guided image editing.