Inversion-free editors such as FlowEdit are fast and naturally structure-preserving — they edit by following the velocity difference between a source and a target trajectory, with no inversion and no reconstruction drift. But they consistently fall short on global attribute changes, such as recoloring an entire scene or shifting a material.
We trace the cause to the high-noise early steps of generation. There, the flow that transports the sample toward the natural-image manifold dominates the comparatively weak text-conditioned direction, so the trajectory stays tied to the source and the global edit never accumulates. By the time the noise level drops and the prompt regains influence, the coarse layout and color are already fixed.
Our goal follows directly: restore the semantic guidance precisely in that early window, in a form that reshapes global structure without disturbing fine detail — and do so on a pretrained model, with no training and no attention surgery.
The figure contrasts one editing step of FlowEdit (a) with ours (b). Both start from the same source latent xtsrc.
FlowEdit follows a single direction. It forms one geometric residual Δvt from the source and target velocity fields and steps along it. Because Δvt compares two different latent points, under a global edit its direction is governed by the manifold-seeking flow rather than by the prompt — so the sample barely deviates from the source.
We add a cleaner source of guidance. Probing the same latent xtsrc under the source vs. the target prompt yields a pure semantic direction vsem: it isolates what the prompt change alone wants to do, with the spatial confound removed.
We keep only its low-frequency part. vsem still carries high-frequency stochastic noise, so a low-frequency wavelet operator Flow retains only its coarse component vlow — the global color, material, and layout — and discards the noisy detail that would corrupt fine structure.
We inject it where it matters, when it matters. The final update combines the geometric term Δvt with vlow, weighted by a time factor that grows with the noise level t (overall strength λ): strong in the early steps where the edit is weakest, fading to zero near the end so detail is left untouched — and reducing exactly to FlowEdit when λ = 0. The bottom row traces the resulting path, smoothly turning the scene (flowers → dog) while the background layout stays in place.
Qualitative results on PIE-Bench. Each pair shows the Source image (left) and Ours (right).
Source | Ours
a woman and a horse → a man and a horse
Source | Ours
a round shape cake with orange frosting on a wooden plate → a star shape cake with orange frosting on a wooden plate
Source | Ours
a panda bear close his mouth → a panda bear open his mouth
Source | Ours
four eggs with drawings on them are in a nest → three eggs with drawings on them are in a nest
Source | Ours
woman with brown hair → woman with blue hair
Source | Ours
an older man sitting on a bench with a dog → an older man sitting on a bench
Source | Ours
black chair in a conference room → blue chair in a conference room
Source | Ours
a chocolate cake with candies on top → a icecream cake with candies on top
@inproceedings{tang2026wavelet,
title = {Wavelet-Guided Semantic Signal Compensation for Inversion-Free Image Editing},
author = {Tang, Anqi and Sun, Wenhao and Liu, Zhaoqiang},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026}
}