Editing real images using a pre-trained text-to-image (T2I) diffusion/flow model often involves inverting the image into its corresponding noise map. The authors introduce FlowEdit, which is inversion-free, optimization-free and model agnostic. The method constructs an ODE that directly maps between the source and target distributions and achieves lower transport cost than inversion approaches, demonstrating state-of-the-art results with Stable Diffusion 3 and FLUX.

Diffusion and flow models enable editing real signals through pre-trained models. Traditional approaches rely on "inversion" — extracting the initial noise vector to generate edited signals with modified conditions like text prompts. However, the editing-by-inversion paradigm often leads to insufficient fidelity to the source image. The authors present FlowEdit, which breaks away from the editing-by-inversion paradigm. Rather than mapping images to noise and back, it constructs a direct path between the source and target distributions. This path is "shorter than that achieved in inversion, and thus maintains better fidelity to the source image".

Most zero-shot editing methods employ image-to-noise inversion as a first step. However, the noise map obtained through naive inversion is generally unsuitable for effective editing. Many methods focus on improving inversion accuracy or extracting "structural information implicitly encoded within the model architecture" and injecting it during sampling. These are typically tailored for specific model architectures, limiting transferability. Some methods use optimization over the image itself utilizing generative priors as loss terms. The authors' approach does not rely on inversion, maps from source to target distributions without traversing through the Gaussian distribution, avoids optimization, and does not intervene in the model internals.

The authors reframe inversion-based editing as a direct path between the source and target distributions. Given forward and reverse trajectories, they define: Z^inv_t = Z^src₀ + Z^tar_t - Z^src_t. Expressing this as an ODE: dZ^inv_t = V^Δ_t dt, where V^Δ = V^tar - V^src. Crucially, "images along this path are noise-free" because noisy images contain roughly the same noise constituent, so the difference vector encompasses the difference only between the clean image predictions. The path constitutes "a sort of autoregressive coarse-to-fine evolution" — early timesteps modify coarse image structures, while the last features to get updated are the fine textures.

Traditional inversion induces undesirable sample pairings between source and target distributions. FlowEdit introduces "many different random pairings, and averages the velocity fields corresponding to each". Rather than relying on the inversion-derived path, it employs an alternative forward process: Ẑ^src_t = (1-t)Z^src₀ + tN_t, where N_t is Gaussian noise. The method constructs the FlowEdit path by solving an ODE where the velocity field is an expectation over multiple noise realizations, creating "separated" modes rather than "intermixed" ones, reducing transport cost.

Practically, discrete timesteps are used, and the expectation is approximated by averaging n_avg model predictions at each timestep. An integer n_max determines the starting timestep for the process: when n_max equals the total steps, the full edit path is traversed and the strongest edit is obtained; when n_max is smaller, the first timesteps are skipped, weakening edits. The covariance function of the noise is chosen so noise becomes independent across timesteps. Rather than sampling source and target noisy images independently, they are "constructed with the same noise instance, which further improves robustness to small values of n_avg".

FlowEdit is evaluated using official weights of SD3 medium and FLUX.1 dev on a dataset of over 70 real 1024x1024 images with over 250 text-image pairs. Compared to ODE inversion, SDEdit, iRFDS, and RF-Inversion, FlowEdit demonstrates good structural preservation of the source image while simultaneously maintaining good adherence to the target text. Quantitatively, using LPIPS (structure preservation) and CLIP (text adherence), FlowEdit achieves a favorable balance in this tradeoff. A controlled experiment on 1000 model-generated images with ground-truth noise maps confirms lower transport cost (1376 vs. 2239 MSE; 0.15 vs. 0.25 LPIPS) and lower FID (51.14 vs. 55.88) compared to exact inversion.

FlowEdit provides an inversion-free, optimization-free and model agnostic method for text-based image editing using pre-trained flow models. The approach constructs a direct ODE between the source and target distributions without traversing standard Gaussian distributions. Evaluations show lower transport costs and thus stronger structure preservation, with state-of-the-art performance across various editing tasks. The authors acknowledge that FlowEdit relies on adding random noise, potentially producing diverse results depending on the random seed, which "can lead to suboptimal results that include artifacts" in some cases.