Researchers compared two ways to tweak image attributes in pretrained neural backbones. The first uses a noisy image, inverts it to a hidden state, edits that state, then reconstructs an image. The second skips the hidden state entirely. Instead it moves the real image in small steps guided by differences in the model’s velocity field, so the image always stays on the training manifold.

On CelebA, the methods that perform inversion – SD3‑inv and FlowEdit – hit higher F1 scores on three of four attributes. However, the inversion‑free FlowEdit preserves identity better, showing lower LPIPS scores and a smaller visual shift from the original image. The experiment used a frozen 2‑Billion‑parameter SD3 backbone, the same for both pipelines, isolating the transport step.