DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing
Desong Yang, Mang Ye · 2026
Read DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing by Desong Yang, Mang Ye (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-02417.txt.
Description
Desong Yang and Mang Ye present DirectEdit, a training-free image editing method for flow-based text-to-image models that removes reconstruction error at its source. Editing usually adds noise to a clean image through inversion, then denoises along separate reconstruction and editing paths. Because the reconstruction path is approximated with noisy latents from mismatched timesteps, drift accumulates and limits fidelity.
The authors analyse inversion inside the flow transformer step by step. DirectEdit aligns the forward paths directly, rather than trying to repair the inversion path as most prior work does. That eliminates the inherent reconstruction error with no extra neural function evaluations. Step-level accuracy in inversion is the paper's central claim, aimed at edits that keep unedited regions faithful.
Catalog record
| CREATOR | Desong Yang, Mang Ye |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.02417