With Linjie Lyu, @vdeschaintre.bsky.social, Yannick Hold-Geoffroy, Milos Hasan, Jae Shin Yoon, Thomas Leimkühler, Christian Theobald.

Code will be released shortly.

With no task-specific training, we show state-of-the-art results on material editing, object removal & insertion, relighting.

Our method can preserve identity of unedited image aspects, and thanks to editing in intrinsic-image space, it provides unprecedented control.

We build on the RGB↔X framework and address its two key limitations of identity drifts and channel entanglement, to unlock its image-editing potential.

Editing appearance, geometry, lighting with precision is easy with a 3D scene representation. But it's so much more difficult with just an image or photograph.

Enter IntrinsicEdit: Precise generative image manipulation in intrinsic space (SIGGRAPH 2025)!

intrinsic-edit.github.io