Original title: Fill My Mirror: Geometry-Constrained Mirror Inpainting
Ofek Basson, Shimon Vainer, Yacov Hel-Or, Ohad Fried · September 2026
Generative image models struggle to draw realistic mirrors. A mirror's reflection changes based on the viewer's angle and the shape of the surrounding room, so models cannot just guess what should appear. Researchers tackled this problem by keeping the room fixed and only filling in the mirror. Their idea is simple: much of what a mirror shows is already visible in the rest of the picture. First, they estimate the geometry of the room. Next, they project parts of the visible room into the mirror. Finally, a diffusion model finishes the reflection using a two-mask strategy that balances these real geometric projections with its learned rules. This training-free method cuts down on visual errors and creates more consistent reflections in complex real-world scenes.
Imagine arranging furniture in a room and trying to predict how it would look from a mirror on the wall. Normally, you'd have to guess based on memory, but if you know the exact positions and angles, you can calculate what the mirror should show. In this paper, the system estimates the room's layout first, then uses that to determine the reflection, rather than letting the AI guess everything from scratch.
This matters for photographers and digital artists who need realistic mirror reflections in edited images. Without geometry constraints, generated mirrors often look distorted or inconsistent, breaking the scene's realism.
The abstract doesn't explicitly list limitations, but the method is limited to the mirror inpainting setting with a fixed scene. It may not handle dynamic environments or require accurate geometry estimation.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Fill My Mirror: Geometry-Constrained Mirror Inpainting — Ofek Basson, Shimon Vainer, Yacov Hel-Or et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.