Kosta Derpanis (sabbatical in Zurich)#932
1/ Presenting ๐ ๐๐๐2๐๐๐๐ง๐ at #CVPR2026! ๐
๐ Poster Session: ExHall A & F 117 ๐ June 5, Friday โฐ 4:00 PM โ 6:00 PM MDT
@TristanAA97 will present the poster. If you are interested in image restoration, diffusion models, or face-guided restoration, please stop by!
8/ This work was done during my internship at @samsungresearch.
Grateful to all my amazing collaborators: Maitreya Suin, @TristanAA97, Sina Honari, Amanpreet Walia, Iqbal Mohomed, @CSProfKGD, @Babak_Taati, and Alex Levinshtein.
7/ Results show that ๐ ๐๐๐2๐๐๐๐ง๐ improves full-scene restoration over strong baselines, with better perceptual quality and more consistent restoration across the face, body, and background.
2/ In ๐ ๐๐๐2๐๐๐๐ง๐, we ask a simple question:
Can a restored face tell us how to restore the entire scene? ๐ง
3/ Reference-based face restoration can recover high-quality facial details, but it often leaves the body and background degraded.
For human-centric images, restoring only the face is not enough.
6/ Our key contributions:
๐น Facial degradation as an oracle ๐น FaDeX for face-derived degradation extraction ๐น MapNet for multi-scale diffusion conditioning ๐น InScene, a benchmark for human-centric restoration with identity references
4/ ๐งฉ Our key idea is to treat the face as a degradation oracle.
We restore the face using identity references, extract degradation cues from the LQ/HQ face pair, and use them to guide full-scene restoration. โจ
5/ ๐ We evaluate in both synthetic and real settings:
๐น Synthetic scenes with controlled degradations ๐น Real images with synthetic degradations ๐น Real captured images with natural degradations This helps test whether face-derived cues generalize beyond controlled benchmarks.
9/ Project page: https://amirhossein-kz.github.io/face2scene/
Paper: https://arxiv.org/abs/2603.16570
#ComputerVision #ImageRestoration #DiffusionModels #FaceRestoration #CVPR2026