NAISS
SUPR
NAISS Projects
SUPR
3D Consistent Multi-Vew Image Editing
Dnr:

NAISS 2026/4-620

Type:

NAISS Small

Principal Investigator:

Yaroslava Lochman

Affiliation:

Chalmers tekniska högskola

Start Date:

2026-03-27

End Date:

2027-04-01

Primary Classification:

20205: Signal Processing

Allocation

Abstract

This project aims to address 3D consistent image and video editing/generation and novel view synthesis. The primary focus is on leveraging the existing foundation image and video editing/generation models. Such models have been trained on extremely large datasets for extremely long times. Consequently, they are able to generate very realistic high-resolution images. However, for multiple views of the same scene, in general, geometric (and sometimes even semantic) consistency is not guaranteed. We are looking into ways to force this consistency without re-training or fine-tuning the large foundation models. We are investigating approaches like guidance of the denoising process, output verification, input augmentation, all with the help of different methods for 3D reconstruction.