3D Consistent Novel View Synthesis

SUPR uses JavaScript for certain functions. We cannot guarantee that you will be able to use the system with JavaScript disabled.

Dnr:

NAISS 2026/4-646

Type:

NAISS Small

Principal Investigator:

Josef Bengtson

Affiliation:

Chalmers tekniska högskola

Start Date:

2026-05-01

End Date:

2027-05-01

Primary Classification:

20205: Signal Processing

Webpage:

https://3d-consistent-editing.github.io/

Allocation

Mimer at C3SE: 2000 GiB
Alvis at C3SE: 800 GPU-h/month

Abstract

This project aims to find ways to improve the ability of Diffusion Models to generate 3D consistent novel views and perform 3D editing. We want to improve performance specifically for non-rigid 3D edits (such as object insertion/removal or other geometry changing edits), utilizing guidance and building on recent advancements in flow-matchning based editing methods and 3D generation pipelines.