Consistent view synthesis without 3D

Paper Title:

Free3D: Consistent Novel View Synthesis without 3D Representation

Published on:

7 December 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Chuanxia Zheng,

Andrea Vedaldi

Bullets

Key Details

•

Encodes target views via distributed per-pixel ray conditioning

•

Improves multi-view consistency with attention & noise sharing

•

Avoids expensive 3D representations used in other works

•

Demonstrates excellent generalization to unseen categories

•

Outperforms recent state-of-the-art models on pose accuracy

Explore the topics in this paper

360-degree view synthesis

attention mechanisms

image-to-image translation

implicit 3d representations

novel view generation

AI generated summary

Consistent view synthesis without 3D

This paper introduces Free3D, a novel approach for generating consistent 360-degree novel views from a single input image, without needing slow and memory-intensive explicit 3D representations. It achieves this through a new per-pixel ray conditioning technique to precisely encode target camera poses, alongside multi-view attention and noise sharing for improved consistency.