Saved in:
Bibliographic Details
Main Authors: Kong, Xin, Watson, Daniel, Strümpler, Yannick, Niemeyer, Michael, Tombari, Federico
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.06579
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911142788464640
author Kong, Xin
Watson, Daniel
Strümpler, Yannick
Niemeyer, Michael
Tombari, Federico
author_facet Kong, Xin
Watson, Daniel
Strümpler, Yannick
Niemeyer, Michael
Tombari, Federico
contents Multi-view diffusion models have shown promise in 3D novel view synthesis, but most existing methods adopt a non-autoregressive formulation. This limits their applicability in world modeling, as they only support a fixed number of views and suffer from slow inference due to denoising all frames simultaneously. To address these limitations, we propose CausNVS, a multi-view diffusion model in an autoregressive setting, which supports arbitrary input-output view configurations and generates views sequentially. We train CausNVS with causal masking and per-frame noise, using pairwise-relative camera pose encodings (CaPE) for precise camera control. At inference time, we combine a spatially-aware sliding-window with key-value caching and noise conditioning augmentation to mitigate drift. Our experiments demonstrate that CausNVS supports a broad range of camera trajectories, enables flexible autoregressive novel view synthesis, and achieves consistently strong visual quality across diverse settings. Project page: https://kxhit.github.io/CausNVS.html.
format Preprint
id arxiv_https___arxiv_org_abs_2509_06579
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CausNVS: Autoregressive Multi-view Diffusion for Flexible 3D Novel View Synthesis
Kong, Xin
Watson, Daniel
Strümpler, Yannick
Niemeyer, Michael
Tombari, Federico
Computer Vision and Pattern Recognition
Multi-view diffusion models have shown promise in 3D novel view synthesis, but most existing methods adopt a non-autoregressive formulation. This limits their applicability in world modeling, as they only support a fixed number of views and suffer from slow inference due to denoising all frames simultaneously. To address these limitations, we propose CausNVS, a multi-view diffusion model in an autoregressive setting, which supports arbitrary input-output view configurations and generates views sequentially. We train CausNVS with causal masking and per-frame noise, using pairwise-relative camera pose encodings (CaPE) for precise camera control. At inference time, we combine a spatially-aware sliding-window with key-value caching and noise conditioning augmentation to mitigate drift. Our experiments demonstrate that CausNVS supports a broad range of camera trajectories, enables flexible autoregressive novel view synthesis, and achieves consistently strong visual quality across diverse settings. Project page: https://kxhit.github.io/CausNVS.html.
title CausNVS: Autoregressive Multi-view Diffusion for Flexible 3D Novel View Synthesis
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2509.06579