Minjun kang - GeoNVS (ECCV26)

GeoNVS: Geometry Grounded Video Diffusion

for Novel View Synthesis

Minjun Kang1 Inkyu Shin2 Taeyeop Lee1 Myungchul Kim1

1KAIST 2 Luma AI

European Conference on Computer Vision (ECCV, Spotlight Oral, Top 1.6%) 2026

Abstract

Novel view synthesis requires strong 3D geometric consistency and the ability to generate visually coherent images across diverse viewpoints. While recent camera-controlled video diffusion models show promising results, they often suffer from geometric distortions and limited camera controllability. To overcome these challenges, we introduce GeoNVS, a geometry-grounded novel-view synthesizer that enhances both geometric fidelity and camera controllability through explicit 3D geometric guidance. Our key innovation is the Gaussian Splat Feature Adapter (GS-Adapter), which lifts input-view diffusion features into 3D Gaussian representations, renders geometry-constrained novel-view features, and adaptively fuses them with diffusion features to correct geometrically inconsistent representations. Unlike prior methods that inject geometry at the input level, GS-Adapter operates in feature space, avoiding view-dependent color noise that degrades structural consistency. Its plug-and-play design enables zero-shot compatibility with diverse feed-forward geometry models without additional training, and can be adapted to other video diffusion backbones. Experiments across 9 scenes and 18 settings demonstrate state-of-the-art performance, achieving 11.3% and 14.9% improvements over SEVA and CameraCtrl, with up to 2x reduction in translation error and 7x in Chamfer Distance.

✨ Motivation

Geometry-based NVS struggles under sparse views, while generative NVS lacks explicit geometric constraints and often misaligns with the target view, motivating a geometry-grounded video diffusion model that combines the strengths of both.

Recent works condition generation on geometry priors to address this limitation, but only as input conditioning, leaving them vulnerable to color rasterization noise.

💡 Proposed Method

Avoid color rasterization errors through feature-space modulation. By operating in intermediate feature space rather than pixel space, GeoNVS mitigates view-dependent color artifacts of geometry priors that can compromise structural consistency in prior methods.
Adaptive fusion of geometry priors. GeoNVS adaptively balances geometry priors and generative cues, leveraging each where it is most reliable through our proposed fusion mechanism.
Plug-and-Play Design. The modular architecture supports zero-shot compatibility with various feed-forward geometry models (e.g., VGGT, DepthSplat, Pi3) without any additional training and can be adapted to different video diffusion backbones (e.g., SEVA, CameraCtrl).