FLAG-4D: Flow-Guided Local-Global Dual-Deformation Model for 4D Reconstruction

Guan Yuan Tan, Ngoc Tuan Vu, Arghya Pal, et al.

[AAAI 2026] | Paper | Project Page

Abstract

We introduce FLAG-4D, a novel framework for generating novel views of dynamic scenes by reconstructing how 3D Gaussian primitives evolve through space and time. Existing methods, often relying on a single MLP for temporal deformation, struggle with complex motions and fine-grained details. FLAG-4D overcomes this with a dual-deformation network: an Instantaneous Deformation Network (IDN) for fine-grained local deformations and a Global Motion Network (GMN) for long-range dynamics, synergistically refined via mutual learning. The GMN robustly integrates dense motion features from a pretrained optical flow backbone, which are temporally fused and aligned with each Gaussian's evolving state via a deformation-guided attention mechanism. Extensive experiments demonstrate that FLAG-4D achieves state-of-the-art performance, producing temporally coherent reconstructions with superior detail preservation.

Getting Started

Code will be released soon!

1. Installation

First, clone the repository:

git clone https://github.com/tgy1221/FLAG-4D.git
cd FLAG-4D