[Paper Review] Grow with the Flow: 4D Reconstruction of Growing Plants with Gaussian Flow Fields

Weihan Luo, Lily Goli+4 more|arXiv (Cornell University)|Feb 9, 2026

Plant Molecular Biology Research0 citations

TL;DR

GrowFlow models plant growth as a continuous 3D Gaussian flow controlled by a neural ODE, enabling coherent, time-elongated reconstruction of growing plants from multi-view timelapse data. It introduces reverse-growth training with 3D Gaussians and a spatio-temporal HexPlane encoder to predict growth dynamics.

ABSTRACT

Modeling the time-varying 3D appearance of plants during growth poses unique challenges: unlike most dynamic scenes, plants continuously generate new geometry as they expand, branch, and differentiate. Existing dynamic scene representations are ill-suited to this setting: deformation fields provide insufficient constraints to yield physically plausible scene dynamics, and 4D Gaussian splatting represents the same physical structures with different Gaussian primitives at different times, breaking temporal consistency. We introduce GrowFlow, a dynamic representation that couples 3D Gaussian primitives with a neural ordinary differential equation to model plant growth as a continuous flow field over geometric parameters (position, scale, and orientation). Our representation enables consistent appearance rendering and models nonlinear, continuous-time growth dynamics with full temporal correspondences for every primitive. To initialize a sufficient set of Gaussian primitives, we first reconstruct the mature plant and then learn a reverse-growth process, effectively simulating the plant's developmental history in reverse. GrowFlow achieves superior image quality and geometric coherence compared to prior methods on a new, multi-view timelapse dataset of plant growth, and provides the first temporally coherent representation for appearance modeling of growing 3D structures.

Motivation & Objective

• Motivate accurate 4D reconstruction of growing plants where new geometry emerges over time.
• Propose a continuous-time dynamical system for plant growth using 3D Gaussian primitives.
• Enable introduction of new structures over time via reverse-growth and a learned flow field.
• Maintain monotonic growth assumptions while achieving high photometric and geometric accuracy.
• Demonstrate state-of-the-art performance on synthetic and real multi-view timelapse plant data.

Proposed method

• Represent the scene with 3D Gaussian splats whose geometric parameters evolve over time.
• Model the time-varying evolution with a velocity field F_phi that governs Gaussian centers, rotations, and scales.
• Use a spatio-temporal HexPlane encoder plus MLP decoders to predict per-Gaussian velocities.
• Train via a multi-stage procedure including boundary reconstruction, piecewise backward integration, and global optimization.
• Adopt reverse-growth: learn to shrink Gaussians backward in time and then reverse the flow to obtain forward growth.

Figure 2 : Method overview. a) Our method first optimizes a set of 3D Gaussians on the fully-grown plant. b) Using the optimized 3D Gaussians from the fully-grown plant, we progressively train the dynamics model to learn the state of the plant at each timestep. After each reconstructed timestep, we

Experimental results

Research questions

• RQ1 Can a continuous-time Gaussian flow model capture non-rigid, topology-changing growth of plants across time?
• RQ2 Does reversing growth to shrink Gaussians facilitate differentiable optimization and stable training?
• RQ3 How does GrowFlow perform in terms of geometric fidelity and photometric quality compared to state-of-the-art baselines on synthetic and real plant timelapse data?
• RQ4 Is a spatio-temporal HexPlane encoding essential for capturing growth trajectories over time?

Key findings

• GrowFlow achieves geometrically coherent growth trajectories with high temporal consistency across seen and interpolated timesteps.
• GrowFlow outperforms baselines in both geometric accuracy (Chamfer Distance) and image quality (PSNR, SSIM, LPIPS) on synthetic data.
• On captured real-world scenes, GrowFlow yields superior novel-view renderings and stable interpolation where baselines struggle.
• A HexPlane-based spatio-temporal encoder and the boundary reconstruction stage are crucial for performance; replacing them degrades results.
• The reverse-growth training enables continuous introduction of new geometry without non-differentiable additions.

Figure 3 : Results on synthetic data. We compare results on both seen and interpolated times. GrowFlow achieves stable, coherent geometry, unlike prior methods that show visually correct RGB but inconsistent deformations. Yellow marks interpolated frames.