MeGAS: Thermomechanical Dynamic Gaussian Splatting for Thermophysical Scene Editing¶

Conference: ECCV2026
arXiv: 2606.23455
Code: To be confirmed
Area: 3D Vision
Keywords: 3D Gaussian Splatting, thermomechanical simulation, physical simulation, scene editing, phase change

TL;DR¶

MeGAS integrates thermodynamic phase-change dynamics (heat conduction, convection, melting/solidification) into the 3D Gaussian Splatting framework, enabling the reconstructed real-world scenes to undergo controllable, physics-based thermomechanical deformation, while maintaining rasterization quality under extreme large deformations through a topology-adaptive rendering strategy.

Background & Motivation¶

Combining physical simulators with neural rendering has emerged as a key trend in recent years. Works such as PhysGaussian and Gaussian Splashing have successfully introduced mechanically-driven dynamics—such as elasticity and fluids—into 3DGS. This allows reconstructed scenes to not only be viewed but also deform under physical laws, generating realistic dynamic content. However, existing methods focus primarily on mechanical processes and neglect a ubiquitous yet invisible physical quantity: temperature. Thermodynamic phenomena such as melting, solidification, and heat conduction are among the most common sources of deformation in nature. Yet, because they couple heat transfer with mechanical responses and involve extreme non-rigid topological changes (such as surface cracking and interior exposure) during phase transitions, they have remained excluded from physical simulation pipelines in neural rendering.

The core difficulties behind this omission stem from two aspects. First, modeling complexity: Thermodynamic dynamics require not only conserving momentum but also simultaneously handling convective-diffusive heat transfer and temperature-controlled material phase changes. The stress-strain constitutive models of solid and molten states are radically different and must be dynamically switched at the particle level. Second, rendering stability: 3DGS and its physical extensions (such as PhysGaussian) assume that geometry undergoes continuous or smooth deformation. However, extreme large deformations during the melting process can stretch Gaussian kernels into needle-like shapes, crack surfaces, and expose redundant internal Gaussians as floaters, causing naive combinations to fail immediately.

The core idea of this work is to assign a temperature attribute to each Gaussian particle in 3DGS, drive the temperature field evolution using a lattice-based thermal advection-diffusion solver (LBM), and trigger the transition from an elastic solid to a viscoplastic fluid constitutive model based on temperature thresholds, thereby achieving physically consistent melting simulation within the MPM framework. Concurrently, a topology-adaptive rendering strategy is designed—incorporating regularization during training, ray-tracing-based interior filling, and implicit-surface-guided adaptive densification—to ensure the output of crack-free, floater-free, and high-quality images under extreme deformation.

Method¶

Overall Architecture¶

The overall pipeline of MeGAS consists of three stages. First, a 3DGS representation of the scene is reconstructed from the input color image sequence. Unlike standard 3DGS, anisotropic regularization and low-contribution Gaussian pruning are incorporated during reconstruction to yield a uniform Gaussian model free of internal floaters. Second, each Gaussian is assigned a temperature attribute. The temperature field is initialized according to a user-specified heat source region, and the thermal advection-diffusion equation is iteratively solved on an MPM-coupled grid using the Lattice Boltzmann Method (LBM) to evolve the temperature distribution. Each particle dynamically switches its constitutive model between an elastic solid model (StVK) and a viscoplastic model (Herschel–Bulkley) based on whether its temperature exceeds the melting point threshold, thereby driving mechanical deformation. Finally, at deformed timesteps, cracking regions are detected via ray tracing, and the Gaussian kernel parameters of the interior filling particles are reconstructed guided by an implicit moving least squares (IMLS) surface fitted from the deformed Gaussians, restoring complete geometric and appearance rendering.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Color Image Sequence"] --> B["Reconstruction: Floater-free<br/>Uniformly Distributed 3DGS"]
    B --> C["Thermodynamic Initialization<br/>User-specified Heat Source + Temp Assignment"]
    C --> D{"MPM per Timestep"}
    D --> E["Grid LBM Solver<br/>Thermal Advection-Diffusion Equation"]
    E --> F["P2G -> Temp Field Update -> G2P"]
    F --> G{"Particle Temp &gt; Melting Point?"}
    G -->|Yes| H["Switch to Viscoplastic Constitutive<br/>(Herschel-Bulkley)"]
    G -->|No| I["Maintain Elastic Solid Constitutive<br/>(StVK)"]
    H --> J["Deformation Gradient Update<br/>Gaussian Covariance/Appearance"]
    I --> J
    J --> K{"Surface Cracking Appears?"}
    K -->|Yes| L["Implicit-Surface-Guided<br/>Adaptive Densification"]
    K -->|No| D
    L --> D
    D -->|After N Steps| M["Output: Rasterized Rendering<br/>Dynamic Temp Field/Appearance/Geometry"]

Key Designs¶

1. Thermomechanically Dynamic Gaussian Representation: Attaching Temperature Attributes to Each Gaussian Particle

Standard 3DGS only stores position, covariance, opacity, and spherical harmonic color parameters. MeGAS additionally stores a scalar temperature $T_i$ on each Gaussian. The temperature field can be projected to any viewpoint using the exact same alpha-blending formula used for color rendering, outputting a thermal map. This introduces crucial editing flexibility: Users can directly "place" heat source regions in the scene—for instance, marking a high-temperature zone with a brush. MeGAS initializes the temperature field accordingly and allows the heat to diffuse autonomously. The renderability of temperature information (visualizing temperature distribution) and its controllability (user-specified heat sources) form the foundation of physically controllable editing. The evolution of the temperature field does not rely on heuristic rules but is driven by a physical solver on a grid.

2. Hybrid LBM-MPM Heat-Force Coupling and Phase-Change Constitutive Switching

Once the temperature field is initialized, MeGAS maps it onto simulation grids and utilizes the Lattice Boltzmann Method (LBM) to simultaneously solve the Navier-Stokes equations (determining the velocity field $\mathbf{u}$) and the advection-diffusion equations (determining the temperature field $T$). It introduces thermal buoyancy via the Boussinesq approximation (where hot regions rise and cold regions sink) to generate natural convection. The pipeline periodically executes the standard MPM sequence: P2G (particle-to-grid transfer of mass and temperature) $\rightarrow$ grid update (velocity, temperature boundaries, LBM diffusion) $\rightarrow$ G2P (grid-to-particle interpolation of temperature and velocity). After each G2P step, every particle checks if its temperature exceeds the material's melting threshold $T_{\mathrm{melt}}$. If it does not, stress is computed using an elastic solid model (StVK); if it does, it switches to a viscoplastic fluid model (Herschel–Bulkley) to calculate stress. Macroscopically, this manifests as "solid regions maintaining their shape while molten regions begin to flow." This physics-based phase-change control renders the melting process progressive, localized, and predictable—only high-temperature regions liquefy while low-temperature regions remain rigid, closely matching real-world melting behavior.

3. Topology-Adaptive Rendering Strategy: Training Regularization + Ray-Tracing Interior Filling + Implicit-Surface-Guided Densification

Extreme deformations during the melting process cause direct applications of PhysGaussian-like 3DGS to suffer from three severe artifacts: needle-like Gaussian kernels, exposed internal floaters, and surface cracking. To address this, MeGAS designs four sequentially acting mechanisms.

The training phase incorporates two key cleaning steps: (1) iterating over all training views to record the maximum rendering contribution weight $w_{\max}$ of each Gaussian across all pixels, and pruning low-contribution Gaussians below a threshold $\tau_{\mathrm{prune}}$ (which are typically redundant internal structures created during reconstruction that remain hidden in static renders but expose themselves as floaters during deformation); (2) incorporating an anisotropic regularization loss $\mathcal{L}_{\mathrm{aniso}}$ to penalize Gaussian kernels with excessively large ratios of maximum to minimum scale (preventing needle-like artifacts) while clamping the scale factor $\mathbf{S}_p \leftarrow \min(\mathbf{S}_p, \tau_{\mathrm{scale}})$ to maintain volume consistency.

Before deformation occurs, volumetric support is required because, after pruning internal Gaussians, the MPM requires particles to carry deformation computations. MeGAS employs a ray-tracing-based filling strategy: dividing the scene bounding box into uniform grids and casting rays along primary axes. If the accumulated opacity of a ray indicates that a grid point lies entirely within the physical volume (i.e., starting from the interior and aligning with the Gaussian normal), the grid point is populated with an internal MPM particle. Compared to density-field-based filling (such as the scheme used in PhysGaussian), this strategy yields more uniform fills without generating hollow structures.

During deformation, cracking regions are filled by particles naturally migrating from interior zones. However, these particles lack valid covariance and appearance parameters. MeGAS detects particles reaching cracking regions at each timestep and fits a local implicit surface from neighboring original Gaussians via Implicit Moving Least Squares (IMLS). The particles are then projected onto this surface to obtain surface normals and rotation matrices, while their colors are averaged from $K$-nearest original Gaussians, thereby endowing these filling particles with surface-aligned, appearance-consistent Gaussian kernel parameters. This mechanism smoothly "sews" together cracked regions without the need for expensive per-frame retraining.

A Complete Example: Statue Melting¶

Taking a plaster statue in a real-world scene as an example: The user first uses a brush to mark a "heat source region" on the head of the statue, setting the temperature to 300°C and the material's melting point to 200°C. MeGAS initializes the temperature field: head particles are set to 300°C, and the rest remain at an ambient temperature of 25°C. At the first timestep, the LBM solver calculates heat conduction from the head down the neck, establishing a temperature gradient; simultaneously, thermal buoyancy induces weak natural convection in the surrounding air. When the neck temperature exceeds the 200°C threshold, the neck particles switch from the StVK elastic solid state to the Herschel–Bulkley viscoplastic state. Under the influence of gravity and mass, they begin to deform downwards. The deformation gradient updates the Gaussian covariance, producing a downward sagging effect for the entire head. As deformation intensifies, micro-cracks appear on the surface of the neck. Once ray tracing detects these cracked areas, IMLS fits the current implicit surface of the head, projects nearby filling particles onto the surface to acquire correct normals and colors, and sews up the cracks. Consequently, every rendered frame preserves photorealistic, multi-view consistent physical behaviors.

Loss & Training¶

The overall optimization objective combines the standard RGB rendering loss of 3DGS with anisotropic regularization: $$\mathcal{L} = \mathcal{L}_{\mathrm{RGB}} + \lambda_{\mathrm{aniso}} \mathcal{L}_{\mathrm{aniso}}$$ where the anisotropic loss penalizes any Gaussian whose ratio of maximum to minimum scale exceeds a threshold $a$. The training of the entire pipeline is split into two phases: The first phase incorporates regularization and pruning during static scene reconstruction to obtain a floater-free, uniformly distributed Gaussian model (approx. 30k steps). In the second phase, this model is loaded, and thermomechanical animation is executed, during which each timestep only requires localized fine-tuning via adaptive densification (approx. 10 seconds per scene) without global optimization.

Key Experimental Results¶

Main Results¶

The paper performs melting style editing on real-world scenes (MipNeRF360, Tanks and Temples datasets) and compares MeGAS with three baselines:

Scene	Method	Physical Realism	Visual Realism	Multi-View Consistency
Average User Rating (Multi-Scene)	DGE	1.538	2.038	2.154
Average User Rating (Multi-Scene)	AutoVFX	2.692	2.154	2.385
Average User Rating (Multi-Scene)	Runway Gen-4.5	2.038	2.154	1.885
Average User Rating (Multi-Scene)	MeGAS	3.731	3.654	3.577

A user study (20+ participants ranking anonymous videos, with a maximum score of 4) demonstrates that MeGAS significantly HTML-outperforms other methods across physical realism, visual realism, and multi-view/temporal consistency.

Ablation Study¶

Configuration	Effect	Description
Full MeGAS	No cracks, no floaters, uniform volume	Full method
w/o Internal Prune	Internal floaters exposed after deformation	Pruning of low-contribution Gaussians
w/o Anisotropy Loss	Needle-like Gaussian kernels cause rendering streaks	Unconstrained max/min scale ratio
w/o Scale Clamping	Inconsistent volume, unnatural deformation	Large discrepancy in Gaussian volumes
w/o Adaptive Densification	Surface cracking cannot be repaired	Failure of implicit surface guidance
w/ Density Field Filling	Uneven internal filling, prone to collapse	Alternative is inferior to ray-tracing filling
w/o Phase-Change Switching	All Gaussians are initially in molten state, early global collapse	Temperature threshold control is critical
w/o Phase-Change (Heat Diffusion Only)	Temperature field evolves normally but no deformation is produced	No melting without thermal-mechanical phase-change coupling

Key Findings¶

Anisotropic regularization and low-contribution pruning are the two most critical components of the entire topology-adaptive strategy; removing them leads to the most severe human-perceptible artifacts in deformation rendering.
Phase-change switching (handling temperature threshold controls rather than global melting) is central to physical plausibility: without it, the scene collapses globally immediately after simulation starts, destroying the visual effect of progressive melting.
Under different combinations of heat source locations and temperatures (varying thermal diffusivity and viscosity), MeGAS can simulate diverse melting modes ranging from "rapid collapse" to "slow dripping," demonstrating excellent controllability.
The framework naturally supports bidirectional phase changes (heating $\rightarrow$ melting $\rightarrow$ removing heat source $\rightarrow$ cooling $\rightarrow$ solidification), proving that it is not restricted solely to the melting phenomenon.

Highlights & Insights¶

Introducing Temperature as a First-Class Citizen in 3DGS: Unlike previous works that only leverage thermal infrared images during reconstruction or employ simplified thermal diffusion estimation, MeGAS treats temperature as an explicitly renderable, editable, and mechanically-deformable physical quantity, opening up an entirely new dimension of editing.
Clever Choice of the LBM Solver: The combination of MPM and LBM allows the same set of grid points to simultaneously solve fluid mechanics (NSE) and thermal advection-diffusion (ADE), avoiding the reinvention of complex coupled solvers. Additionally, the inherent parallelism of LBM facilitates subsequent GPU acceleration.
Three-stage "Training Pruning $\rightarrow$ Filling $\rightarrow$ Adaptive Post-Deformation Sewing" Strategy for Extreme Large Deformation: Instead of trying to perform global optimization during deformation, this approach removes instabilities (internal Gaussians and needle-like kernels) from the data during the training phase, actively provides volumetric support in the pre-filling phase, and performs only localized minimal stitching in the deformation phase. This "prevention-first, local-repair" mindset is highly practical.
IMLS Implicit Surface Sewing Scheme: Projecting internal filling particles onto an implicit surface fitted from the deformed Gaussians to re-orient Gaussian kernel parameters is elegant and efficient (requiring no extra training data and using pure geometric computation).

Limitations & Future Work¶

Material appearance changes induced by phase transitions are currently unmodeled: When melting, an object's surface color and reflectivity (PBR attributes) change significantly, but the spherical harmonic colors in MeGAS are fixed during static reconstruction and do not update dynamically during phase changes. The authors suggest that combining the model with video diffusion models (e.g., Stable Video Diffusion, Wan) could enhance the final appearance rendering.
The model heavily relies on the physical parameter settings of the MPM simulator (thermal diffusivity, viscosity, melting point, etc.), presenting a barrier for non-expert users. There is currently a lack of means to automatically estimate these parameters.
The adaptive densification at each timestep requires ray tracing and IMLS fitting, which incurs heavy computational overhead (typically taking several seconds per frame as reported in the paper) and remains far from real-time interaction.
Currently, only melting is demonstrated as the primary showcase for thermophysical editing; broader scenarios such as welding, evaporation, and combustion remain to be explored.

vs PhysGaussian: PhysGaussian first combined elasto-plastic MPM with 3DGS, but it only considers mechanically driven deformations and assumes minor deformations. MeGAS builds on this by introducing temperature attributes and phase-change constitutive switching. It establishes handling extreme large deformations as its core design goal and successfully resolves PhysGaussian's failure under large deformations through its topology-adaptive rendering strategy.
vs Gaussian Splashing / FieryGS: The former focuses on fluid dynamics (liquid splashing), while the latter keys in on fire and smoke synthesis. MeGAS instead addresses solid-liquid phase transitions, which is a temperature-driven physical process, thus belonging to a different physical domain.
vs Thermal3D-GS / NTR-Gaussian / ETGS: These works reconstruct temperature fields under thermal infrared views using 3DGS but are limited to static scenes or heat conduction without phase changes. Their focus is "reconstruction" rather than "editing." MeGAS "drives scene deformation using temperature," which is fundamentally physics-simulation-driven content generation.
vs DGE / AutoVFX / Runway Gen-4.5: Editing tools based on diffusion models or LLM pipelines inherently lack physical realism (video generation lacks multi-view consistency and physical constraints). In contrast, MeGAS achieves unparalleled physical consistency and multi-view stability through underlying explicit physical simulations.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ [First to combine thermodynamic phase transitions with 3DGS; both problem formulation and design schemes exhibit strong originality]
Experimental Thoroughness: ⭐⭐⭐⭐ [Comprehensive evaluation on real and synthetic scenes alongside user studies, though lacking quantitative comparisons with SOTA methods (due to the baselines' inability to support thermodynamic editing, leaving only user studies)]
Writing Quality: ⭐⭐⭐⭐⭐ [Clear motivation, detailed methodology description, and thorough ablation studies. The progressive ablation visualization in Fig.9 is exceptionally intuitive]
Value: ⭐⭐⭐⭐⭐ [Paves a brand-new direction for physics-integrated neural rendering—namely, temperature-driven modeling—marking a significant step towards a truly comprehensive world model of "editable physical worlds"]