Emergent Mind

Feature Splatting: Language-Driven Physics-Based Scene Synthesis and Editing

Published Apr 1, 2024 in cs.CV , cs.AI , cs.GR , and cs.LG


Scene representations using 3D Gaussian primitives have produced excellent results in modeling the appearance of static and dynamic 3D scenes. Many graphics applications, however, demand the ability to manipulate both the appearance and the physical properties of objects. We introduce Feature Splatting, an approach that unifies physics-based dynamic scene synthesis with rich semantics from vision language foundation models that are grounded by natural language. Our first contribution is a way to distill high-quality, object-centric vision-language features into 3D Gaussians, that enables semi-automatic scene decomposition using text queries. Our second contribution is a way to synthesize physics-based dynamics from an otherwise static scene using a particle-based simulator, in which material properties are assigned automatically via text queries. We ablate key techniques used in this pipeline, to illustrate the challenge and opportunities in using feature-carrying 3D Gaussians as a unified format for appearance, geometry, material properties and semantics grounded on natural language. Project website: https://feature-splatting.github.io/
Feature splatting pipeline combines images into a Gaussian representation encompassing scene's geometry, texture, semantics.


  • Feature Splatting combines 3D Gaussian primitives with semantics from vision-language models for enhancing scene synthesis.

  • Incorporates a particle-based physics simulator for dynamic manipulation grounded in natural language queries.

  • Improves scene editing and manipulation through algorithmic enhancements and a novel feature fusion method.

  • Extends applications in virtual reality, graphics, and digital content creation through intuitive, language-driven interaction.


The paper introduces Feature Splatting, a novel approach that bridges the gap between static 3D scene representations and dynamic scene synthesis through the incorporation of language-driven physics and semantics. Leveraging the power of 3D Gaussian primitives enhanced with rich semantic features distilled from large-scale vision-language models, this technique enables a more intuitive and semantically enriched interaction with 3D scenes. By integrating a particle-based physics simulator, Feature Splatting facilitates the seamless manipulation of physical properties and dynamic behaviors within a scene, all grounded in natural language queries. This advancement paves the way for more interactive and user-friendly applications in graphics, virtual reality, and beyond.

Key Contributions

The paper presents several significant contributions to the field of dynamic scene synthesis:

  1. Introduction of Feature Splatting: A method for augmenting static 3D scenes with semantics and physics, allowing for language-driven manipulation and editing of scene dynamics.

  2. Algorithmic and Systematic Enhancements: The development includes an adapted Material Point Method (MPM) physics engine for Gaussian-based representations and a novel feature fusion method utilizing features from multiple foundation models for accurate scene decomposition.

  3. Demonstration of Practical Applications: Through comprehensive experimentation and analysis, the authors illustrate the effectiveness of Feature Splatting as a tool for automatic, language-grounded scene editing.

Technical Insights

Feature Distillation and Scene Decomposition: The process starts with distilling high-quality, object-centric vision-language features into each 3D Gaussian. This step utilizes text queries to semi-automatically segment the scene, distinguishing between various objects and their components based on semantic attributes derived from linguistic inputs. The system leverages a combination of features from notable vision models such as CLIP, DINOv2, and SAM, addressing the challenges of low-resolution and noisy 2D feature maps through innovative pooling and refining techniques.

Physics-Based Dynamic Synthesis: Building on the static scene augmented with semantic features, the paper introduces a method to infuse physical dynamics through a particle-based simulation. Material properties are assigned semi-automatically via text queries, allowing for a nuanced control over the physical behavior (e.g., rigidity, elasticity) of objects within the scene. This unified approach to handling appearance, semantics, geometry, and physics in a single format significantly advances the capabilities for dynamic scene editing and synthesis.

Implications and Future Directions

Feature Splatting's integration of physics-based dynamics and rich semantic understanding opens up new avenues for creating highly interactive and responsive 3D environments. This research holds potential implications not only for enhancing virtual reality experiences and game development but also for advancing applications in robot vision, simulation-based training, and digital content creation.

Looking ahead, further exploration into refining the semantic-depth of the language models used, expanding the range of physical phenomena that can be simulated, and improving the efficiency of the feature distillation process could broaden the applicability and accessibility of this innovative approach. Additionally, integrating more advanced natural language processing capabilities might allow for even more nuanced and context-aware scene manipulations.


The introduction of Feature Splatting represents a significant step forward in the synthesis and editing of dynamic 3D scenes. By combining the strengths of physics-based simulation and vision-language models, this approach offers a powerful tool for creating richly interactive virtual environments. As this field continues to evolve, the convergence of language, vision, and physics holds the promise of unlocking unprecedented levels of creativity and immersion in digital content creation.

Get summaries of trending AI papers delivered straight to your inbox

Unsubscribe anytime.

Test Your Knowledge

You answered out of questions correctly.

Well done!