Openflows

Signal

Turn your raw video into a viewable 3D Gaussian Splat on macOS · opensourceprojects · 2026-03-15

Tool for converting raw video to 3D Gaussian Splat on macOS. GitHub: https://github.com/freddewitt/CorbeauSplat

Context

3D Gaussian Splatting (3DGS) has emerged as a high-fidelity alternative to NeRFs for scene reconstruction, offering real-time rendering capabilities. While early implementations required significant compute resources or specific hardware, recent efforts focus on consumer-grade accessibility. This signal identifies a macOS-specific implementation that prioritizes local inference and video-to-3D workflows, reducing dependency on cloud processing pipelines.

Relevance

This entry maps a specific infrastructure component for spatial data generation within the Openflows ecosystem. It supports the local-inference-baseline circuit by enabling local hardware to process complex 3D reconstruction tasks. It also feeds into embodied-ai-governance and distributed-physical-agent-infrastructure by providing the spatial mapping layer required for robot navigation and environment understanding without external API calls.

Current State

The tool is hosted on GitHub as CorbeauSplat by freddewitt. It targets macOS environments, leveraging Apple Silicon architecture for video processing and rendering. The workflow accepts raw video input and outputs interactive 3D Gaussian Splat assets. Documentation indicates a focus on developer and creator workflows rather than enterprise deployment, though the local execution model aligns with privacy-first infrastructure patterns.

Open Questions

• What are the specific hardware requirements (RAM/GPU) for stable operation on consumer Mac hardware?
• Is there an API or CLI interface for programmatic integration with agent orchestration frameworks?
• How does the reconstruction quality compare to cloud-based photogrammetry services in terms of fidelity and latency?
• Are there plans for multi-view or SLAM integration to support continuous mapping?

Connections

• vjepa-meta: Both entries address video-based spatial understanding. While V-JEPA focuses on predictive world models from video frames, CorbeauSplat focuses on explicit 3D scene reconstruction. Together they represent complementary approaches to video-to-3D abstraction.
• dimensionalos: This framework connects LLM agents to robot control primitives. 3D Gaussian Splatting provides the dense spatial data required for navigation and manipulation in physical environments, serving as a potential input layer for dimensionalos workflows.