Vanguard Studio

An AI-Native Production Platform for the "Unified Scene" Era.

Vanguard Studio is a prototype production environment that unifies live footage, generative video (Wan 2.2), 3D Gaussian Splatting, real-time physics, and semantic segmentation (SAM 3) into a single interactive workspace.

Project Goal

To build an AI-first, real-time 3D "Unified Scene" application that eliminates "round-trip" workflows by treating all media types as first-class, interactive citizens.

Architecture & Technology Choices

We implement a dual-stack architecture to balance performance with iteration speed.

1. Native Stack (Performance Reference)

• Language: C++20
• UI Toolkit: Qt 6 (QML/C++) for native performance and dockable panels.
• Renderer: Vulkan (primary) for low-level control and compute shaders; OpenGL fallback.
• Physics: Nvidia Flex (via plugin) or PhysX.
• Justification: Essential for high-fidelity real-time rendering, handling large 3DGS scenes, and minimizing GC overhead during complex physics simulations.

2. Hybrid Stack (Rapid Iteration)

• Shell: Electron
• UI: React + Tailwind CSS
• 3D Viewport: Three.js (WebGL2)
• Justification: Allows for rapid UI prototyping, easier access to web-based AI APIs, and a lower barrier to entry for contributors familiar with web technologies.

3. Inference Server

• Stack: Python 3.11 + FastAPI + gRPC
• Role: Hosts heavy AI models (SAM 3, Wan 2.2, PickStyle) outside the main UI process to prevent freezing and allow for remote deployment.

Directory Structure

• ui/ - Electron + React hybrid application.
• native/ - C++ Qt + Vulkan reference implementation.
• core/ - Shared schemas (.vscene) and logic.
• inference/ - Python AI model server.
• plugins/ - Integrations for Blender, Unity, etc.
• docs/ - Architecture and planning documentation.
• demo_data/ - Sample assets and acceptance tests.

Setup & Build

Prerequisites

• Node.js v18+
• CMake 3.20+
• Qt 6.5+
• Python 3.11+
• Vulkan SDK
• CUDA Toolkit (for inference)

Hybrid (Electron)

cd ui
npm install
npm start

Native (C++)

cd native
mkdir build && cd build
cmake ..
make
./VanguardNative

Inference Server

cd inference
pip install -r requirements.txt
python server.py

License & Ethics

• License: MIT (Code), CC-BY-NC (Demo Assets).
• Model Usage: Users must adhere to the licenses of the underlying models (Wan 2.2, SAM 3).
• Telemetry: Opt-in only.
• Content Provenance: Exports include metadata tags indicating AI generation tools used.