README.md

---
title: Nobility v1 — Beryl AI Labs
emoji: 🎭
colorFrom: green
colorTo: yellow
sdk: gradio
sdk_version: 3.50.2
app_file: app.py
pinned: true
license: apache-2.0
hardware: zero-gpu
short_description: Qwen-Native Audio-Driven Avatar Pipeline
tags:
  - avatar
  - qwen
  - real-time
  - diffusion
  - digital-human
  - nobility-v1
  - beryl-ai-labs
---

# Nobility v1 — Beryl AI Labs

**The first end-to-end Qwen-native audio-driven avatar pipeline.**

*"We need to have a face to face."*

---

## What This Space Validates

This Space runs **Layers 1-3** of the Nobility v1 pipeline live:

| Layer | Model | Status |
|-------|-------|--------|
| L1 Brain | Qwen3 via NVIDIA NIM | ✅ Live |
| L2 Director | Qwen3 Shot Planner via NIM | ✅ Live |
| L3 Drift Scorer | Qwen3-VL Identity Lock via NIM | ✅ Live |
| L4 Engine | Wan2.2-S2V 4-step Causal Student | ⏳ Distillation Pending |
| L5 Decode | Stream-VAE + FlashVSR | ⏳ With L4 |

## How To Use

1. Get a free NIM API key at [build.nvidia.com](https://build.nvidia.com)
2. Paste it in the API key field
3. Upload an identity reference image (optional)
4. Talk to Amanda

## The Five Research Contributions

- **RC-01** Qwen-Native Closed-Loop Architecture
- **RC-02** Audio-First Causal Distillation (4-step student from 50-step teacher)
- **RC-03** Cross-Modal Drift Scoring — 0.3 Deviation Rule with melanin-weighted enforcement
- **RC-04** BitNet W1.58A8 + MoE Compound Weight Reduction
- **RC-05** Infinity-RoPE for indefinite avatar session length

## The Moat

Every competitor chains 4+ vendors. Nobility v1 runs one co-designed family.
Wan-S2V was trained using Qwen-VL captions. Wan2.2 ships with Qwen2.5 prompt extenders.
The family is designed to interlock. We're the first to run the full stack.

---

**Beryl AI Labs** · New Orleans · [tyronne-os/nobility-v1](https://github.com/tyronne-os/nobility-v1)