Summary

DEMON is a streaming diffusion engine for real-time audio generation, built upon ACE-Step v1.5. It targets technical users needing continuous, low-latency audio synthesis with dynamic control. The engine delivers bit-identical streaming output, allowing parameters to be modified on-the-fly for interactive audio creation.

How It Works

DEMON employs a streaming architecture with a ring buffer advancing multiple in-flight generations per tick. End-to-end TensorRT acceleration optimizes DiT decoder and VAE performance. It supports per-frame parameter modulation (scalars/curves) that are hot-mutable mid-stream, alongside a dynamically resizable ring buffer depth for flexible throughput.

Quick Start & Requirements

Clone the repo, run uv sync, then uv run demon-setup to install dependencies, download ACE-Step v1.5 checkpoints (~18 GB), and build TensorRT engines. Launch the web demo via uv run python -u -m demos.realtime_motion_graph_web.run. Requires an NVIDIA GPU (RTX 3090/4090/5090 tested), ~40 GB disk space, and Node.js 20+ (for web demo). Full details in docs/INSTALL.md. Hosted instance: music.daydream.live.

Highlighted Details

• Bit-Identical Streaming: Output quality matches batch processing via ring buffer and step-advancement.
• Hot-Updatable Controls: Parameters like velocity, guidance, LoRA strength, and pipeline depth are dynamically modifiable mid-generation.
• Heterogeneous Slots: Simultaneously manages multiple generation slots with independent parameters for complex blended outputs.
• End-to-End TensorRT: Accelerates DiT decoder (refit-enabled for LoRA) and VAE, including a fixed 1s windowed VAE decode for low latency.

Maintenance & Community

Originally created by Ryan Fosdick, maintained by Daydream Live and contributors. Builds upon the MIT-licensed ACE-Step codebase. Specific community channels are not detailed.

Licensing & Compatibility

Distributed under AGPL-3.0-or-later, requiring network-distributed modifications to be open-sourced. ACE-Step components retain their MIT license. Commercial use is possible if AGPL network terms are met.

Limitations & Caveats

Strict NVIDIA GPU requirement (RTX 3090/4090/5090 tested). TensorRT engine builds can be time-consuming and are specific to duration/architecture. The AGPL license's strong copyleft may impact closed-source commercial adoption over networks.