DEMON
open-source real-time music diffusion engine, 25Hz local GPU
Last verified:
What is DEMON?
DEMON (Diffusion Engine for Musical Orchestrated Noise) is a real-time diffusion engine that transforms the AI music denoising process into a playable musical instrument. Built on ACE-Step 1.5 and StreamDiffusion's ring-buffer architecture with TensorRT acceleration, it enables real-time streaming music generation at approximately 100 milliseconds latency - fast enough to play like a live instrument. The engine exposes denoising parameters as live performance controls, allowing musicians, DJs, and producers to remix songs and loops interactively.
Key features include real-time remixing of songs, denoise/structure/timbre strength adjustment, reference track swapping, prompt blending with parameter scheduling curves, LoRA hotswapping with runtime strength adjustment, per-frame source blending, MIDI controller mapping to any knob, vocal stem cutting/pasting with melformer, and XL model support. The engine sustains up to 12.3 decoder completions per second for 60-second music on a single consumer GPU (RTX 5090), or 11.3 generations per second at production ring-depth of 4. It provides six per-frame control curves at 25 Hz for continuous frame-level control of music diffusion.
DEMON is designed for DJs and producers who want to transform source loops into raw material for live performance, as well as developers and researchers working on continuous frame-level control of music diffusion. The tool is open-source and local-first, running on a single consumer GPU with NVIDIA CUDA support. A web-based live demo is available at music.daydream.live for users who want to try it without installing.
The engine includes specialized mechanisms like per-slot heterogeneous denoise scheduling, shared mutable per-step state for next-tick parameter effect, and windowed VAE decode exploiting receptive-field analysis for an 8x decode speedup. The author has also trained approximately 200 LoRA/Dora models for ACE-Step 1.5 and 1.5XL, released in batches. A VST plugin is currently in development.
DEMON pricing
Pricing model: Freemium
DEMON is completely free and open-source. It is a local-first application that runs on your own GPU at no cost. There is no paid tier or subscription required for the core tool. A web-based live demo is available at music.daydream.live for trying without installation. The code is available on GitHub at github.com/daydreamlive/DEMON.
DEMON pros
- Real-time response at approximately 100ms latency - playable like a musical instrument
- Open-source and free to use locally
- Runs on a single consumer GPU (RTX 5090 capable of 12.3 generations/second)
- TensorRT acceleration for optimized performance
- MIDI controller mapping to any parameter knob
- LoRA hotswapping with runtime strength adjustment
- Real-time remixing of entire songs or continuous loop remixing
- Per-frame source blending for transformation-strength control
- Six per-frame control curves at 25 Hz for fine-grained control
- Reference track swapping during generation
- Prompt blending with parameter scheduling curves
- Vocal stem cutting/pasting with melformer integration
- Distilled DreamVAE for 8x faster decode speedup
- Web demo available at music.daydream.live without installation
- Approximately 200 community LoRA/Dora models available for ACE-Step 1.5
DEMON cons
- Not pure continuous music - optimized for remixing loops or entire songs
- ACEStep begins and ends songs, limiting pure autoggressive generation
- XL support less stable with VRAM pressure issues
- Requires NVIDIA GPU with CUDA support (not CPU-only optimized)
- Lyrics/vocals feature not yet implemented (marked SOON)
- Spectral quality improvements not yet implemented (marked SOON)
- Learning curve for real-time performance control parameters
- Local GPU hardware requirement may exclude some users
Frequently asked questions about DEMON
What is DEMON and what does it do?
DEMON (Diffusion Engine for Musical Orchestrated Noise) is a real-time diffusion engine that makes the AI music denoising process playable as a live musical instrument. It transforms the denoising parameters into live performance controls, allowing you to remix songs and loops in near real-time with approximately 100ms latency. Built on ACE-Step 1.5 and StreamDiffusion's ring-buffer architecture with TensorRT acceleration, it sustains real-time streaming music generation on a single consumer GPU.
What are the system requirements to run DEMON?
DEMON requires an NVIDIA GPU that supports CUDA for TensorRT acceleration. It is optimized for consumer GPUs like the RTX 5090, which can achieve 12.3 decoder completions per second for 60-second music. The engine runs locally on your machine and is local-first, meaning you need the hardware to run it rather than relying on cloud inference.
Is DEMON free to use?
Yes, DEMON is completely free and open-source. The code is available on GitHub at github.com/daydreamlive/DEMON. It runs locally on your own GPU at no cost. There is no subscription or paid tier required for the core functionality. A web-based demo is also available at music.daydream.live for users who want to try it without installing.
Can I control DEMON with MIDI controllers?
Yes, DEMON supports MIDI controller mapping to any knob. The engine exposes denoising parameters as live performance controls, and MIDI controllers can be mapped to any parameter for tactile real-time control during performance. This makes it suitable for live musical performance and DJ sets.
What is the latency/response time of DEMON?
DEMON responds in approximately 100 milliseconds, which is fast enough to actually play like a musical instrument. This low latency is achieved through TensorRT acceleration, the ring-buffer architecture, and specialized mechanisms like per-slot heterogeneous denoise scheduling and shared mutable per-step state that enable next-tick parameter effect.
Can DEMON generate songs with lyrics and vocals?
Not yet. According to the project documentation, lyrics/vocals functionality is marked as 'SOON' and is not currently implemented. The engine is currently optimized for remixing existing songs or continuously remixing loops, rather than generating complete songs with vocals from scratch.
What is LoRA hotswapping in DEMON?
LoRA hotswapping allows you to swap in different LoRA (Low-Rank Adaptation) models in real-time during generation, with runtime strength adjustment. The author has trained approximately 200 LoRA/Dora models for ACE-Step 1.5 and 1.5XL, which are released in batches and available on CivitAI. This lets you change music styles mid-generation without stopping.
Does DEMON support continuous music generation?
Not in the pure sense. DEMON is optimized for remixing either an entire song or continuously remixing a loop. ACEStep correctly 'begins' and 'ends' songs, so this is not pure continuous autogressive music generation. For continuous music generation, other tools like Autogression may perform better. The loop mode works fine for continuous remixing scenarios.
What is the per-frame source blending feature?
Per-frame source blending is a sampling-time control on the standard SDE re-noise step that provides a framewise transformation-strength axis. This complements scalar denoise scheduling and allows you to control how strongly the output transforms from the source audio at each frame. It is one of the four key mechanisms contributed by DEMON.
Is there a VST plugin available for DEMON?
A VST plugin is currently in development but not yet released. The project team has mentioned it as part of what's next for DEMON. In the meantime, you can use MIDI controllers directly with DEMON for real-time performance control, or use the web demo at music.daydream.live.