DEMON¶

Development line: project:demon · thread demon
Last event: 2026-06-03 · 1 dated since 2026-06-03 · Researched: 2026-09-06 · confidence: medium

What it is¶

DEMON turns ACE-Step v1.5 into a live music generator and transformer on GPU. It is an audio counterpart to StreamDiffusion for performers, creative-tool developers, and researchers.

• Ring-buffer diffusion streams finished audio while source audio, prompts, LoRAs, and curves change during playback.
• The web interface exposes prompt blending, timbre and structure references, automation, MIDI mapping, recording, and a Session API.

Development line¶

• 2026-06-03 — DEMON public resource set documented. The core project had no GitHub Releases record. The separately versioned demonTD TouchDesigner client released v0.2.11 and v0.2.12 for hosted-session and socket-protocol fixes.

What changed¶

• 2026-04-22: DreamVAE was published as a standalone ACE-Step decoder asset.
• 2026-05-27: the DEMON paper published the ring-buffer and TensorRT runtime with RTX 5090 measurements.
• 2026-06-02: DreamVAE was positioned as part of DEMON research while remaining a decoder-only asset.
• 2026-06-03: the core project had no GitHub Releases record; the separately versioned demonTD TouchDesigner client released v0.2.11 and v0.2.12 for hosted-session and socket-protocol fixes.
• 2026-06-04: demonTD v0.2.13 added an audio-output-device picker.
• 2026-06-09: demonTD v0.2.15 added automatic LoRA-trigger handling and prompt-B wiring. demonTD v0.2.16 moved pacing work off the UI thread to address choppy audio.
• 2026-06-11: demonTD v0.2.17 consolidated connection and audio-bleed fixes and documented bundle installation.

How to use this¶

1. For a no-install trial, open the hosted instance at music.daydream.live; it is the stated path for users without a local GPU. — https://huggingface.co/daydreamlive/DEMON
2. For a local web workflow, use an NVIDIA GPU, uv, and about 40 GB free disk. Clone daydreamlive/DEMON, run uv sync, and run uv run demon-setup to obtain supported ACE-Step v1.5 assets and build minimal TensorRT engines. — https://github.com/daydreamlive/DEMON
3. Launch uv run python -u -m demos.realtime_motion_graph_web.run and open http://localhost:6660. If setup skipped engines, use -- --accel compile rather than the default all-TensorRT launch. — https://github.com/daydreamlive/DEMON
4. Load source audio, prompts, and optional references in the web interface. Use prompt blending, strength, LoRAs, curves, MIDI, or recording according to the performance goal. — https://huggingface.co/daydreamlive/DEMON
5. Add compatible optional LoRAs as .safetensors files under $ACESTEP_MODELS_DIR/loras/ (default ~/.daydream-scope/models/demon/loras/) and adjust their strength in the live session. — https://huggingface.co/daydreamlive/DEMON

Best practices¶

• Keep the supplied ACE-Step v1.5 checkpoints and DEMON setup together; do not substitute arbitrary checkpoints or paths for a TensorRT configuration that expects the pinned model layout. — https://github.com/daydreamlive/DEMON
• Start with the TensorRT windowed VAE decoder plus a compiled decoder to avoid a full engine build; it is the documented low-cost route to streaming decode. — https://huggingface.co/daydreamlive/DEMON
• Rebuild TensorRT plans after changing GPU, CUDA, driver, or TensorRT version, because plans are hardware- and version-specific. — https://github.com/daydreamlive/DEMON
• Treat prompt, source, and scalar-denoise changes as ring-buffered controls: make changes deliberately and judge audio after propagation rather than expecting every control to be audible in the same tick. — https://daydreamlive.github.io/DEMON/
• If using DreamVAE separately, pair its decoder with the stock ACE-Step encoder; only the decoder is distilled. — https://huggingface.co/daydreamlive/DreamVAE
• For the TouchDesigner integration, install the release ZIP with demonTD.tox beside its vendor directory; the bare .tox is only for an installation that already has matching dependencies. — https://github.com/daydreamlive/demonTD

Superseded by this¶

• 2026-06-04: For demonTD, instructions to use Edit -> Preferences -> Audio -> Audio Device -> None are obsolete. demonTD v0.2.13 says that setting does not exist and provides an audio-output-device picker. Source: https://github.com/daydreamlive/demonTD/releases
• 2026-06-09: For demonTD, treating a selected LoRA as automatically effective without its trigger word is obsolete. demonTD v0.2.15 prepends stored primary triggers and re-sends the prompt when the LoRA changes. Source: https://github.com/daydreamlive/demonTD/releases

Still unknown¶

• No official core-engine GitHub Release identifies 2026-06-03 as a versioned DEMON release. We should not treat that date as a core-model or core-runtime release date.
• The Civitai URL https://civitai.com/models/2416425/acestep-loras did not provide a verifiable model card during research. Exact files, versions, dates, triggers, and compatibility with DEMON remain unverified; treat those LoRAs as optional external assets rather than a default dependency.
• The published core-code license is AGPL-3.0-or-later, while the Hugging Face DEMON card and project page display MIT. The sources do not reconcile that difference; confirm the applicable distribution license before modification, network deployment, or redistribution.
• The Daydream project page calls the XL Turbo base 5B, while the upstream ACE-Step repository calls its XL DiT 4B. The sources do not explain the parameter-count difference.
• Performance figures are author-reported RTX 5090 measurements under stated 60-second, TensorRT, depth, and step settings. No independent cross-hardware benchmark or universal latency and VRAM floor was found.

Sources¶

Agent brief¶

• Subject: project:demon, thread demon, 1 dated events 2026-06-03 → 2026-06-03.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.