Daydream

Overview

Daydream generates original audio samples in real time, from sound you already have. Feed DreamSampler a riff, a vocal, or a field recording and it produces new royalty-free material with synth-style control — chop it, perform it, and drop it into a track without touching a sample library or clearing rights.

The distinction it draws is against batch-mode tools: rather than writing a prompt and waiting for a render, you move a control and hear the result in roughly 81ms. The model responds to a performance instead of replacing it.

What you can build

• Original samples and stems — generate new material from your own audio rather than splitting what is already there
• Performed, not prompted — map the model's latent space to faders, knobs, and MIDI hardware, with control propagating at 25Hz
• Sessions inside a DAW — a native VST3 plugin automates generation curves from standard envelope lanes and macro knobs
• Agentic and generative workflows — an MCP server and API expose every parameter for programmatic control

Developer surface

• DEMON — the diffusion-native audio engine behind every Daydream instrument, open source and self-hostable, with sub-100ms parameter latency on a single consumer GPU
• Supported models — ACE-Step v1.5 (2B turbo and 5B XL turbo) today, with Stable Audio 3.0 and Magenta RT2 announced
• Runs anywhere — a browser app needing no install or GPU, a VST3 plugin, and a TouchDesigner operator, all on hosted infrastructure
• Scope — Daydream's real-time video project, also open source, for running generation from any source locally or in the cloud

Powered by Livepeer

Daydream's hosted inference is powered by Livepeer's GPU network. Workloads are routed to independent orchestrators, giving builders elastic, cost-efficient real-time inference without a centralized provider.