Daydream

Overview

Daydream is an open-source, local-first platform for real-time generative AI video. Its flagship project, Daydream Scope, runs autoregressive video diffusion models on your own GPU — or in the cloud — and exposes a node-based workflow for composing live, interactive pipelines from text, video, and live camera inputs.

It's built for developers and creative technologists who want continuous, real-time AI video instead of asynchronous batch generation.

What you can build

• Real-time generative video pipelines — compose text-to-video, video-to-video, and live camera workflows from a node-based graph
• Interactive installations and live visuals — drive output reactively with audio, MIDI, OSC, or DMX
• Custom nodes and models — extend Scope with your own pipelines, LoRAs, and inference logic in Python
• Tooling integrations — share output to TouchDesigner, OBS, Resolume, Unity, Ableton Live, and other creative tools via Spout, Syphon, and NDI

Developer surface

• Self-hosted runtime — Scope runs locally as a server on localhost:8000 with a WebRTC-based API for programmatic, low-latency control
• Bundled diffusion models — ships with StreamDiffusion V2, LongLive, Krea Realtime, RewardForcing, and MemFlow
• Open-source Python codebase — fork it, build custom nodes, contribute pipelines back upstream
• Cloud inference — run pipelines on remote GPUs when local hardware isn't available

Powered by Livepeer

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