Interactive serving#

FlashDreams serving keeps a world-model session alive while inputs and outputs stream through the application loop.

Serving models#

Live input
Application controls or sensor updates arrive continuously.
Warm session
Pipeline and cache state persist across updates.
Model step
Encoder, transformer, scheduler, and decoder advance the world.
Streamed output
Frames or latent output return without closing the session.

Reference integrations#

Serving implementation references#

  • Serving for serving API concepts and component mapping.

  • Inference pipeline overview for runner/pipeline execution flow.

  • integrations/lingbot/lingbot/webrtc for the WebRTC serving stack.