Skip to content

FlashDreams

FlashDreams is an inference and serving runtime for turning world models into live, controllable simulations. FlashDreams runs models in a continuous loop, carrying state forward and streaming frames while new actions or sensor inputs change what happens next step.

Whether the application is a game world, an autonomous-vehicle simulator, robotic policy testing, or a virtual training environment, FlashDreams is the runtime to make it possible.

Get Started! GitHub Contribute
FlashDreams quick intro animation

Explore FlashDreams

Choose a section to get started or browse the project.

Get StartedInstall FlashDreams and run your first model. ModelsBrowse default supported models. DemosExplore our interactive experiences and model interfaces. DocumentationCheckout our CLI, Demo API, Inferencing API, tooling, or related guides. CommunityFind contribution documentation and support channels.

Why FlashDreams?

Serving World Models is Hard

Building a static clip generator is not enough for world-model serving.

Offline video generation makes one pass from prompt through model to clip. Online serving receives low-latency input from a user or policy, passes it through FlashDreams encoding, state update, model inference, and decoding, and feeds low-latency output back to the input continuously.

Worlds evolve over time

World models are taught to generate and evolve an environment over time. They support robotics-policy testing, autonomous-vehicle simulation, interactive worlds, and other applications.

World models support autonomous systems using sensor input, robotics policy testing, and interactive worlds with live controls.

Serving has two parts

  1. Inference of a video model that consumes inputs such as clock time and sensors while evolving world state.
  2. A separate application that coordinates inputs, model state, GPU inference, and output.
World-model serving has two parts. Model Inference consumes time and sensors to evolve world state. In Application Coordination, controls update world state, a connected upper path feeds world state back to the controls, the GPU produces stacked frames, those frames stream to the monitor, and the monitor also feeds back to the controls.

FlashDreams was built for both parts:

  1. The Inferencing API helps model designers implement stateful world-model inference.
  2. The Demo API helps application developers build experiences around the constraints of world-model serving.
Model Inference maintains state and cache for GPU inference. The Demo API coordinates controls, sensors, and presentation. Actions and state flow to Model Inference; streamed output flow to the Demo API.

Performance

Each tile shows speedup over an existing implementation of the same model. Both run the same weights on the same GPU, so the gain comes from the FlashDreams runtime alone.
2.12× Self-Forcing speedup 3.10× LingBot-World speedup 1.40× Wan2.1 speedup 1.42× FlashVSR speedup



Try FlashDreams!

The Get Started guide walks from a fresh checkout to running OmniDreams, an interactive driving world-model demo built on FlashDreams.

Supported Models

Streaming and autoregressive model implementations emit per-step output with sub-second latency once warm; bidirectional model implementations are kept as full-block parity references. Each model page carries the canonical invocation, the checkpoint source, and the per-implementation knobs.
OmniDreamsInteractive world simulator for autonomous vehicles. Cosmos3 Nano Sim AutoDriving video with multiple cameras and optional LiDAR. Cosmos3 Nano Sim BimanualController-driven robot-manipulation video. Self-ForcingAutoregressive text-to-video based on Wan 2.1. Causal-ForcingAutoregressive text/image-to-video based on Wan 2.1. Causal Wan 2.2Autoregressive text-to-video based on Wan 2.2 from FastVideo. LingBot-WorldCamera-controllable image-to-video world model. Waypoint 1.5Interactive world model controlled by keyboard and mouse. SANA-WMCamera-controlled world model with streaming and bidirectional variants. FlashVSRStreaming video super-resolution. SwiftVRReal-time streaming video restoration with 2× and 4× presets. Wan 2.1Bidirectional text-to-video and image-to-video generation. Wan 2.2 TI2V-5BBidirectional text-and-image-to-video generation. Cosmos-Predict2.5Bidirectional Text2World, Image2World, and Video2World generation.