Skip to content

Overview

FlashDreams runs a growing family of world and video models (text-to-video, image-to-video, camera-controlled, ...). Model overviews identify currently supported models.

Running a model

All models run via our flashdreams-run-v2 CLI command 

cd flashdreams

# Almost all models download their own model weights from Hugging Face, please set `HF_TOKEN` to avoid rate limiting.
export HF_TOKEN=<your-hugging-face-token>

uv run flashdreams-run-v2 <Demo Preset> <System Arguments> -- <Demo Arguments>

  • <Demo Preset> is a preset that contains the model/demo pair to run. Refer to the model-overviews below to find a preset.

  • <System Arguments> are arguments that apply to all flashdreams-run-v2 demos, reference is here: System arguments.

  • <Demo Arguments> are arguments that apply to a particular demo. Refer to a particular demo overview.

Models - Streaming and autoregressive

These models advance a video or world incrementally for responsive, stateful generation.

NVIDIA OmniDreamsHDMap-conditioned streaming world model for autonomous driving.
Self-ForcingAutoregressive text-to-video based on Wan 2.1.
Causal-ForcingReal-time causal text- and image-to-video generation.
Causal Wan 2.2Autoregressive video generation based on Wan 2.2.
LingBot-WorldCamera-controllable streaming 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.

Models - Bidirectional generation

These models render a complete clip in one pass and provide full-sequence reference implementations.

SANA-WMFull-sequence camera-controlled image-to-video generation.
Wan 2.1Text-to-video and image-to-video generation.
Wan 2.2 TI2V-5BText-and-image-to-video generation in one full-clip rollout.
Cosmos-Predict2.5Text2World, Image2World, and Video2World generation.

Models - Upscaling and restoration

These models enhance an existing video stream rather than generating a scene from scratch.

FlashVSROne-step streaming video super-resolution.
SwiftVRReal-time streaming video restoration with 2× and 4× presets.

Adding your own model

Follow Create a model for the model package, then Integrate model adapter for a demo to register a runnable application.

  • Follow Get Started for the shortest path to running a model on your own hardware.
  • The API guides cover the architecture behind the models you can run today.
  • Community lists the channels to use if a process on this page does not run on your hardware.
  • Browse the source on GitHub at NVIDIA/flashdreams.