FlashDreams is a high-performance inference and serving library for interactive autoregressive video and world models. It began as the optimized runtime behind the NVIDIA OmniDreams closed-loop demo for GTC 2026 and has grown into a general platform for real-time world-model applications across gaming, autonomous vehicles, robotics, simulated or virtual environments, and more.

https://github.com/user-attachments/assets/2b000ce9-effe-4cc9-a227-5b4619413e4d

System Requirements

• NVIDIA GPU with 80 GB VRAM or more (e.g. H100 80GB), see notes below.
• NVIDIA driver from the R580 series or newer (default compatible with CUDA 13.x)
• CUDA 13.x by default (PyTorch 2.11.0+cu130 and the nvidia-*-cu13 libraries are resolved by uv sync. A system CUDA toolkit is needed only for the developer extras and is included in nvidia/cuda:13.2.1-cudnn-devel-ubuntu24.04)
• Python >= 3.10
• PyTorch >= 2.11.0+cu130 (>= 2.9 for bare PyPI library install)
• Linux x86-64 or arm64
• 100 GB+ free storage space recommended for environment and model checkpoints.
• Docker with the NVIDIA Container Toolkit (optional, only for the container workflow)

Development and testing were performed on GPUs with 80 GB of VRAM or more. Inference can fail (out-of-memory) on consumer and even enthusiast GPUs. Per-model GPU and VRAM requirements are listed on each model page in the model gallery.

Quickstart

The complete setup is in the installation guide. Assuming uv is installed, the shortest viable path is:

git clone https://github.com/NVIDIA/flashdreams.git
cd flashdreams
uv sync --extra runners
export HF_TOKEN=<your-hf-token>
uv run flashdreams-run --help

Note for developers/maintainers you would want to run uv sync --extra dev --extra runners instead.

Select CUDA Version

FlashDreams defaults to the CUDA 13 PyTorch stack (the cuda13 dependency group is activated automatically via default-groups):

• Linux: uv sync pulls torch>=2.9 from PyPI, whose Linux wheels are the CUDA 13 build (the +cu130 local-version tag is just stripped on PyPI).
• Windows: uv sync pulls torch from NVIDIA's cu130 index (https://download.pytorch.org/whl/cu130) — PyPI's Windows torch wheel is CPU-only.

CUDA 12.8 is available as an opt-in side profile (Linux only) via the cuda12 dependency group:

uv sync --group cuda12 --extra runners

The two groups are declared mutually exclusive in pyproject.toml, so uv deactivates the default cuda13 group automatically when --group cuda12 is passed.

Then launch your first model by following the Get Started guide. For example, the Self-Forcing T2V v2 application is:

uv run --project integrations_v2/self_forcing \
    flashdreams-run-v2 t2v-self-forcing-wan2.1-t2v-1.3b \
    --output-path artifacts/t2v-self-forcing-wan2.1-t2v-1.3b.mp4 -- \
    --prompt "A cat surfing." --total-blocks 7

You can also install FlashDreams as a library from PyPI:

pip install flashdreams

Try the interactive driving demo

Drive a world model in real time with the unified OmniDreams local-window or webrtc launch mode. See the interactive demo guide.

Supported models

FlashDreams ships first-party integrations under integrations_v2/. Each model has a dedicated docs page with runner slugs, multi-GPU commands, and (where available) profiling benchmarks.

See the model gallery and the new method guide to add your own.

Developer guides

• Architecture
• Inference pipeline overview
• Config system
• Add a new method

For day-to-day development:

uv sync --extra dev --extra runners
uv run --group lint pre-commit run -a
uv run pytest -m "not manual"

See DEV.md for repository-specific workflow notes.

Contributing

For how to contribute, see CONTRIBUTING.md. New integrations, bug reports, feature requests, performance tuning, and documentation edits are all welcome.

Use GitHub Issues to report defects or request improvements.

Join us on the NVIDIA Omniverse Discord to share your results and take part in technical discussion! Channel: #flashdreams

Security

To report a potential security vulnerability, follow the coordinated disclosure process in SECURITY.md.

License

FlashDreams is released under the Apache License 2.0. Third-party components and their licenses are listed in THIRD-PARTY-NOTICES and NOTICE. The repository is REUSE-compliant; see REUSE.toml and LICENSES/.

Citation

If FlashDreams is useful in your research or product, please cite the project:

@misc{flashdreams2026,
  title        = {FlashDreams: High-performance inference and serving for
                  interactive autoregressive video and world models},
  author       = {{FlashDreams Contributors}},
  year         = {2026},
  howpublished = {\url{https://github.com/NVIDIA/flashdreams}},
}

@misc{nvidia2026omnidreams,
  title         = {{NVIDIA} {OmniDreams}: Real-Time Generative World Model for Closed-Loop Autonomous Vehicle Simulation},
  author        = {Basant, Aarti and Kar, Amlan and Paschalidou, Despoina and Wei, Fangyin and Ferroni, Francesco and Garcia Cobo, Guillermo and Turki, Haithem and Ling, Huan and Seo, Jaewoo and Lucas, James and Wu, Jay Zhangjie and Wang, Jialiang and Lorraine, Jonathan and Gao, Jun and He, Kai and Tothova, Katarina and Xie, Kevin and Tyszkiewicz, Micha{\l} and Wu, Qi and de Lutio, Riccardo and Li, Ruilong and Fidler, Sanja and Kim, Seung Wook and Shen, Tianchang and Cao, Tianshi and Pfaff, Tobias and Lew, William and Wu, Xindi and Ren, Xuanchi and Lu, Yifan and Zhang, Yuxuan and Gojcic, Zan and Wang, Zian},
  year          = {2026},
  eprint        = {2606.03159},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2606.03159},
  url           = {https://arxiv.org/abs/2606.03159},
}