NVIDIA FlashDreams — Third-Party Notices Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. This file identifies third-party software components that are included in, redistributed with, depended upon by, or otherwise present in this repository, and reproduces the attribution notices required by the corresponding upstream licenses. The primary FlashDreams source is licensed under the Apache License, Version 2.0 — see LICENSE at the repository root. Per-file SPDX identifiers and the REUSE 3.3 manifest at REUSE.toml are the machine-readable counterparts of the information below. Full license texts for the OSS source redistributions listed below are reproduced under LICENSES/; runtime dependency license texts remain available from the upstream URLs listed below. A package/project license does not by itself enumerate the terms of its bundled native libraries or tools. Most dependencies are consumed unmodified at runtime through Python's import system (i.e. dynamic linking via the interpreter); their source is not redistributed in this repository. Licenses are identified by SPDX identifier where available, or by their upstream agreement name. Model weights and example assets downloaded separately retain their own terms; FlashDreams's Apache-2.0 license does not replace those terms. Source-level redistributions (third-party source code physically present in this repository) are disclosed in the "Source-level redistributions" section at the end of this file. Additional third-party libraries are installed as runtime dependencies by `uv` / `pip` (resolving from PyPI and NVIDIA package indexes) or pulled from Hugging Face Hub on first run. Those packages and their downloaded model assets are not redistributed with this repository and carry their own license terms from their respective distribution locations. Review the license terms of these packages before use in your environment. ================================================================================ Direct runtime dependencies ================================================================================ boto3 Apache-2.0 https://github.com/boto/boto3 botocore Apache-2.0 https://github.com/boto/botocore einops MIT https://github.com/arogozhnikov/einops filelock MIT https://github.com/tox-dev/filelock ftfy Apache-2.0 https://github.com/rspeer/python-ftfy h5py BSD-3-Clause https://github.com/h5py/h5py huggingface-hub Apache-2.0 https://github.com/huggingface/huggingface_hub imgui-bundle MIT https://github.com/pthom/imgui_bundle loguru MIT https://github.com/Delgan/loguru numpy BSD-3-Clause https://github.com/numpy/numpy nvidia-ml-py BSD-3-Clause https://pypi.org/project/nvidia-ml-py/ psutil BSD-3-Clause https://github.com/giampaolo/psutil pyyaml MIT https://github.com/yaml/pyyaml safetensors Apache-2.0 https://github.com/huggingface/safetensors slangpy Apache-2.0 https://github.com/shader-slang/slangpy torch BSD-3-Clause https://pytorch.org torchvision BSD-3-Clause https://github.com/pytorch/vision tqdm MIT AND MPL-2.0 https://github.com/tqdm/tqdm transformers Apache-2.0 https://github.com/huggingface/transformers transformer-engine Apache-2.0 https://github.com/NVIDIA/TransformerEngine triton-windows MIT https://github.com/woct0rdho/triton-windows tyro MIT https://github.com/brentyi/tyro urllib3 MIT https://github.com/urllib3/urllib3 ================================================================================ Reference architectures ================================================================================ Wan 2.1 / Wan 2.2 Apache-2.0 https://github.com/Wan-Video/Wan2.1 FlashDreams's Wan integrations implement the architectures published by the Wan-Video team. Model weights are downloaded from upstream at runtime and are not redistributed in this repository. ================================================================================ Optional application: apps/crazy_robotaxi ================================================================================ ruamel-yaml MIT https://sourceforge.net/projects/ruamel-yaml ================================================================================ Optional applications: apps/interactive_drive and apps/omnidreams_game_engine ================================================================================ Pillow MIT-CMU https://github.com/python-pillow/Pillow pyarrow Apache-2.0 https://github.com/apache/arrow ================================================================================ Optional application: apps/v2v ================================================================================ av BSD-3-Clause https://github.com/PyAV-Org/PyAV BSD-3-Clause covers the Python bindings. Bundled FFmpeg libraries and codec dependencies retain their own terms; preserve the licenses of the actual installed or redistributed artifact. imageio-ffmpeg BSD-2-Clause https://github.com/imageio/imageio-ffmpeg Used only to generate test fixtures for this application. BSD-2-Clause covers the Python wrapper. Bundled FFmpeg executables carry additional FFmpeg and codec terms, including GPL-enabled builds; preserve the licenses of the actual installed or redistributed artifact. Big Buck Bunny CC-BY-3.0 https://peach.blender.org/about/ (c) copyright 2008, Blender Foundation / www.bigbuckbunny.org The v2v application downloads the Blender Foundation's big_buck_bunny_480p_h264.mov encode at runtime and processes an excerpt as demo input. The source video is licensed under the Creative Commons Attribution 3.0 license (https://creativecommons.org/licenses/by/3.0/) and is not redistributed in this repository. ================================================================================ Optional application: apps/handdreams ================================================================================ Pillow MIT-CMU https://github.com/python-pillow/Pillow Scene image decoding uses Pillow without modifying or redistributing its source. Pillow was already used by the original hand demo; the independent V2 integration declares it directly. ================================================================================ Optional application support: apps/handdreams/controller_input ================================================================================ websockets BSD-3-Clause https://github.com/python-websockets/websockets The optional webxr extra uses the unmodified WebSocket client for local Quest Browser DevTools automation. No upstream source is redistributed. This package is also used by the existing Isaac Teleop dependency stack. isaacteleop NOASSERTION https://github.com/NVIDIA/IsaacTeleop Public PyPI wheel metadata declares Apache-2.0. The 1.4.142 wheel also contains CloudXR native components; the complete wheel's license conclusion remains NOASSERTION pending the applicable Isaac Teleop/CloudXR review. HandDreams runs isaacteleop[cloudxr,retargeters-lite]==1.4.142 as an isolated subprocess. `flashdreams-controller-input setup isaac-teleop` installs the public PyPI wheels and complete dependency closure into a dedicated virtual environment using the platform-specific hash-pinned manifests in controller_input/isaac_teleop/. FlashDreams does not copy Isaac Teleop or CloudXR binaries into its GitHub release or install them in the main workspace environment. The operator must explicitly accept the NVIDIA CloudXR EULA at setup and again via --accept-eula before CloudXR startup; environment creation does not imply startup acceptance. The terms applicable to this runtime are linked at: https://github.com/NVIDIA/IsaacCapture/blob/main/deps/cloudxr/CLOUDXR_LICENSE Public package: https://pypi.org/project/isaacteleop/1.4.142/ Review upstream distribution notices before use or redistribution. Isaac Lab is not part of HandDreams and is not an installation dependency. ================================================================================ Optional integration: integrations_v2/flashvsr ================================================================================ FlashVSR Apache-2.0 https://github.com/OpenImagingLab/FlashVSR The integration uses the OpenImagingLab implementation and downloads JunhaoZhuang/FlashVSR-v1.1 checkpoints and a prompt tensor separately. The checkpoint model card declares Apache-2.0: https://huggingface.co/JunhaoZhuang/FlashVSR-v1.1/blob/27561b186ded3402d7c975f4fd722e2885b6135f/README.md Source attribution and the parity-test source pin are recorded below. ninja Apache-2.0 https://github.com/ninja-build/ninja The Python distribution wrapper and bundled Ninja executable can carry separate Apache/BSD terms; retain the selected artifact's licenses. ================================================================================ Optional integrations: integrations_v2/cosmos_interactive and cosmos3_nano_sim_auto ================================================================================ The LiDAR codec uses PyTorch FlexAttention and the vendored vLLM-Omni components attributed under Source-level redistributions below. scipy BSD-3-Clause https://github.com/scipy/scipy Native prepared-scene LiDAR projection uses SciPy rotation utilities. TAEHV MIT https://github.com/madebyollin/taehv The streaming super-decoder implements the TAEHV architecture using the shared FlashDreams TAEHV components. Decoder weights are downloaded separately from upstream; no model weights are included in the package. The Cosmos Interactive runtime contains the native MoT network, checkpoint mapping, text tokenizer wiring and LiDAR codec. The model and scene integrations use the existing PyTorch, einops, safetensors, transformers, loguru, SciPy, Pillow, imageio and Ludus dependencies disclosed in this document. Training checkpoints and prepared scene assets have their own distribution terms and are not included in these source packages. ================================================================================ Optional integration: integrations_v2/omnidreams ================================================================================ mediapy Apache-2.0 https://github.com/google/mediapy nvidia-cudnn-frontend MIT https://github.com/NVIDIA/cudnn-frontend opencv-python-headless Apache-2.0 https://github.com/opencv/opencv-python shapely BSD-3-Clause https://github.com/shapely/shapely ludus-renderer Apache-2.0 (NVIDIA, internal distribution) cmake Apache-2.0 https://github.com/scikit-build/cmake-python-distributions Python distribution wrapper; bundled CMake uses BSD-3-Clause: https://github.com/Kitware/CMake imageio BSD-2-Clause https://github.com/imageio/imageio nvidia-cudnn-cu13 LicenseRef-NvidiaProprietary https://developer.nvidia.com/cudnn nvidia-nvcomp-cu12 LicenseRef-NvidiaProprietary https://github.com/NVIDIA/nvcomp pandas BSD-3-Clause https://github.com/pandas-dev/pandas PyNvVideoCodec MIT https://developer.nvidia.com/pynvvideocodec setuptools MIT https://github.com/pypa/setuptools NVIDIA PhysX 5.9.0 BSD-3-Clause https://github.com/NVIDIA-Omniverse/PhysX ``ludus-renderer`` downloads the pinned PhysX source archive at first use, verifies its SHA-256, and builds it into a platform-local cache. The source and resulting native module are not redistributed in this repository. The OmniDreams runners read and write .mp4 video by invoking the external "ffmpeg" command-line tool as a subprocess (through mediapy). ffmpeg is a user-supplied system prerequisite: it is neither redistributed in this repository nor linked into any FlashDreams binary, and no FlashDreams package declares it as a runtime dependency. It is therefore not part of the dependency inventory above; install it separately (e.g. "apt install ffmpeg"). NVIDIA OmniDreams single-view native builds use script-managed, pinned source downloads under integrations_v2/omnidreams/impl/omnidreams_singleview/3rdparty/. Those downloaded trees are ignored by Git and are not source-level redistributions in this repository. CUTLASS BSD-3-Clause https://github.com/NVIDIA/cutlass Pinned by integrations_v2/omnidreams/impl/omnidreams_singleview/ thirdparty_sources.json at f3fde58372d33e9a5650ba7b80fc48b3b49d40c8. The sync script removes CUTLASS's python/CuTeDSL subtree because the native build does not use it and upstream licenses it separately. The script also applies the project-owned SM120 TMA descriptor patch under integrations_v2/omnidreams/impl/omnidreams_singleview/patches/cutlass/. SageAttention Apache-2.0 https://github.com/thu-ml/SageAttention Pinned by integrations_v2/omnidreams/impl/omnidreams_singleview/ thirdparty_sources.json at d1a57a546c3d395b1ffcbeecc66d81db76f3b4b5. SpargeAttn Apache-2.0 https://github.com/thu-ml/SpargeAttn Pinned by integrations_v2/omnidreams/impl/omnidreams_singleview/ thirdparty_sources.json at bfd980b781784c04ad6a53e7ee657c0645d99171. ================================================================================ Optional post-processing: RTX Video Super Resolution ================================================================================ nvidia-vfx LicenseRef-NvidiaProprietary https://pypi.org/project/nvidia-vfx/ The ``rtx-postprocess`` extra installs NVIDIA's proprietary Python bindings for RTX Video Super Resolution and deblur. It is optional, dynamically loaded only when an RTX preset is selected, and is not redistributed as source in this repository. ================================================================================ Optional integration: integrations_v2/lingbot ================================================================================ LingBot-World v1 Apache-2.0 https://github.com/Robbyant/lingbot-world The v1 fast checkpoint model card also declares Apache-2.0: https://huggingface.co/robbyant/lingbot-world-fast/blob/b7a1db2da9bd3ba675ee01c228f41db0309b981e/README.md LingBot-World v2 CC-BY-NC-SA-4.0 https://github.com/Robbyant/lingbot-world-v2/blob/1895d300d8ac936401689b26389f51cbd36530eb/LICENSE.txt The v2 14B and 1.3B causal-fast checkpoints declare the same terms: https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast/blob/5c33dd40b213598c418fd25bff30fdbd23fd38a7/README.md https://huggingface.co/robbyant/lingbot-world-v2-1.3b-causal-fast/blob/7e36a5f919f86cb4255cc9bfc30adb44963fbde1/README.md Checkpoints are downloaded separately. The example-data path also downloads from the v2 repository, including when using v1 presets; its applicable provenance/grant must be checked separately. The v2 public terms require noncommercial use, attribution and share-alike for shared adaptations. They do not establish commercial-use rights. aiohttp Apache-2.0 AND MIT https://github.com/aio-libs/aiohttp aiortc BSD-3-Clause https://github.com/aiortc/aiortc multidict Apache-2.0 https://github.com/aio-libs/multidict opencv-python-headless Apache-2.0 https://github.com/opencv/opencv-python ================================================================================ Optional integration: integrations_v2/sana_wm ================================================================================ sentencepiece Apache-2.0 https://github.com/google/sentencepiece ================================================================================ Source-level redistributions ================================================================================ -------------------------------------------------------------------------------- Cosmos Interactive decoder architecture adaptation -------------------------------------------------------------------------------- Path: integrations_v2/cosmos_interactive/cosmos_interactive/_taehv_super.py License: Apache-2.0 AND MIT (see LICENSES/Apache-2.0.txt and LICENSES/MIT.txt) Copyright (c) 2025 Ollin Boer Bohan Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. Upstream: https://github.com/madebyollin/taehv The decoder adapts the TAEHV-super layout for FlashDreams streaming caches and checkpoint loading. The decoder checkpoint is pinned to upstream revision 62f7591f59dfbb4c3c02b7a621d180a9eeaba26c and downloaded separately. The following third-party source is physically present in this repository (not just consumed at runtime). Each file retains its original copyright notice inline; license texts are reproduced under LICENSES/ or in the credited source files. -------------------------------------------------------------------------------- HPG-2011 NVIDIA CUDA Rasterizer port -------------------------------------------------------------------------------- Path: integrations_v2/omnidreams/impl/ludus-renderer/ludus_renderer/_cpp/cudaraster/ License: BSD-3-Clause (see LICENSES/BSD-3-Clause.txt), with the file-specific exceptions listed below. Copyright (c) 2009-2026, NVIDIA Corporation. All rights reserved. The cudaraster subtree is a port of NVIDIA Research's HPG-2011 CUDA Rasterizer (Laine & Karras). It is NVIDIA-authored and is redistributed under the 3-clause BSD license carried in integrations_v2/omnidreams/impl/ludus-renderer/ludus_renderer/_cpp/cudaraster/LICENSE, with the following file-specific terms: CudaRasterKernels.cu: Apache-2.0 (local integration code). cudaraster_fw_stub.cpp: Apache-2.0 AND BSD-3-Clause (Apache port additions and retained BSD framework portions). Original BSD notices are preserved. The tile-level rasterization pipeline (triangle setup, bin / coarse / fine raster) is upstream code; the host wrapper, viewport / multi-image plumbing, and deterministic tiebreaker are local modifications. See integrations_v2/omnidreams/impl/ludus-renderer/ludus_renderer/_cpp/cudaraster/PORT_NOTES.md for the full list of deviations. -------------------------------------------------------------------------------- LodePNG -------------------------------------------------------------------------------- Path: integrations_v2/omnidreams/impl/ludus-renderer/ludus_renderer/_cpp/ cudaraster/framework/3rdparty/lodepng/{lodepng.h, lodepng.cpp} License: Zlib (see LICENSES/Zlib.txt) Copyright (c) 2005-2011 Lode Vandevenne Upstream: https://lodev.org/lodepng/ LodePNG is a PNG codec implementation embedded by the upstream cudaraster framework. The inline notice in lodepng.h preserves the upstream copyright notice as required by the Zlib license; LICENSES/Zlib.txt reproduces the license text in full. -------------------------------------------------------------------------------- Cosmos-Drive-Dreams camera helpers -------------------------------------------------------------------------------- Path: integrations_v2/omnidreams/impl/conditioning/world_scenario/ camera_base.py, ftheta.py, pinhole.py License: Apache-2.0 (see LICENSES/Apache-2.0.txt) Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved. Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. Upstream: https://github.com/nv-tlabs/Cosmos-Drive-Dreams Modified for FlashDreams; the per-file modification and copyright notices are retained. The reviewed upstream comparison is revision 915f79f2e34b5420c654ca9851e466c823409459, under cosmos-drive-dreams-toolkits/utils/camera/ (base.py, ftheta.py, pinhole.py). This comparison pin is not asserted to be the original-copy revision. -------------------------------------------------------------------------------- SwiftVR adapted source -------------------------------------------------------------------------------- Path: integrations_v2/swiftvr/impl/ attention.py, pipeline.py, encoder/__init__.py, decoder/__init__.py, decoder/network.py, transformer/__init__.py, transformer/network.py License: Apache-2.0 (see LICENSES/Apache-2.0.txt) Copyright (c) 2026 SwiftVR Authors. Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. Upstream: https://github.com/H-oliday/SwiftVR/tree/5ca168cef6ca7200f135fdfea85e5e13d12c5b53 These seven files adapt SwiftVR's encoder, transformer, decoder and streaming pipeline to FlashDreams. Their upstream copyrights, Apache license identifiers and NVIDIA modification notices remain in the files. Model weights are obtained separately and are not included by this source attribution. -------------------------------------------------------------------------------- LingBot-World v1 camera helpers -------------------------------------------------------------------------------- Path: integrations_v2/lingbot/impl/encoder/utils.py License: Apache-2.0 (see LICENSES/Apache-2.0.txt) Upstream: https://github.com/Robbyant/lingbot-world/blob/a43bec7f8091c83e9b30b16b912f6fc906236fa6/wan/utils/cam_utils.py Camera interpolation, SE(3), intrinsic and Plücker-ray helpers are adapted for FlashDreams. This is a reviewed comparison revision, not an assertion of the original-copy revision. The v2 model terms above do not replace this v1 source grant. -------------------------------------------------------------------------------- FlashVSR decoder helpers -------------------------------------------------------------------------------- Path: integrations_v2/flashvsr/impl/decoder/network.py License: Apache-2.0 (see LICENSES/Apache-2.0.txt) Upstream: https://github.com/OpenImagingLab/FlashVSR/blob/b527c6f285fb30df530f5febc8b45764a789c961/examples/WanVSR/utils/TCDecoder.py The decoder adapts TCDecoder helpers, including IdentityConv2d and decoder deepening. The comparison revision above is the upstream pin used by integrations_v2/flashvsr/tests/parity_check/run.sh. FlashDreams modifies these helpers for its decoder/cache interfaces. -------------------------------------------------------------------------------- Material for MkDocs KaTeX example -------------------------------------------------------------------------------- Path: docs/source/javascripts/katex.js License: MIT upstream portions; Apache-2.0 NVIDIA modifications (full upstream notice retained in the source file). Copyright (c) 2016-2024 Martin Donath Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. Upstream: https://github.com/squidfunk/mkdocs-material/blob/f0c99d89da61a247318f1a9471a082cfb981b3e0/docs/reference/math.md The initializer adapts the upstream example with formatting changes. -------------------------------------------------------------------------------- vLLM-Omni LiDAR inference blocks -------------------------------------------------------------------------------- Path: integrations_v2/cosmos_interactive/cosmos_interactive/_lidar_codec/ transformer_vae.py, rope.py, encoding.py, neighborhood_attention.py License: Apache-2.0 (see LICENSES/Apache-2.0.txt), with MIT terms retained for k-diffusion-derived portions of transformer_vae.py. Copyright contributors to the vLLM-Omni project Copyright (c) 2022 Katherine Crowson Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. Version: 8ef7df57219688835f3752accb2c3b2f444b4058 (cosmos3_gtc_berlin_release) Upstream: https://github.com/MaciejBalaNV/vllm-omni/tree/8ef7df57219688835f3752accb2c3b2f444b4058/vllm_omni/diffusion/models/cosmos3/lidar_encoder License URL: https://github.com/MaciejBalaNV/vllm-omni/blob/8ef7df57219688835f3752accb2c3b2f444b4058/LICENSE The pinned release snapshot provides the VAE encoder/decoder, 3D rotary positions, polar coordinates and sparse neighborhood attention. FlashDreams changes VAE imports to package-relative imports and computes constructor shape metadata on CPU to support meta-device inspection. Other vendored files are unchanged. network.py adapts these modules to the existing FlashDreams checkpoint and streaming interfaces. No vLLM runtime dependency, model weights, or sample data is redistributed by this source change. The codec's complete upstream Apache license is retained in that directory's LICENSE.vllm-omni; its NOTICE accompanies the copied source. The shared Cosmos package also records this redistribution in its own LICENSE, NOTICE and THIRD-PARTY-NOTICES. The VAE credits k-diffusion's image_transformer_v2.py. Its full inherited MIT notice, including Copyright (c) 2022 Katherine Crowson, is reproduced in the codec directory's LICENSE.k-diffusion. The MIT text is pinned at https://github.com/crowsonkb/k-diffusion/blob/e95849c23340e78c4bca5ef6cea6dd60611acad8/LICENSE. The standalone geometry sampler and spinning-LiDAR projector are not covered by this upstream attribution. Their prior implementation bodies have been removed from the public integration; its independent geometry and projection replacements are enabled. Those first-party replacements are separate from the vLLM-Omni redistribution. ================================================================================