Inquiry on Log-Based 3D Scene Generation Using Omniverse NuRec

Hello,

We’re consdering using NVIDIA Omniverse NuRec to generate 3D scenes.

After reviewing the documentation https://docs.nvidia.com/nurec/index.html

we have several questions.

Questions

  1. How should log data be prepared for NuRec generation?
  2. Is the LiDAR data is necessary for NuRec generation?
  3. Is output of NuRec USDZ file which can directly be used to AlpaSim?What components are typically included in the generated USDZ scene assets?
  4. When using NuRec to convert log data into 3D scene assets, does each scene require a separate reconstruction/training process (e.g., 3DGRUT training)?
  5. After reconstructing a scene with NuRec, is it possible to modify weather or environmental conditions (rain, snow, fog, lighting) during simulation?
  6. There is a Fixer model that can post-process reconstructed data, but is this intended to improve video quality (e.g., resolution)?

We would appreciate it if you could provide answers to the above questions.

Thank you.

NuRec expects well-calibrated, time-synchronized multi‑sensor logs (cameras, optionally LiDAR, plus ego pose) and outputs a USDZ scene package that can be reused across simulators, with Fixer improving reconstruction artifacts rather than raw video resolution. carla.readthedocs

Log data preparation

To prepare logs for NuRec you typically need:

  • Calibrated camera intrinsics and extrinsics for all views, with stable poses over the log. voxel51
  • Time synchronization between cameras, LiDAR (if present), and ego pose/IMU so each frame bundle aligns in time. carla.readthedocs
  • A consistent world / vehicle coordinate system and complete metadata (timestamps, sensor IDs, vehicle pose trajectory). voxel51
  • Sufficient parallax and coverage: multiple viewpoints over the area you want reconstructed, not just a single static camera. developer.nvidia

A typical example is an AV-style log with 4–8 surround cameras, optional LiDAR, and GNSS/IMU over a 10–30 s segment covering the scene of interest. developer.nvidia

Is LiDAR necessary?

  • NuRec can reconstruct scenes from camera-only logs; LiDAR is not strictly required. docs.nvidia
  • When available, LiDAR point clouds help constrain geometry, improving structure, depth consistency, and robustness in texture‑poor regions. arxiv
  • For indoor or short‑range robotics scenes, multi‑view RGB plus accurate poses are often sufficient; for long‑range AV highway/urban scenes, LiDAR tends to improve large‑scale layout and distant structures. sciencedirect

So you can run NuRec without LiDAR, but you should expect better metric accuracy and fewer geometric artifacts when LiDAR is included.

NuRec outputs and USDZ contents

NuRec writes a USDZ bundle as its primary artifact. This package normally contains: docs.nvidia

  • A USD scene graph (the neural scene asset with geometry/volume, materials, and bindings to the neural renderer). docs.nvidia
  • A trained reconstruction checkpoint (e.g., NeRF / 3DGS / 3DGUT weights) referenced by the USD schema. docs.nvidia
  • A JSON track file with ego and rig trajectories and sequence metadata (frame ordering, timestamps, etc.). docs.nvidia
  • Optionally, an OpenDRIVE (XODR) map if you provide one; NuRec does not infer XODR itself. docs.nvidia

NVIDIA notes that these USDZ assets are designed to load directly into Omniverse‑based apps (Isaac Sim, CARLA integration via NuRec, robotics/AV platforms with OmniNuRecVolumeAPI). Whether you can “directly” use them in AlpaSim depends on AlpaSim’s USD/USDZ and NuRec integration; if AlpaSim is an Omniverse Kit–based app with OmniNuRecVolumeAPI support, you can generally load the USDZ and control rendering, otherwise you may need an import or conversion step. linkedin

Typical scene components in the USDZ

  • Neural volume representation of the environment (road, buildings, foliage, static objects). carla.readthedocs
  • Materials/appearance parameters used by the NuRec renderer (neural radiance field or 3D Gaussians configuration). developer.nvidia
  • Coordinate frames and transforms for world, ego vehicle, and sensor rigs. voxel51
  • Optional semantic groupings or prims used for simulation hooks (e.g., driveable area volumes, occluder tags), depending on your pipeline. edge-ai-vision

These are scene assets; dynamic agents (cars, pedestrians) are usually simulated separately, not baked as controllable USD prims in the NuRec volume. edge-ai-vision

Per‑scene reconstruction / training

  • Each distinct scene you want to reconstruct from logs is processed separately: you pick a segment of data and run a reconstruction (which internally trains the neural representation for that segment). carla.readthedocs
  • You do not typically run a global “once for all scenes” 3DGUT/NeRF training; the training is per scene/log, producing one checkpoint and USDZ per scenario. developer.nvidia
  • Workflows can be automated to run many scenes in batch, but conceptually it is still one reconstruction per scene region you care about. voxel51

Some pipelines may reuse pre‑trained world models for priors, but NuRec’s shipped process still optimizes each reconstruction to its specific sensor logs. edge-ai-vision

Changing weather and environment in simulation

  • NuRec itself reconstructs the scene as observed (lighting, weather, foliage, etc.); the neural asset encodes that appearance. carla.readthedocs
  • To change weather, time of day, or fog during simulation, NVIDIA pairs NuRec scenes with world foundation models and tools like Cosmos Transfer, which can generate scene variations (e.g., different weather and lighting) on top of reconstructed assets. edge-ai-vision
  • Within Omniverse/Isaac Sim, you can also adjust sensor rendering parameters (exposure, noise, fog volumes, rain particle systems) at the simulator level, effectively changing rendered conditions even when the underlying neural asset was trained under one condition. voxel51

So yes, you can modify environmental conditions for simulation, but it is done via simulator and generative‑model tooling around NuRec rather than retuning the original reconstruction.

Fixer model purpose

  • NuRec Fixer is a generative post‑processing model that removes reconstruction artifacts and restores missing detail in under‑constrained regions of the neural scene. paralleldomain
  • It is diffusion/transformer‑based, built on NVIDIA Cosmos, and is applied either during reconstruction or as a post‑process on rendered frames to inpaint, sharpen, and clean up surfaces and textures (e.g., road markings, building facades). paralleldomain
  • While perceived resolution and sharpness improve, Fixer is not a generic “video super‑resolution” model; its objective is simulation‑grade scene fidelity and artifact removal rather than arbitrary video upscaling. paralleldomain

An example workflow is: reconstruct with NuRec → render sensor views from the neural scene → run Fixer on those rendered frames to reduce smearing, fill in occluded or sparse regions, and make the scene more realistic for AV or robotics training. paralleldomain

To tailor this to your pipeline, what simulator stack are you planning to target first (Isaac Sim, CARLA, AlpaSim only, or a custom Omniverse app)?

Thank you very much for your detailed response.

We are considering using NuRec for AV (Autonomous Vehicle) applications, and we are planning to evaluate the use of AlpaSim.

In addition, we would like to ask two further questions:

  • When LiDAR is used together with cameras, how much improvement in reconstruction or measurement accuracy can typically be expected?

  • If we would like to use NuRec to reconstruct 3D scenes from our log data, which types of log data are required or recommended for this purpose?

We would appreciate your guidance on these points.

Well LIDAR is always going to help tremendously with accuracy. What is more accurate than a laser… nothing!
As for the log data, see the first paragraph of my response above.

Hello Richard,

I have a question regarding the note in the NuRec pre‑created nvidia/PhysicalAI-Autonomous-Vehicles · Datasets at Hugging Face, which states:

“Note: We currently support NuRec reconstruction only for the Hyperion 8.1 sensor rig.”

Could you please confirm whether this means that NuRec reconstruction is strictly limited to the Hyperion 8.1 sensor rig (e.g., the full configuration with 12 cameras, 9 LiDARs, etc.),
or whether other sensor configurations—such as a camera‑only sensor rig—can also be used for NuRec reconstruction, with the understanding that reconstruction quality or performance may be degraded compared to Hyperion 8.1?

Thank you very much for your clarification.

Best regards,
Siyuan

I am sorry but that I do not know. I would ask these questions on the actual Hugging Face repo to get direct help there.
nvidia/PhysicalAI-Autonomous-Vehicles · Discussions

Other sensor configurations can be used in NuRec (H8 and H8.1 are included in the data set) with the caveat you stated (reconstruction quality or performance may be degraded compared to Hyperion 8.1).

Thanks !

Hello aperec, Thank you very much for your answer. I understand the caveat, and I appreciate the clarification.

Thank you!

Hello aperec,

I have an additional question regarding the use of camera-only data.

If there are any issues, such as degradation in reconstruction quality,
could you please let us know any methods to mitigate or resolve these problems?

Thank you very much for your clarification.

Best regards,

Ryoya

@ryoya_ogino As stated above, yes, there will be significant degradation of final data, if not using LiDar. It is designed to take both inputs. If you are having issues then all we can recommend is to get LiDar data, or to use very very high quality stereo photos and a lot of them.

Hello Rechard,

Thank you very much for your answer.

I understand degradation and recommendation of getting LiDAR data.

I also appreciate the clarification.

Your answer is very helpful. At the same time, I would like to ask a related question: How can we use the video output by Cosmos Predict (which seemingly does not provide camera internal and external parameters) for Nurec reconstruction?

@idlqy As far as I know and understand, there is really no link at all between Cosmos and NuRec. They are very different systems. Cosmos is a Visual AI model, which is incredibly powerful, but it does not “output” any kind of 3D files or ingestible data, other than the visual. There is no workflow from Cosmos to NuRec. NuRec really requires a lot of super clean, super organized data to work well. Nothing from Cosmos would be useful.