6.0.0
Operating System
Ubuntu 24.04
Detailed Description
I’m reconstructing a dynamic scene from a single fixed-position 1080p/30fps video using NVIDIA’s ViPE → GEN3C → Lyra pipeline. Lyra outputs standard 3DGS .ply files at several timestamps throughout the video (“bullet time” snapshots), each containing ~2M Gaussians.
What works:
When I load the raw gaussians_0.ply output directly in SuperSplat, the scene renders clearly.
What doesn’t work:
I converted the same fixed .ply to NuRec USDZ using:
python -m threedgrut.export.scripts.ply_to_usd <input.ply> --output_file <output.usdz>
(with NuRecExporter(export_cameras=False) since I have no associated dataset/camera poses — patched directly in ply_to_usd.py)
When I load the resulting .usdz in Isaac Sim, the OmniNuRecFieldAsset prim (with density_field and emissive_color_field) is created correctly in the stage, but the rendered result is significantly blurrier/streakier than the SuperSplat view of the identical (opacity-fixed) data even at the same approximate camera angle, and even after letting RTX Real-Time 2.0 accumulate frames.
I also tried the alternative ParticleField3DGaussianSplat (UsdVol “standard”/“lightfield”) schema by manually authoring the prim myself (positions, scales as exp(scale_i), orientations as normalized quaternions, opacities as sigmoid(opacity), SH degree-0 coefficients from f_dc_0/1/2, with required extent attribute). That path rendered (after fixing missing extent/Boundable requirements), but more blurry than NuRec, with worse streak/noise artifacts when rotating the camera further from the source viewpoint, consistent with what’s described in the other thread comparing these two pipelines (forum thread 373260). So for now I am sticking with NeRec
Questions:
- Is there a known minimum angular/baseline coverage requirement for Lyra-generated 3DGS scenes to reconstruct cleanly across a wide viewing range, or is this purely a function of how many/how-wide the GEN3C trajectories are?
- Independent of the data issue above , is there a recommended renderer setting (samples-per-pixel, denoiser, kernel/falloff parameter) specific to
OmniNuRecFieldAssetin Isaac Sim 6.0 that would meaningfully close the quality gap vs SuperSplat, even on data with good viewing coverage? - Is
ParticleFieldKernelGaussianEllipsoidAPIexpected to be explicitly applied viaprim.ApplyAPI(...)forParticleField3DGaussianSplatprims, or is it implicitly assumed by the renderer? I found that explicitly applying it viaApplyAPIcaused fewer Gaussians to render (didn’t investigate further), so I’m unsure if I’m using it correctly or if applying it manually is even the intended workflow. - Is there anything else I can do to make this better
Screenshots or Videos
in IsaacSim:
In Supersplat:

