horizonrobotics Scene3D Deployment
Use the repository installer first. It initializes thirdparty/pano2room,
installs txt2panoimg without dependencies, and builds fused-ssim and
tiny-cuda-nn.
bash install.sh scene3d
scene3d-cli --help
tiny-cuda-nn
tiny-cuda-nn reads TCNN_CUDA_ARCHITECTURES, not
TORCH_CUDA_ARCH_LIST.
- RTX 4090:
TCNN_CUDA_ARCHITECTURES=89. - RTX 5090:
TCNN_CUDA_ARCHITECTURES=120with nvcc 12.8 or newer.
If a VCS install reports missing fmt, CUTLASS, or CMakeRC files, clone the
source recursively and install its torch binding:
git clone --recursive https://github.com/NVlabs/tiny-cuda-nn /tmp/tiny-cuda-nn
python -m pip install --no-build-isolation \
/tmp/tiny-cuda-nn/bindings/torch
Reduce MAX_JOBS if compilation runs out of memory.
basicsr and txt2panoimg
Keep torch 2.8 with torchvision 0.23. The installer intentionally uses
--no-deps for txt2panoimg, and gen_scene3d.py applies the compatibility
patch before importing basicsr. Do not downgrade torchvision to fix
torchvision.transforms.functional_tensor.
Blackwell double backward
On RTX 5090, tiny-cuda-nn's eager HashGrid double backward can fail with
CUDA error: an illegal memory access was encountered during Pano2Mesh depth
optimization. The scene3d installer prepares tiny-cuda-nn's runtime CUDA
headers, and EmbodiedGen enables its exact JIT gradient path only on Blackwell.
Rerun bash install.sh scene3d after updating an existing checkout.
Quick Reference
| Error | Root Cause | Fix |
|---|---|---|
Could NOT find fmt or CUTLASS |
tiny-cuda-nn submodules missing | Use a recursive clone |
functional_tensor missing |
txt2panoimg pulled incompatible deps | Reinstall through install.sh scene3d |
compute_120 unsupported |
nvcc older than 12.8 | Load the cu128 hook |
| HashGrid backward illegal memory access | Eager double backward on Blackwell | Reinstall scene3d to enable exact JIT gradients |
| Build OOM | Too many compiler jobs | Set MAX_JOBS=4 |