Title: Poor .splat result locally – looks like point cloud, unlike Colab output

[hugonbgg]

Hi! I'm testing OpenSplat locally using the banana dataset, but I'm getting a very different result compared to the example shown in the Colab notebook. Instead of a smooth and dense Gaussian Splat rendering, the output looks more like a sparse point cloud with very low visual quality.

Environment:
System: Debian 12 (Bookworm) on WSL
GPU: GTX 1050 Ti (Dell G3)
CUDA: 12.6 (nvcc --version shows V12.6.85)
Libtorch: libtorch-cxx11-abi-shared-with-deps-2.7.1+cu126
OpenSplat: compiled from source

Command used:
../OpenSplat/build/opensplat bananav2/ -o banana.splat -n 2000

Differences observed:
Output file size
Local: ~500 KB
Colab: ~7 MB

Execution logs
Colab:
Step 10: 0.203031 (0%)
...
Step 2000: 0.0424626 (100%)
Added 8926 gaussians, new count 23167
Culled 1277 gaussians, remaining 21890
Wrote cameras.json
Wrote banana.splat

Local:
Step 10: 0.340408 (0%)
...
Step 2000: 0.323012 (100%)
Wrote cameras.json
Wrote banana.splat

As you can see, locally there is no message about adding or culling Gaussians, and the loss remains high even after 2000 steps.

Question:
Could this be caused by a mismatch in dependencies (e.g. OpenCV, Libtorch), build flags, or some configuration missing locally?
Is there anything specific about the Colab setup that enables proper Gaussian Splatting behavior?

Thanks in advance for any insights!