Blurry Gaussian Splat Results with splatfacto-w-light Using ns-process-data metashape Conversion

[jitesh3023]

I'm using Metashape to align cameras and then running ns-process-data metashape, followed by training with splatfacto-w-light. Although the camera alignment looks correct in Metashape, the resulting Gaussian Splat is extremely blurry and lacks detail when viewed in Viser, as if the Gaussians never converge—even after 30,000 iterations. Most camera positions look reasonable, but a few frames are incorrectly placed. I'm currently not sure what's causing this; I'm assuming it might be due to an incorrect transformation or coordinate mismatch during the Metashape-to-Nerfstudio conversion, but I could be wrong. This issue does not occur with small datasets (e.g., around 50 images), where the results are sharp and accurate. The problem only seems to arise with larger datasets. My processing command is:
process_command = [
"ns-process-data", "metashape",
"--data", data_dir,
"--xml", camera_xml,
"--output-dir", output_dir,
"--max_dataset_size", "-1"
]
And my training command is:
train_command = [
"ns-train", "splatfacto-w-light",
"--pipeline.datamanager.images-on-gpu", "False",
"--pipeline.datamanager.cache-images", "disk",
"--pipeline.model.sh-degree", "2",
"--pipeline.model.num-downscales", "0",
"--pipeline.model.enable_bg_model", "False",
"--pipeline.model.enable_alpha_loss", "False",
"--max-num-iterations", "30000",
"--viewer.quit-on-train-completion", "True",
"--data", output_dir,
"--output-dir", output_dir,
"nerfstudio-data",
"--downscale-factor", "1",
"--train-split-fraction", "1"
]
Could this be related to a scaling or transformation mismatch when using Metashape XML files with larger datasets? Any guidance would be appreciated.

I referred this ->
Metashape

1. All images must use the same sensor type (but multiple sensors are supported).
2. Align your images using Metashape. File -> Workflow -> Align Photos...
   metashape alignment
3. Export the camera alignment as a xml file. File -> Export -> Export Cameras...
   metashape export
4. Convert the data to the nerfstudio format.
   ns-process-data metashape --data {data directory} --xml {xml file} --output-dir {output directory}
5. Train with nerfstudio!
   ns-train nerfacto --data {output directory}

Screen.Recording.2025-08-04.at.10.03.24.PM.mp4