NeRF Synthetic dataset training fails - PSNR decreases instead of increasing

[Hudabey]

I'm trying to reproduce baseline results on the NeRF Synthetic dataset (lego scene) but training completely fails:

Observed behavior:

• Iteration 7000: PSNR 5.97 dB
• Iteration 30000: PSNR 2.14 dB
• PSNR decreases instead of increases

Expected behavior:

• Should reach ~32 dB on lego scene

What works:

• Tanks & Temples: 25.84 dB ✅ (works perfectly)
• MipNeRF360: 25.09 dB ✅ (works perfectly)

Dataset source:
Downloaded from: https://github.com/yenchenlin/nerf-pytorch (nerf_example_data.zip)
Official link is dead: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/input/nerf_synthetic.zip

Questions:

1. What's the correct NeRF Synthetic dataset source/format?
2. Should I preprocess the dataset differently?
3. Are there known issues with this dataset?

Environment:

• 3DGS commit: [run: cd /workspace/gaussian-splatting && git rev-parse HEAD]
• PyTorch 2.8.0+cu128
• CUDA 12.8

[JCpeny]

This question has been solved, the reason for low PSNR is that lots of artifacts occur when rendering. Turn to the follow two issue to get more help:
https://github.com/graphdeco-inria/gaussian-splatting/issues/1038
https://github.com/graphdeco-inria/gaussian-splatting/issues/1124