Significant Performance Discrepancy Between Latest and Older Versions on the Same Dataset

[irisfreesiri]

Hi,

I recently checked out two different versions of the Gaussian Splatting project on my machine:

Latest version: Commit 54c035f (released on October 31, 2024)
Older version: Commit d9fad7b (released on February 13, 2024)
I ran both versions on the same dataset (coffee), but the results differed significantly. Specifically, the latest version produced much lower scores compared to the older version.

training result form latest version:

training result form older version:

This makes me wonder if there were any changes introduced between these versions that might have affected the performance. Unfortunately, I couldn't identify any obvious reasons for this discrepancy from the commit history.

Is anyone actively maintaining this project? If so, could you clarify if there were intentional changes that might explain the difference in results?

Any help or insights would be greatly appreciated.

Thank you!

[caoletian]

Checkout these issues #1038 #1124

[irisfreesiri]

Checkout these issues #1038 #1124

Thank you for taking the time to respond to previous questions.

I noticed that in issue #1124, the phenomenon of black backgrounds with white flares in the NeRF dataset was mentioned. Additionally, it seems that training in the "train" scene has become significantly slower compared to before. Since the last update aimed to accelerate training speed and the most recent update was about three months ago, we are curious to know if there are any plans for follow-up fixes or improvements?

Thank you again for your time and looking forward to your response!