Why is the result different from the demo?

[soohyoen]

Thank you for your great work, I got good results by inputting 8 images on your demo page.
However, when I operate your code directly, the results are very different.
Artifact looks worse when turned into code than demo
What is the problem?

• Code

• Demo

[kairunwen]

Hi, The demo employs lower initial learning rates, while the code sets higher values for parameters like xyz_lr and feature_lr. The difference may stem from these initial rates. Try lowering the rates in training_setup.

[soohyoen]

@kairunwen
Thank you so much for your quick response. Can you tell me how much difference there is?

[kairunwen]

10 times.

[soohyoen]

@kairunwen

Hello, thank you so much for your kind response.

However, the problem is not solved yet.
I reduced learning_rate by about 10, but the result is as shown in the picture.
I think it has a soft shape when I see the results of demo.

What I want is things that have less artifact in the gaussian form like I ran in demo.

Is there an answer other than learning_rate?

• code
  

• demo

[starwalkin-C]

@soohyoen Can you inspect the result of your rendering from render.py and see if the result is consistent with your visualization?

[soohyoen]

@kairunwen @starwalkin-C

The results after render.py look neat.
But they're not neat when you see them on superplat.
Is the instantsplat-2dgs on demo?

[kairunwen]

@soohyoen Hi, could you provide the image corresponding to this Gaussian? I need it to debug this issue.

[soohyoen]

@kairunwen
Thank you so much for your answer

Attached is the picture you rendered after learning.

• Train view
  

• Test view

There were eight train views and six test views and a few were sampled.

But what's a little strange is that the first picture continues from 1 to 7 in the train view I attached.

If you need additional information, please request it

[kairunwen]

Hi @soohyoen, after debugging on my own data, I found that this issue is indeed caused by a high learning rate (lr). You can add some constraints on the scale or simply lower the learning rate to address it. (Reference: #42 (comment)).

If you could share your original images to my email: wenkairun@gmail.com, it would greatly help with further debugging.