Quality and GPU memory usage issues

[Linkersem]

Hello, @Snosixtyboo @ameuleman my device is 4090 24G.

First,when using the SIBR viewer to view my trained model (model size is 4G), I found that the gpu memory is about 22G, if this is the case, if I have a larger scene, 24G of video memory will not be able to withstand it, but the paper says that you can try to render 88G size files, which makes me feel a little confused here.

Second,I have a set of data collected from a large scene, which is denser than the sample data you provided. After training, I found that the details were not well represented compared to 3dgs, especially some fonts were very blurry.
I have conducted some analysis on the previous issue and found that re triangulating the scene after segmenting it into chunks can result in a very sparse point cloud, which makes it difficult to reconstruct some areas. Therefore, I will replace the triangulated data with the original colmap point cloud data. However, this leads to a lot of white fog in the final trained model

May I ask what the reason is? How should this phenomenon be resolved?

[ameuleman]

Hi,

1. The viewer loads only a part of the hierarchy into GPU memory, following the --budget parameter. Nodes will be loaded and removed from GPU memory as the user moves through the scene.
2. Discussion on issue 61

[Linkersem]

Hello, thank you for your reply!

1. I understand what you said, but I would like to ask if there is a quantitative indicator for this? For example, if I have a 4G model, how much GPU memory do I need? For example, my 4G model cannot run on one of my 2060 8G devices and will crash, but it can be browsed normally on a 4090 24G device.
   I hope to get a more specific conclusion, such as the relationship between gpu memory and model size? For example, I also have a 500M model, which can be rendered normally on a 2060 8G device.

2. I will continue to discuss this in Weird result #61

Thank you again for your patient reply!