The High Cost of Continuing 3DGS Training Based on Model Prediction Results

I've found that both the point clouds predicted by the model and the 3DGS point clouds are pixel-level, which means that inputting just a few dozen images will generate millions of points. These points appear overly dense yet inaccurate, and this causes the subsequent 3DGS training to take much longer. Does anyone have successful experience dealing with this situation? What if we input thousands of images? It feels like the direct generation of point clouds by the model or the use of 3DGS point clouds is a dead end.

No files, tests, or entry points are named. Start by reviewing the model-predicted and 3DGS point-cloud workflow and measuring training with a few dozen images; the issue needs an agreed approach and concrete acceptance criteria before a newcomer can determine what done looks like.

Written by the indexing model from the issue text.