Tips on large resolutions

[LaFeuilleMorte]

Hi, I'm recently using gsplat to train 8K resolution images. My gaussian model can fit well on 2K resolution. But when I'm training on 4K or even higher resolution 8K, the loss will drop very slowly and the metrics are way too worse than training in 2K resolution. And the rendering quality is also much worse. To solve this problem, I tried the following measures:

1. Use multi-scale training: Use the raw image and its downsampled ones as training views: (1, 2, 4, 8) as the downsample ratio.
2. Use smaller densify grad threshold but triggers severe floaters problems.
3. Warm up, training from 4X smaller resolution from the beginning, and then train on the high resolution images.

Both the above measures has no significant improvements on this problem. Do you have any advices on this problem?