280*518

[chf-git027]

Is the resolution limited by the model itself? If we want to obtain outputs with higher resolution, is it necessary to redesign and retrain the model? I have noticed that current feed-forward Gaussian Splatting (GS) methods mostly produce outputs with relatively low resolution. Is this because higher resolutions introduce certain difficulties that make the models hard to train effectively?

[primebo1]

Hi, thanks for your question. To answer your points:

1. The current model design indeed uses a fixed input–output size. We encourage you to try using our model trained on 280×518 resolution to directly run inference on smaller or larger images to see how stable the performance is under different resolutions. In addition, for obtaining higher-resolution images, we believe this can be achieved by adjusting the projection camera parameters and rendering from the existing Gaussians.

2. Increasing the image resolution will typically produce more Gaussian clusters, which makes the supervision during training more difficult.

[chf-git027]

The generated GS positions are converted into a point cloud for visualization, as shown in the figure. The result looks similar to the GS positions obtained by previous feed-forward GS methods. However, in my understanding, the GS positions produced by this method should be consistent with the real environment, similar to DGGT. Why does my visualization not look like that?

[primebo1]

Hi, I assume your visualization is in the ego-vehicle coordinate system (the point cloud appears to radiate outward). Have you tried visualizing the point cloud in the world coordinate system?