Improving quality of views

[pooncs]

Hi I am trying to increase the quality of the views that I am getting.

So I am training about 30 images at 4K resolution with run.py, and rendering the image with the same resolution for 35k steps, which the loss stop decreasing much.
In the paper, it seems like the rendered images were able to match the ground truth (gt) quite well.

Shown here is the (zoomed in, about 8x) gt which I am trying to get close to

After rendering (zoomed in (~8x) to the details)

Increasing the scale and reducing aabb_scale to 1 seemed to help. I have also changed the input image resolution to 2K and 1K, but the rendered image seemed similar. So I was wondering if there is some internal limiter to the input image resolution, or grid size scaling/stepping that is causing this (since i do not really need the render to show in real-time)?

Thanks!