Antialiased Rendering Mode Causes Disappearance of Many Small Gaussians

[22331168]

I trained a model using simple_trainer.py and viewed it with simple_viewer. When I switched the Anti-Aliasing mode in the UI from "classic" to "antialiased", many fine details—such as bicycle spokes and leaves—disappeared entirely.

This behavior seems inconsistent with the results shown in the original mip-Splatting paper. The issue is not specific to my model: I observed the same problem with models trained using the standard gaussian_splatting codebase as well.

I'm using the default 2D epsilon value of 0.3, and adjusting it (either increasing or decreasing) does not resolve the problem.

My hypothesis is that in antialiased mode, small Gaussians undergo significant opacity attenuation after 2D filtering, causing their opacity to drop below visible thresholds. However, based on the mip-Splatting results, this level of detail loss shouldn't occur—the visual degradation appears much more severe than expected.

Any insights or suggestions would be greatly appreciated