@simonbethke I tried to address the false transparency using the skin/void method (see https://superspl.at/scene/0765075f). Not perfect (somehow created a 'swimming' navel that I need to diagnose), but seems to mostly fix the false transparency. The splatter was borrowing splats from the background it couldn't photometrically distinguish as belonging to the front or the back.

@simonbethke, I think the false transparency comes from a lighting model the SH bands can't quite capture. Currently trying to give the diffusion process more hints to create a globally fixed lighting source rather than one that travels with the camera. The alternative is likely forcing a mostly opaque 'skin' by creating a body shaped void with high placement loss. The trainer isn't really meant to be general purpose. It's also CUDA only, as it implements some recent papers that allow leveraging of the tensor cores which is not trivial to implement in Brush' backend agnostic framework. Brush had become the slowest step in the process by quite a wide margin and there wasn't much left to optimize.

Wow, thats the last thing I expected: A new brush-replacement :D However, your splats often have plenty of false-transparency.