Scenes trained with MCMC method fail after conversion to .ksplat

[williefalloon]

Hi, this issue has been occurring for some time and I haven't seen it mentioned here so want to bring it to your attention.

If these conditions are met, the scene will fail to load, or take a very long time to load.

• Trained using MCMC profile in PostShot (default training profile in PS)
• The scene has a high (but well under published limits) splat count
• Conversion to ksplat

I have some files below with pass and fail cases from my testing.

• PASS: https://drive.google.com/file/d/1n2ALxAAr1noyx-bZjZPJXl8QtfykwTRh/view?usp=sharing

  • MCMC profile
  • 3m splats
  • .ply file

• FAIL: https://drive.google.com/file/d/1fpvMebP12oWcLIy6tMnXaLGYJJlUamcI/view?usp=sharing

  • Same MCMC scene
  • 3m splats
  • .ksplat (converted on Mkellogg demo site)

• PASS: https://drive.google.com/file/d/1V69PjuEUwHSe_JSCPQcM18FsVlXNiKsj/view?usp=sharing

  • MCMC scene
  • Approx 1m splats
  • .ksplat (converted on demo site)

Points to note:

• It seems that the critical point is the ksplat conversion.
• SH level settings do not affect pass or fails.
• Scenes trained with the ADC profile in PostShot and cropped do not have the issue.
• The scenes often do load eventually, but CPU usage is at max for quite a long time before it succeeds.

[mkkellogg]

I think it might have something to do with the dimensions of your scene, which appear to be somewhat large, which can be addressed by changing the "block size" parameter when converting (bigger block sizes work better for scenes with larger dimensions). I used a block size of 20 with a bucket size of 512 and that seemed to work well. My .ksplat compression scheme is fairly janky, so it doesn't automatically handle non-optimal conditions very well 🙂

[mkkellogg]

I'll close this for now, but let me know if you know any more help