NVIDIAGameWorks / kaolin-wisp

NVIDIA Kaolin Wisp is a PyTorch library powered by NVIDIA Kaolin Core to work with neural fields (including NeRFs, NGLOD, instant-ngp and VQAD).
Other
1.46k stars 132 forks source link

Various bug fixes and quality improvements #167

Closed tovacinni closed 1 year ago

tovacinni commented 1 year ago

List of changes:

  1. New config arguments (especially in hash grid nerf) which improve quality significantly
  2. Adaptive ray batching (improves nerf quality a lot)
  3. bg_color is now a tuple (0.0, 0.0, 0.0) not a string
  4. Supporting other validation splits --valid-split
  5. apex based fused Adam
  6. Bug fix in rays for dist_min, inproper normalization in channel_fn
  7. 2D support for hash grid
  8. MultiTable utility class for working with hash grids (and similar) easier
  9. App renderer will now try to make FPS stable for better UX
  10. Fixing iteration count
  11. Fix tracker's tensorboard support