Config Notes

April 26, 2026 ยท View on GitHub

The released training config is:

  • configs/dino_v3_large_sat_0906.json

Important fields:

  • opt.backbone: The image encoder backbone. This release uses dinov3-large-sat.
  • opt.representation_type: The 3D representation type. This release uses oneplane.
  • opt.pad: Add spatial padding mode before decoding.
  • opt.network.no_hidden_states: Uses only the final DINO features.
  • opt.network.no_cls_token: Removes the CLS token branch from the decoder input.
  • opt.network.color_channels: Feature width used by the renderer and super-resolution head.
  • opt.network.sr_factor: Whether the 2x super-resolution head is enabled.
  • opt.network.point_sampling_kwargs.num_points: Number of coarse ray samples.
  • opt.network.point_sampling_kwargs.num_importance: Number of importance samples.
  • opt.network.ray_marching_kwargs.density_clamp_mode: Density activation used during ray marching.
  • opt.network.triplane.dim: Plane feature dimension.
  • opt.network.triplane.ch_mult: Decoder channel multiplier schedule.

JSON does not support inline comments, so this file is the released parameter note for the kept config.