{
  "format_version": 1,
  "model_type": "vggt-diff",
  "model_family": "nvs",
  "checkpoint_file": "vggt-diff.safetensors",
  "base_model": "Wan-AI/Wan2.1-I2V-14B-480P",
  "geometry_encoder": "facebook/VGGT-Omega",
  "resolution": [192, 336],
  "source_views": 6,
  "target_views": {
    "type": "variable_independent_views",
    "maximum_during_training": 16
  },
  "vae": {
    "frozen": true,
    "individual_encoding": true,
    "temporal_target_compression": 1
  },
  "conditioning": {
    "geometry_features": "frozen",
    "camera_normalization": "source_anchor",
    "temporal_plucker_packing": false,
    "point_track_consistency": true
  },
  "training": {
    "dataset": "DL3DV clean-10K with benchmark overlap excluded",
    "available_scenes": 9510,
    "finetune_mode": "full",
    "release_checkpoint": "100K effective steps",
    "original_fixed_compute_checkpoint": "18K effective steps",
    "optimization_status": "Best checkpoint in the reported 18K-to-100K half-resolution sweep; SSIM and LPIPS remained improving at 100K"
  }
}