config.json

{
  "model_type": "sam3d_objects",
  "dtype": "bfloat16",
  "hidden_size": 1024,
  "num_heads": 16,
  "num_blocks": 24,
  "cond_channels": 1024,
  "latent_channels": 8,
  "latent_resolution": 16,
  "resolution": 64,
  "io_channels": 128,
  "decoder_channels": 768,
  "decoder_heads": 12,
  "decoder_blocks": 12,
  "window_size": 8,
  "structure_channels": [
    512,
    128,
    32
  ],
  "structure_res_blocks": 2,
  "dino_hidden_size": 1024,
  "dino_heads": 16,
  "dino_layers": 24,
  "dino_image_size": 518,
  "image_size": 518,
  "point_size": 256,
  "point_patch_size": 8,
  "point_channels": 512,
  "point_heads": 16,
  "ss_steps": 25,
  "slat_steps": 25,
  "ss_guidance": 7.0,
  "slat_guidance": 1.0,
  "ss_rescale_t": 3.0,
  "slat_rescale_t": 1.0,
  "downsample_ss_dist": 1,
  "slat_mean": [
    0.12211431,
    0.37204156,
    -1.26521907,
    -2.05276058,
    -3.10432536,
    -0.11294304,
    -0.85146744,
    0.45506954
  ],
  "slat_std": [
    2.37326008,
    2.13174402,
    2.2413953,
    2.30589401,
    2.1191894,
    1.8969511,
    2.41684989,
    2.08374642
  ],
  "depth_model": {
    "model_type": "moge3",
    "encoder": {
      "backbone": "dinov2_vitl14",
      "intermediate_layers": [
        5,
        11,
        17,
        23
      ],
      "dim_out": 1024
    },
    "neck": {
      "dim_in": [
        1026,
        2,
        2,
        2,
        2
      ],
      "dim_out": null,
      "dim_res_blocks": [
        1024,
        256,
        128,
        64,
        32
      ],
      "num_res_blocks": [
        0,
        2,
        2,
        2,
        0
      ],
      "res_block_in_norm": "none",
      "res_block_hidden_norm": "none",
      "resamplers": [
        "conv_transpose",
        "conv_transpose",
        "conv_transpose",
        "bilinear"
      ]
    },
    "points_head": {
      "dim_in": [
        1024,
        256,
        128,
        64,
        32
      ],
      "dim_out": [
        null,
        null,
        null,
        null,
        3
      ],
      "dim_res_blocks": [
        1024,
        256,
        128,
        64,
        32
      ],
      "num_res_blocks": [
        0,
        1,
        1,
        1,
        0
      ],
      "res_block_in_norm": "none",
      "res_block_hidden_norm": "none",
      "resamplers": [
        "conv_transpose",
        "conv_transpose",
        "conv_transpose",
        "bilinear"
      ]
    },
    "normal_head": {
      "dim_in": [
        1024,
        256,
        128,
        64,
        32
      ],
      "dim_out": [
        null,
        null,
        null,
        null,
        3
      ],
      "dim_res_blocks": [
        1024,
        256,
        128,
        64,
        32
      ],
      "num_res_blocks": [
        0,
        1,
        1,
        1,
        0
      ],
      "res_block_in_norm": "none",
      "res_block_hidden_norm": "none",
      "resamplers": [
        "conv_transpose",
        "conv_transpose",
        "conv_transpose",
        "bilinear"
      ]
    },
    "mask_head": {
      "dim_in": [
        1024,
        256,
        128,
        64,
        32
      ],
      "dim_out": [
        null,
        null,
        null,
        null,
        1
      ],
      "dim_res_blocks": [
        1024,
        256,
        128,
        64,
        32
      ],
      "num_res_blocks": [
        0,
        1,
        1,
        1,
        0
      ],
      "res_block_in_norm": "none",
      "res_block_hidden_norm": "none",
      "resamplers": [
        "conv_transpose",
        "conv_transpose",
        "conv_transpose",
        "bilinear"
      ]
    },
    "scale_head": {
      "dims": [
        1024,
        1024,
        1024,
        1
      ]
    },
    "num_tokens_range": [
      1200,
      3600
    ],
    "refiner": {
      "encoder_channels": 1026,
      "model_channels": [
        32,
        64,
        128,
        256,
        512
      ],
      "encoder_blocks_per_level": 1,
      "decoder_blocks_per_level": 1,
      "bottleneck_blocks": 1,
      "downsample_factors": [
        2,
        2,
        2,
        2
      ],
      "encoder_downsample": 16,
      "in_channels": 3,
      "out_channels": 1
    },
    "refiner_depth_resolution": 256
  }
}