TripoSR_config.yaml

cond_image_size: 512

image_tokenizer_cls: TripoSR.models.tokenizers.image.DINOSingleImageTokenizer
image_tokenizer:
  pretrained_model_name_or_path: "facebook/dino-vitb16"

tokenizer_cls: TripoSR.models.tokenizers.triplane.Triplane1DTokenizer
tokenizer:
  plane_size: 32
  num_channels: 1024

backbone_cls: TripoSR.models.transformer.transformer_1d.Transformer1D
backbone:
  in_channels: ${tokenizer.num_channels}
  num_attention_heads: 16
  attention_head_dim: 64
  num_layers: 16
  cross_attention_dim: 768

post_processor_cls: TripoSR.models.network_utils.TriplaneUpsampleNetwork
post_processor:
  in_channels: 1024
  out_channels: 40

decoder_cls: TripoSR.models.network_utils.NeRFMLP
decoder:
  in_channels: 120 # 3 * 40
  n_neurons: 64
  n_hidden_layers: 9
  activation: silu

renderer_cls: TripoSR.models.nerf_renderer.TriplaneNeRFRenderer
renderer:
  radius: 0.87 # slightly larger than 0.5 * sqrt(3)
  feature_reduction: concat
  density_activation: exp
  density_bias: -1.0
  num_samples_per_ray: 128