refine/refine.yaml

name: "refine"
tag: "${rmspace:${system.prompt_processor.prompt},_}"
exp_root_dir: "outputs"
seed: 0

data_type: "random-camera-datamodule"
data:
  batch_size: 1
  width: 64
  height: 64
  camera_distance_range: [2.5, 3.0]
  fovy_range: [40, 70]
  elevation_range: [-10, 60]
  light_sample_strategy: "dreamfusion"
  eval_camera_distance: 3.5
  eval_fovy_deg: 70.
  eval_elevation_deg: 10

system_type: "dreamfusion-system"
system:
  geometry_type: "implicit-volume"
  geometry:
    radius: 1.0
    normal_type: finite_difference
    finite_difference_normal_eps: 0.01

    density_bias: 0.0
    density_activation: trunc_exp

    pos_encoding_config:
      otype: Volume
      channel: 32
      resolution: 64

    mlp_network_config:
      otype: VanillaMLP
      activation: ReLU
      output_activation: none
      n_neurons: 256
      n_hidden_layers: 4
      bias: True

  material_type: "diffuse-with-point-light-material"
  material:
    ambient_only_steps: 0
    albedo_activation: scale_-11_01

  background_type: "neural-environment-map-background"
  background:
    color_activation: scale_-11_01

  renderer_type: "nerf-volume-renderer"
  renderer:
    radius: ${system.geometry.radius}
    num_samples_per_ray: 512

  prompt_processor_type: "deep-floyd-prompt-processor"
  prompt_processor:
    pretrained_model_name_or_path: "DeepFloyd/IF-I-XL-v1.0"
    prompt: ???
    no_view_dependent_prompt: true

  guidance_type: "deep-floyd-guidance"
  guidance:
    pretrained_model_name_or_path: "DeepFloyd/IF-I-XL-v1.0"
    guidance_scale: 20.
    weighting_strategy: sds
    min_step_percent: 0.02
    max_step_percent: 0.98

  loggers:
    wandb:
      enable: false
      project: 'threestudio'
      name: None

  loss:
    lambda_sds: 1.
    lambda_orient: 1.
    lambda_sparsity: 0.
    lambda_opaque: 0.0
  optimizer:
    name: Adam
    args:
      lr: 1.e-2
      betas: [0.9, 0.99]
      eps: 1.e-15
    params:
      geometry.encoding:
        lr: 1.0e-2
      geometry.density_network:
        lr: 1.0e-6
      geometry.feature_network:
        lr: 1.0e-3

trainer:
  max_steps: 1000
  log_every_n_steps: 1
  num_sanity_val_steps: 0
  val_check_interval: 100
  enable_progress_bar: true
  precision: 16-mixed

checkpoint:
  save_last: true
  save_top_k: -1
  every_n_train_steps: ${trainer.max_steps}