White edges at the boundary during DMTET geometry refine

[DevLinyan]

I am training a model with DMTET in the second stage using zero123, but during the training process, I noticed that there are white edges at the boundary between the foreground and the background. I am not sure what could be causing this issue. And it occurs in many samples.

Here are the first stage training script and the geometry refine training script.

name: "zero123"
tag: "${data.random_camera.height}_${rmspace:${basename:${data.image_path}},_}"
exp_root_dir: "outputs"
seed: 0

data_type: "single-image-datamodule"
data: # threestudio/data/image.py -> SingleImageDataModuleConfig
  image_path: ./load/images/hamburger_rgba.png
  height: [128, 256, 512]
  width: [128, 256, 512]
  resolution_milestones: [200, 300]
  default_elevation_deg: 15.0
  default_azimuth_deg: 0.0
  default_camera_distance: 3.8
  default_fovy_deg: 20.0
  requires_depth: ${cmaxgt0orcmaxgt0:${system.loss.lambda_depth},${system.loss.lambda_depth_rel}}
  requires_normal: ${cmaxgt0:${system.loss.lambda_normal}}
  random_camera: # threestudio/data/uncond.py -> RandomCameraDataModuleConfig
    height: [64, 128, 256]
    width: [64, 128, 256]
    batch_size: [6,4,2] #[12, 8, 4]
    resolution_milestones: [200, 300]
    eval_height: 512
    eval_width: 512
    eval_batch_size: 1
    elevation_range: [-10, 80]
    azimuth_range: [-180, 180]
    camera_distance_range: [3.8, 3.8]
    fovy_range: [20.0, 20.0] # Zero123 has fixed fovy
    progressive_until: 0
    camera_perturb: 0.0
    center_perturb: 0.0
    up_perturb: 0.0
    light_position_perturb: 1.0
    light_distance_range: [7.5, 10.0]
    eval_elevation_deg: ${data.default_elevation_deg}
    eval_camera_distance: ${data.default_camera_distance}
    eval_fovy_deg: ${data.default_fovy_deg}
    light_sample_strategy: "dreamfusion"
    batch_uniform_azimuth: False
    n_val_views: 30
    n_test_views: 120

system_type: "zero123-system"
system:
  geometry_type: "implicit-volume"
  geometry:
    radius: 2.0
    normal_type: "analytic"

    # use Magic3D density initialization instead
    density_bias: "blob_magic3d"
    density_activation: softplus
    density_blob_scale: 10.
    density_blob_std: 0.5

    # coarse to fine hash grid encoding
    # to ensure smooth analytic normals
    pos_encoding_config:
      otype: HashGrid
      n_levels: 16
      n_features_per_level: 2
      log2_hashmap_size: 19
      base_resolution: 16
      per_level_scale: 1.447269237440378 # max resolution 4096
    mlp_network_config:
      otype: "VanillaMLP"
      activation: "ReLU"
      output_activation: "none"
      n_neurons: 64
      n_hidden_layers: 2

  material_type: "diffuse-with-point-light-material"
  material:
    ambient_only_steps: 100000
    textureless_prob: 0.05
    albedo_activation: sigmoid

  background_type: "solid-color-background" # unused

  renderer_type: "nerf-volume-renderer"
  renderer:
    radius: ${system.geometry.radius}
    num_samples_per_ray: 512
    return_comp_normal: ${cmaxgt0:${system.loss.lambda_normal_smooth}}
    return_normal_perturb: ${cmaxgt0:${system.loss.lambda_3d_normal_smooth}}

  prompt_processor_type: "dummy-prompt-processor" # Zero123 doesn't use prompts
  prompt_processor:
    pretrained_model_name_or_path: ""
    prompt: ""

  guidance_type: "stable-zero123-guidance"
  guidance:
    pretrained_config: "./load/zero123/sd-objaverse-finetune-c_concat-256.yaml"
    pretrained_model_name_or_path: "stable_zero123.ckpt"
    vram_O: ${not:${gt0:${system.freq.guidance_eval}}}
    cond_image_path: ${data.image_path}
    cond_elevation_deg: ${data.default_elevation_deg}
    cond_azimuth_deg: ${data.default_azimuth_deg}
    cond_camera_distance: ${data.default_camera_distance}
    guidance_scale: 3.0
    min_step_percent: [50, 0.7, 0.3, 200]  # (start_iter, start_val, end_val, end_iter)
    max_step_percent: [50, 0.98, 0.8, 200]

  freq:
    ref_only_steps: 0
    guidance_eval: 0

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

  loss:
    lambda_sds: 0.1
    lambda_rgb: [100, 500., 1000., 400]
    lambda_mask: 50.
    lambda_depth: 0. # 0.05
    lambda_depth_rel: 0. # [0, 0, 0.05, 100]
    lambda_normal: 0. # [0, 0, 0.05, 100]
    lambda_normal_smooth: [100, 7.0, 5.0, 150, 10.0, 200]
    lambda_3d_normal_smooth: [100, 7.0, 5.0, 150, 10.0, 200]
    lambda_orient: 1.0
    lambda_sparsity: 0.5 # should be tweaked for every model
    lambda_opaque: 0.5

  optimizer:
    name: Adam
    args:
      lr: 0.01
      betas: [0.9, 0.99]
      eps: 1.e-8

trainer:
  max_steps: 1200
  log_every_n_steps: 1
  num_sanity_val_steps: 0
  val_check_interval: 100
  enable_progress_bar: true
  precision: 32

checkpoint:
  save_last: true # save at each validation time
  save_top_k: -1
  every_n_train_steps: 200 # ${trainer.max_steps}

The second stage

# zero123-geometric


name: "zero123_geom"
tag: "${data.random_camera.height}_${rmspace:${basename:${data.image_path}},_}_prog${data.random_camera.progressive_until}"
exp_root_dir: "outputs"
seed: 0

data_type: "single-image-datamodule"
data: # threestudio/data/image.py -> SingleImageDataModuleConfig
  image_path: ./load/images/hamburger_rgba.png
  height: 512
  width: 512
  default_elevation_deg: 15.0
  default_azimuth_deg: 0.0
  default_camera_distance: 3.8
  default_fovy_deg: 20.0
  requires_depth: ${cmaxgt0orcmaxgt0:${system.loss.lambda_depth},${system.loss.lambda_depth_rel}}
  requires_normal: ${cmaxgt0:${system.loss.lambda_normal}}
  random_camera: # threestudio/data/uncond.py -> RandomCameraDataModuleConfig
    height: 256 # Zero123 operates at 256x256
    width: 256
    batch_size: 8
    resolution_milestones: []
    eval_height: 512
    eval_width: 512
    eval_batch_size: 1
    elevation_range: [-10, 45]
    azimuth_range: [-180, 180]
    camera_distance_range: [3.8, 3.8]
    fovy_range: [20.0, 20.0] # Zero123 has fixed fovy
    progressive_until: 0
    camera_perturb: 0.0
    center_perturb: 0.0
    up_perturb: 0.0
    light_position_perturb: 1.0
    light_distance_range: [7.5, 10.0]
    eval_elevation_deg: ${data.default_elevation_deg}
    eval_camera_distance: ${data.default_camera_distance}
    eval_fovy_deg: ${data.default_fovy_deg}
    light_sample_strategy: "dreamfusion"
    batch_uniform_azimuth: False
    n_val_views: 30
    n_test_views: 120

system_type: "zero123-system"
system:
  refinement: true
  geometry_convert_from: ???
  geometry_convert_inherit_texture: true
  geometry_type: "tetrahedra-sdf-grid"
  geometry:
    radius: 2.0 # consistent with coarse
    isosurface_resolution: 128
    isosurface_deformable_grid: true
    pos_encoding_config: # consistent with coarse, no progressive band
      otype: HashGrid
      n_levels: 16
      n_features_per_level: 2
      log2_hashmap_size: 19
      base_resolution: 16
      per_level_scale: 1.4472692374403782 # max resolution 4096
    mlp_network_config:
      otype: "VanillaMLP"
      activation: "ReLU"
      output_activation: "none"
      n_neurons: 64
      n_hidden_layers: 2
    fix_geometry: false # optimize grid sdf and deformation

  # material_type: "no-material" # unused
  # material:
  #   n_output_dims: 0

  material_type: "diffuse-with-point-light-material"
  material:
    ambient_only_steps: 10000
    textureless_prob: 0.

  background_type: "solid-color-background" # unused

  # renderer_type: "nerf-volume-renderer"
  # renderer:
  #   radius: ${system.geometry.radius}
  #   num_samples_per_ray: 512
  #   return_comp_normal: ${gt0:${system.loss.lambda_normal_smooth}}
  #   return_normal_perturb: ${gt0:${system.loss.lambda_3d_normal_smooth}}

  renderer_type: "nvdiff-rasterizer"
  renderer:
    context_type: cuda

  prompt_processor_type: "dummy-prompt-processor" # Zero123 doesn't use prompts
  prompt_processor:
    pretrained_model_name_or_path: ""
    prompt: ""

  guidance_type: "stable-zero123-guidance"
  guidance:
    pretrained_config: "./load/zero123/sd-objaverse-finetune-c_concat-256.yaml"
    pretrained_model_name_or_path: "stable_zero123.ckpt"
    vram_O: ${not:${gt0:${system.freq.guidance_eval}}}
    cond_image_path: ${data.image_path}
    cond_elevation_deg: ${data.default_elevation_deg}
    cond_azimuth_deg: ${data.default_azimuth_deg}
    cond_camera_distance: ${data.default_camera_distance}
    guidance_scale: 3.0
    min_step_percent: 0.02
    # min_step_percent: [0, 0.4, 0.02, 2000]  # (start_iter, start_val, end_val, end_iter)
    max_step_percent: 0.5
    # max_step_percent: [0, 0.85, 0.85, 2000]

  freq:
    ref_only_steps: 0
    guidance_eval: 0  # if "alternate", this must not be a multiple of system.freq.n_ref

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

  loss:
    lambda_sds: 1.0
    lambda_rgb: 10000.0
    lambda_mask: 100.0
    lambda_depth: 0.
    lambda_depth_rel: 0. # [0.0, 0.0, 1.0, 10000]
    lambda_normal: 0. # [0, 0, 0.05, 100]
    lambda_normal_smooth: 0.
    lambda_3d_normal_smooth: 0.
    lambda_normal_consistency: [50, 10000.0, 1000.0, 51] # 10000.
    lambda_laplacian_smoothness: 0.
    lambda_orient: 0.
    lambda_sparsity: 0.
    lambda_opaque: 0.

  optimizer:
    name: Adam
    args:
      lr: 0.01
      betas: [0.9, 0.99]
      eps: 1.e-8

trainer:
  max_steps: 200
  log_every_n_steps: 1
  num_sanity_val_steps: 0
  val_check_interval: 50
  enable_progress_bar: true
  precision: 16-mixed

checkpoint:
  save_last: true # save at each validation time
  save_top_k: -1
  every_n_train_steps: ${trainer.max_steps}