model_params:
  appearance_feature_extractor_params: # the F in the paper
    image_channel: 3
    block_expansion: 64
    num_down_blocks: 2
    max_features: 512
    reshape_channel: 32
    reshape_depth: 16
    num_resblocks: 6
  motion_extractor_params: # the M in the paper
    num_kp: 21
    backbone: convnextv2_tiny
  warping_module_params: # the W in the paper
    num_kp: 21
    block_expansion: 64
    max_features: 512
    num_down_blocks: 2
    reshape_channel: 32
    estimate_occlusion_map: True
    dense_motion_params:
      block_expansion: 32
      max_features: 1024
      num_blocks: 5
      reshape_depth: 16
      compress: 4
  spade_generator_params: # the G in the paper
    upscale: 2 # represents upsample factor 256x256 -> 512x512
    block_expansion: 64
    max_features: 512
    num_down_blocks: 2
  stitching_retargeting_module_params: # the S in the paper
    stitching:
      input_size: 126 # (21*3)*2
      hidden_sizes: [128, 128, 64]
      output_size: 65 # (21*3)+2(tx,ty)
    lip:
      input_size: 65 # (21*3)+2
      hidden_sizes: [128, 128, 64]
      output_size: 63 # (21*3)
    eye:
      input_size: 66 # (21*3)+3
      hidden_sizes: [256, 256, 128, 128, 64]
      output_size: 63 # (21*3)