vae:
  use_downsample: true
  num_latents: 256
  point_feats: 3
  out_dim: 1
  embed_dim: 64
  width: 768
  heads: 12
  num_encoder_layers: 8
  num_decoder_layers: 16
  init_scale: 0.25
  qkv_bias: false
  use_ln_post: true
  use_udf_extraction: true
  token_scales:
  - 128.0
  - 256.0
  - 384.0
  - 512.0
  - 640.0
  - 1024.0
  - 2048.0
  token_probability:
  - 0.025
  - 0.025
  - 0.025
  - 0.025
  - 0.05
  - 0.2
  - 0.65