{
  "_class_name": "AutoencoderKL",
  "in_channels": 3,
  "out_channels": 3,
  "latent_channels": 4,
  "block_out_channels": [128, 256, 512],
  "layers_per_block": 2
}