vae/config.json

{
  "_class_name": "AutoencoderKLMagvit",
  "_diffusers_version": "0.30.1",
  "act_fn": "silu",
  "block_out_channels": [
    128,
    256,
    512,
    512
  ],
  "cache_compression_vae": false,
  "cache_mag_vae": false,
  "ch": 128,
  "ch_mult": [
    1,
    2,
    4,
    4
  ],
  "down_block_types": [
    "SpatialDownBlock3D",
    "SpatialTemporalDownBlock3D",
    "SpatialTemporalDownBlock3D",
    "SpatialTemporalDownBlock3D"
  ],
  "force_upcast": true,
  "in_channels": 3,
  "latent_channels": 4,
  "layers_per_block": 2,
  "mid_block_attention_type": "3d",
  "mid_block_num_attention_heads": 1,
  "mid_block_type": "MidBlock3D",
  "mid_block_use_attention": true,
  "mini_batch_decoder": 2,
  "mini_batch_encoder": 8,
  "norm_num_groups": 32,
  "num_attention_heads": 1,
  "out_channels": 3,
  "sample_size": 256,
  "scaling_factor": 0.18215,
  "slice_compression_vae": true,
  "slice_mag_vae": true,
  "spatial_group_norm": false,
  "tile_overlap_factor": 0.25,
  "tile_sample_min_size": 384,
  "up_block_types": [
    "SpatialUpBlock3D",
    "SpatialTemporalUpBlock3D",
    "SpatialTemporalUpBlock3D",
    "SpatialTemporalUpBlock3D"
  ],
  "upcast_vae": false,
  "use_gc_blocks": null,
  "use_tiling": true,
  "use_tiling_decoder": false,
  "use_tiling_encoder": false
}