Ref2VA/video_vae/source/config.json

{
  "_class_name": "AutoencoderKLLegacy",
  "_diffusers_version": "0.32.2",
  "causal_decoder": false,
  "causal_encoder": true,
  "ch": 128,
  "ch_mult": [
    1,
    2,
    2,
    4,
    4,
    8
  ],
  "embed_dim": 24,
  "in_channels": 3,
  "num_res_blocks": 2,
  "num_res_blocks_decoder": null,
  "out_ch": 3,
  "padding_mode": "reflect",
  "padding_mode_t": null,
  "pixel_norm_type": "imagenet",
  "scaling_factor": 1.0,
  "shift_factor": 0.0,
  "space_down": [
    2,
    2,
    2,
    2,
    1,
    1
  ],
  "space_up": [
    1,
    2,
    2,
    2,
    2,
    1
  ],
  "time_down": [
    1,
    2,
    2,
    1,
    1,
    1
  ],
  "time_up": null,
  "use_3d_conv": true,
  "use_t_isolated_gn": true,
  "use_vit_decoder": true,
  "vae_ratio": 16,
  "vae_ratio_t": 4,
  "vit_decoder_kwargs": {
    "dim_head": 64,
    "ffn_activation_fn": "silu",
    "ffn_use_gated": true,
    "heads": 32,
    "norm_affine": true,
    "norm_type": "rms_norm",
    "num_layers": 36,
    "qk_norm_affine": false,
    "qk_norm_type": "rms_norm",
    "rope_dim_ratio": 0.75,
    "rope_theta": 100.0
  },
  "z_channels": 24,
  "zq_ch_decoder": null,
  "zq_ch_encoder": null
}