configs/model/sd.yaml

sd_base:
  symbol: sd
  find_unused_parameters: true

sd_autoencoder:
  type: autoencoderkl
  args:
    embed_dim: 4
    monitor: val/rec_loss
    ddconfig:
      double_z: true
      z_channels: 4
      resolution: 256
      in_channels: 3
      out_ch: 3
      ch: 128
      ch_mult: [1, 2, 4, 4]
      num_res_blocks: 2
      attn_resolutions: []
      dropout: 0.0
    lossconfig:
      target: torch.nn.Identity
  # pth: pretrained/kl-f8.pth
  hfm: ['shi-labs/versatile-diffusion-model', 'pretrained_pth/kl-f8.pth']

sd_t2i:
  super_cfg: sd_base
  type: sd_t2i
  args:
    first_stage_config: MODEL(sd_autoencoder)
    cond_stage_config: MODEL(clip_text_frozen)
    unet_config: MODEL(openai_unet_sd)
    beta_linear_start: 0.00085
    beta_linear_end: 0.012
    num_timesteps_cond: 1
    timesteps: 1000
    scale_factor: 0.18215
    use_ema: true

sd_t2i_noema:
  super_cfg: sd
  args:
    use_ema: false

#####################
# sd with full clip #
#####################

sd_t2i_fullclip_backward_compatible:
  super_cfg: sd_t2i
  args:
    cond_stage_config: MODEL(clip_frozen_encode_text_noproj)

sd_t2i_fullclip_backward_compatible_noema:
  super_cfg: sd_t2i_noema
  args:
    cond_stage_config: MODEL(clip_frozen_encode_text_noproj)

sd_t2i_fullclip:
  super_cfg: sd_t2i
  args:
    cond_stage_config: MODEL(clip_frozen_encode_text)

sd_variation:
  super_cfg: sd_t2i
  type: sd_variation
  args:
    cond_stage_config: MODEL(clip_vision_frozen_justin)