model:
  target: ldm.models.diffusion.ddpm.LatentDiffusion
  params:
    linear_start: 0.00085
    linear_end: 0.0120
    shift_scale: 2
    num_timesteps_cond: 1
    log_every_t: 200
    timesteps: 1000
    first_stage_key: "triplane"
    cond_stage_key: "caption"
    image_size: 32
    channels: 8
    cond_stage_trainable: false
    conditioning_key: crossattn
    monitor: val/loss_simple_ema
    scale_factor: 0.5147210212065061
    use_ema: False
    learning_rate: 5e-5

    unet_config:
      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
      params:
        image_size: 32
        in_channels: 8
        out_channels: 8
        model_channels: 320
        attention_resolutions: [4, 2, 1]
        num_res_blocks: 2
        channel_mult: [ 1, 2, 4, 4 ]
        num_heads: 8
        use_spatial_transformer: True
        context_dim: 768
        transformer_depth: 1
        use_checkpoint: True
        legacy: False

    first_stage_config:
      target: model.triplane_vae.AutoencoderKLRollOut
      params:
        embed_dim: 8
        learning_rate: 1e-5
        norm: False
        renderer_type: eg3d
        ddconfig:
          double_z: true
          z_channels: 8
          resolution: 256
          in_channels: 32
          out_ch: 32
          ch: 128
          ch_mult:
          - 2
          - 4
          - 4
          - 8
          num_res_blocks: 2
          attn_resolutions: [32]
          dropout: 0.0
        lossconfig:
          kl_weight: 1e-5
          rec_weight: 1
          latent_tv_weight: 2e-3
        renderer_config:
          rgbnet_dim: -1
          rgbnet_width: 128
          sigma_dim: 12
          c_dim: 20

    cond_stage_config:
      target: ldm.modules.encoders.modules.FrozenCLIPTextEmbedder