scripts/sampling/configs/svd_xt_image_decoder.yaml

model:
  target: sgm.models.diffusion.DiffusionEngine
  params:
    scale_factor: 0.18215
    disable_first_stage_autocast: True
    ckpt_path: checkpoints/svd_xt_image_decoder.safetensors

    denoiser_config:
      target: sgm.modules.diffusionmodules.denoiser.Denoiser
      params:
        scaling_config:
          target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise

    network_config:
      target: sgm.modules.diffusionmodules.video_model.VideoUNet
      params:
        adm_in_channels: 768
        num_classes: sequential
        use_checkpoint: True
        in_channels: 8
        out_channels: 4
        model_channels: 320
        attention_resolutions: [4, 2, 1]
        num_res_blocks: 2
        channel_mult: [1, 2, 4, 4]
        num_head_channels: 64
        use_linear_in_transformer: True
        transformer_depth: 1
        context_dim: 1024
        spatial_transformer_attn_type: softmax-xformers
        extra_ff_mix_layer: True
        use_spatial_context: True
        merge_strategy: learned_with_images
        video_kernel_size: [3, 1, 1]

    conditioner_config:
      target: sgm.modules.GeneralConditioner
      params:
        emb_models:
        - is_trainable: False
          input_key: cond_frames_without_noise
          target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder
          params:
            n_cond_frames: 1
            n_copies: 1
            open_clip_embedding_config:
              target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder
              params:
                freeze: True

        - input_key: fps_id
          is_trainable: False
          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
          params:
            outdim: 256

        - input_key: motion_bucket_id
          is_trainable: False
          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
          params:
            outdim: 256

        - input_key: cond_frames
          is_trainable: False
          target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder
          params:
            disable_encoder_autocast: True
            n_cond_frames: 1
            n_copies: 1
            is_ae: True
            encoder_config:
              target: sgm.models.autoencoder.AutoencoderKLModeOnly
              params:
                embed_dim: 4
                monitor: val/rec_loss
                ddconfig:
                  attn_type: vanilla-xformers
                  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

        - input_key: cond_aug
          is_trainable: False
          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
          params:
            outdim: 256

    first_stage_config:
      target: sgm.models.autoencoder.AutoencoderKL
      params:
        embed_dim: 4
        monitor: val/rec_loss
        ddconfig:
          attn_type: vanilla-xformers
          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

    sampler_config:
      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
      params:
        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
          params:
            sigma_max: 700.0

        guider_config:
          target: sgm.modules.diffusionmodules.guiders.LinearPredictionGuider
          params:
            max_scale: 3.0
            min_scale: 1.5