configs/inference/config_motionctrl_cmcm.yaml

model:
  base_learning_rate: 3.0e-5
  target: sgm.motionctrl.camera_motion_control.CameraMotionControl
  params:
    ckpt_path: /group/30098/zhouxiawang/env/share/weights/svd/stable-video-diffusion-img2vid/svd.safetensors
    scale_factor: 0.18215
    input_key: video
    no_cond_log: true
    en_and_decode_n_samples_a_time: 1
    use_ema: false
    disable_first_stage_autocast: true

    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:
        num_frames: 14
        adm_in_channels: 768
        num_classes: sequential
        use_checkpoint: false
        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
          ucg_rate: 0.1
          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
                # version: "/apdcephfs_cq3/share_1290939/vg_zoo/dependencies/OpenCLIP-ViT-H-14-laion2B-s32B-b79K/blobs/9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4"

        - 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
          ucg_rate: 0.1
          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.AutoencodingEngine
      params:
        loss_config:
          target: torch.nn.Identity
        regularizer_config:
          target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer
        encoder_config: 
          target: sgm.modules.diffusionmodules.model.Encoder
          params:
            attn_type: vanilla
            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
        decoder_config:
          target: sgm.modules.autoencoding.temporal_ae.VideoDecoder
          params:
            attn_type: vanilla
            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
            video_kernel_size: [3, 1, 1]

    # loss_fn_config:
    #   target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
    #   params:
    #     batch2model_keys: ['RT']  
    #     loss_weighting_config:
    #       target: sgm.modules.diffusionmodules.loss_weighting.VWeighting
    #     sigma_sampler_config:
    #       target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
    #       params:
    #         p_mean: 1.0
    #         p_std: 1.6

    sampler_config:
      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
      params:
        num_steps: 25
        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
          params:
            sigma_max: 700.0

        guider_config:
          target: sgm.modules.diffusionmodules.guiders.LinearPredictionGuider
          params:
            num_frames: 14
            max_scale: 2.5
            min_scale: 1.0