configs/inference/svd_mv.yaml

model:
  base_learning_rate: 1.0e-05
  target: sgm.models.video_diffusion.DiffusionEngine
  params:
    ckpt_path: ckpts/svd_xt.safetensors
    scale_factor: 0.18215
    disable_first_stage_autocast: true
    scheduler_config:
      target: sgm.lr_scheduler.LambdaLinearScheduler
      params:
        warm_up_steps:
        - 1
        cycle_lengths:
        - 10000000000000
        f_start:
        - 1.0e-06
        f_max:
        - 1.0
        f_min:
        - 1.0
    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
          ucg_rate: 0.2
          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: true
          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
          params:
            outdim: 256
        - input_key: motion_bucket_id
          is_trainable: true
          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
          params:
            outdim: 256
        - input_key: cond_frames
          is_trainable: false
          ucg_rate: 0.2
          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: true
          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
    sampler_config:
      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
      params:
        num_steps: 30
        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
          params:
            sigma_max: 700.0
        guider_config:
          target: sgm.modules.diffusionmodules.guiders.LinearPredictionGuider
          params:
            max_scale: 2.5
            min_scale: 1.0
            num_frames: 24
    loss_fn_config:
      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
      params:
        batch2model_keys:
        - num_video_frames
        - image_only_indicator
        loss_weighting_config:
          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
          params:
            sigma_data: 1.0
        sigma_sampler_config:
          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
          params:
            p_mean: 0.3
            p_std: 1.2
data:
  target: sgm.data.objaverse.ObjaverseSpiralDataset
  params:
    root_dir: /mnt/mfs/zilong.chen/Downloads/objaverse-ndd-samples
    random_front: true
    batch_size: 2
    num_workers: 16
    cond_aug_mean: -0.0