configs/inference/default.yaml

source_image: ./default.png
driving_audio: default.wav

weight_dtype: fp16

data:
  n_motion_frames: 2
  n_sample_frames: 16
  source_image:
    width: 512
    height: 512
  driving_audio:
    sample_rate: 16000
  export_video:
    fps: 25

inference_steps: 40
cfg_scale: 3.5

audio_ckpt_dir: ./pretrained_models/hallo

base_model_path: ./pretrained_models/stable-diffusion-v1-5

motion_module_path: ./pretrained_models/motion_module/mm_sd_v15_v2.ckpt

face_analysis:
  model_path: ./pretrained_models/face_analysis

wav2vec:
  model_path: ./pretrained_models/wav2vec/wav2vec2-base-960h
  features: all

audio_separator:
  model_path: ./pretrained_models/audio_separator/Kim_Vocal_2.onnx

vae:
  model_path: ./pretrained_models/sd-vae-ft-mse

save_path: ./.cache

face_expand_ratio: 1.1
pose_weight: 1.1
face_weight: 1.1
lip_weight: 1.1

unet_additional_kwargs:
  use_inflated_groupnorm: true
  unet_use_cross_frame_attention: false
  unet_use_temporal_attention: false
  use_motion_module: true
  use_audio_module: true
  motion_module_resolutions:
    - 1
    - 2
    - 4
    - 8
  motion_module_mid_block: true
  motion_module_decoder_only: false
  motion_module_type: Vanilla
  motion_module_kwargs:
    num_attention_heads: 8
    num_transformer_block: 1
    attention_block_types:
      - Temporal_Self
      - Temporal_Self
    temporal_position_encoding: true
    temporal_position_encoding_max_len: 32
    temporal_attention_dim_div: 1
  audio_attention_dim: 768
  stack_enable_blocks_name:
    - "up"
    - "down"
    - "mid"
  stack_enable_blocks_depth: [0,1,2,3]
  

enable_zero_snr: true

noise_scheduler_kwargs:
  beta_start: 0.00085
  beta_end: 0.012
  beta_schedule: "linear"
  clip_sample: false
  steps_offset: 1
  ### Zero-SNR params
  prediction_type: "v_prediction"
  rescale_betas_zero_snr: True
  timestep_spacing: "trailing"

sampler: DDIM