base_model_path: ./pretrained_weights/sd-image-variations-diffusers
checkpointing_steps: 1000
data:
  crop_scale:
  - 0.8
  - 1.2
  do_center_crop: false
  meta_paths:
  - /workspace/develop/video/data/ubc_tiktok-dropout0.03/ubc-meta.json
  - /workspace/develop/video/data/tiktok-dance/good-meta.json
  - /workspace/develop/video/data/20240321/meta.json
  - /workspace/develop/video/data/20240327/meta.json
  - /workspace/develop/video/data/20240506/meta.json
  - /workspace/develop/video/data/20240509/meta.json
  ref_augment:
    downsample:
      min_scale_logit: -1.2
      p: 0.3
    pan:
    - 0.04
    - 0.02
    rotate: 8
    scale:
    - 0.9
    - 1.2
  sample_margin: 30
  train_bs: 4
  train_height: 1152
  train_width: 768
enable_zero_snr: true
exp_name: stage1-openpg-nopaf
image_encoder_path: ./pretrained_weights/sd-image-variations-diffusers/image_encoder
noise_offset: 0.05
noise_scheduler_kwargs:
  beta_end: 0.012
  beta_schedule: scaled_linear
  beta_start: 0.00085
  clip_sample: false
  num_train_timesteps: 1000
  steps_offset: 1
openpose_guider:
  block_out_channels:
  - 96
  - 192
  enable: true
  exclude_paf: true
  model_path: ./pretrained_weights/body_pose_model.pth
output_dir: /workspace/camus/train
pose_guider_pretrain: false
resume_from_checkpoint: latest
save_dir: /workspace/camus/train/20240508-stage1-openpg-nopaf
save_model_epoch_interval: 1
seed: 12580
snr_gamma: 5.0
solver:
  adam_beta1: 0.9
  adam_beta2: 0.999
  adam_epsilon: 1.0e-08
  adam_weight_decay: 0.01
  enable_xformers_memory_efficient_attention: true
  gradient_accumulation_steps: 1
  gradient_checkpointing: false
  learning_rate: 1.0e-05
  lr_scheduler: constant
  lr_warmup_steps: 1
  max_grad_norm: 1.0
  max_train_steps: 100000
  mixed_precision: fp16
  scale_lr: false
  use_8bit_adam: false
uncond_ratio: 0.1
vae_model_path: ./pretrained_weights/sd-vae-ft-mse
val:
  special_steps:
  - 200
  validation_steps: 1000
validation:
  metric:
    batch_size: 4
    generated_frames:
    - 16
    - 45
    - 98
    - 150
    - 188
    - 220
    - 268
    - 300
    guidance_scale: 2.8
    ref_frame: 28
    seed: 42
    size:
    - 640
    - 960
    steps: 20
    videos:
    - configs/inference/metric/91HzMhq7eOS.mp4
    - configs/inference/metric/A1T-Ea-FlQS.mp4
    - configs/inference/metric/A1ubDo0PbQS.mp4
    - configs/inference/metric/A1YNmKj0sCS.mp4
  pose_image_paths:
  - configs/inference/ref_images/anyone-3-partial.png
  - configs/inference/ref_images/91c+SL7Cg7S-98.png
  - configs/inference/ref_images/91c+SL7Cg7S-150.png
  - configs/inference/ref_images/91c+SL7Cg7S-220.png
  ref_image_paths:
  - configs/inference/ref_images/anyone-1.png
  - configs/inference/ref_images/anyone-2.png
  - configs/inference/ref_images/anyone-3.png
  - configs/inference/ref_images/anyone-11.png
weight_dtype: fp16