config.yaml

# pretrained_model_path: models/model_scope_diffusers
pretrained_model_path: cerspense/zeroscope_v2_576w
output_dir: results
prior_preservation: true
prior_lambda: .5
trainable_modules:
- cross
- temporal
- temp_convs
max_train_steps: 3000
learning_rate: 5.0e-06
checkpointing_steps:  [1000, 1500, 2000, 2500, 3000]
validation_steps: [1000, 1500, 2000, 2500, 3000]
initializer_token: pll
placeholder_token: sks
subject: best
train_data:
  width: 384
  height: 384
  use_bucketing: true
  sample_start_idx: 1
  fps: 24
  frame_step: 1
  n_sample_frames: 16
  fallback_prompt:
    -
  path: data/jester/Drumming_Fingers/resized_videos10
  train_weights: all
  add_token: false
  class_path: data/jester/prior/resized_videos/
  class_prompt:
  sampling: high-level
validation_data:
  prompt: data/prompts/jester.txt
  sample_preview: true
  num_frames: 16
  width: 384
  height: 384
  num_inference_steps: 25
  guidance_scale: 9
dataset_types:
- folder
extra_unet_params:
  learning_rate: 1.0e-05
  adam_weight_decay: 0.0001
train_batch_size: 1
scale_lr: false
lr_scheduler: constant
lr_warmup_steps: 0
adam_beta1: 0.9
adam_beta2: 0.999
adam_weight_decay: 0.01
adam_epsilon: 1.0e-08
max_grad_norm: 1.0
gradient_accumulation_steps: 1
resume_from_checkpoint: null
mixed_precision: fp16
use_8bit_adam: false
enable_xformers_memory_efficient_attention: false
enable_torch_2_attn: true
cached_latent_dir: null
kwargs:
cache_latents: false
gradient_checkpointing: true
offset_noise_strength: 0.1
train_text_encoder: false
use_offset_noise: false