default.yaml

sd_version: "1.5" # In ['2.1', '2.0', '1.5', 'depth']
model_key: null # Specify model key. If set, ignore sd_version.
input_path: "path/to/video" # Accept .mp4, .gif file or a folder with png/jpg sequence
work_dir: "workdir"

height: 512
width: 512

inversion:
  save_path: "${work_dir}/latents"
  prompt: "xxxx"
  n_frames: null # null for inverting all frames
  steps: 50 # Inversion steps
  save_intermediate: False # Save intermediate latents. Required when using PnP.
  save_steps: 50
  use_blip: False # Use prompt created by BLIP.
  recon: False # Reconstruct the input video from inverted latents.
  control: "none" # Apply ControlNet in inversion. Choices ['tile', 'softedge', 'depth', 'canny', 'none']
  control_scale: 1.0
  batch_size: 8
  force: False # Force run inversion even inverted latents have been found
  # float_precision: "fp32" # "fp16" or "fp32"

generation:
  control: "pnp" # Apply which control in generation. 
  # Choices:
  # 'pnp' (Plug-and-Play), args:
  pnp_attn_t: 0.5
  pnp_f_t: 0.8

  # ['tile', 'softedge', 'depth', 'canny'] (Controlnet), args:
  control_scale: 1.0
  
  # 'none' (No control when using sd2-depth model)

  # Sample args:
  guidance_scale: 7.5 # CFG scale
  n_timesteps: 50
  negative_prompt: "ugly, blurry, low res"
  prompt: null
    # style: "xxx"
    # object: "xxx"
    # background: "xxx"

  latents_path: "${work_dir}/latents"
  output_path: "${work_dir}"

  chunk_size: 4 # Number of frames in a video chunk
  chunk_ord: "mix-4" # Process video chunks in which order. From ['seq', 'rand', 'mix-#'].

  # VidToMe args. See details in "src/vidtome/patch.py, apply_patch()"
  local_merge_ratio: 0.9
  merge_global: True
  global_merge_ratio: 0.8
  global_rand: 0.5
  align_batch: True

  frame_range: [0, 32, 1] # start, end, interval. == [0, 32] == [32]
  frame_ids: null # Specify frame indexes to edit. It will override frame_range.
  save_frame: True

  # LoRA configs
  use_lora: False
  # If use LoRA, add parameters below such as:
  # lora:
  #   pretrained_model_name_or_path_or_dict: null
  #   lora_weight_name: null
  #   lora_adapter: null
  #   lora_weight: 1.0



seed: 123
device: "cuda"
# base_config: "configs/default.yaml" # Set a base config file here
float_precision: "fp16" # "fp16" or "fp32"
enable_xformers_memory_efficient_attention: True