Skip to content

Latest commit

 

History

History
70 lines (48 loc) · 1.72 KB

manual.md

File metadata and controls

70 lines (48 loc) · 1.72 KB

Manual

Arguments for executing HandRefiner.py:

  • --input_dir

    input directory containing images to be rectified

  • --input_img

    input image to be rectified

  • --out_dir

    output directory where the rectified images will be saved to

  • --log_json

    file where the mpjpe values will be logged to

  • --strength

    control strength for ControlNet

  • --depth_dir

    directory where the depth maps will be saved to. Leaving it empty will disable this function

  • --mask_dir

    directory where the masks will be saved to. Leaving it empty will disable this function

  • --eval (True/False)

    whether evaluate the mpjpe error in fixed control strength mode, currently only works for batch size of 1.

  • --finetuned (True/False)

    whether use finetuned ControlNet trained on synthetic images as introduced in the paper

  • --weights

    path to the SD + ControlNet weights

  • --num_samples

    batch size

  • --prompt_file

    prompt file for multi-image rectification Format for prompt file:

    {"img": filename, "txt": prompt}

    Example:

    {"img": "img1.jpg", "txt": "a woman making a hand gesture"}
{"img": "img2.jpg", "txt": "a man making a hand gesture"}
{"img": "img3.jpg", "txt": "a man making a thumbs up gesture"}
  • --prompt

    prompt for single image rectification

  • --n_iter

    number of generation iteration for each image to be rectified. In general, for each input image, n_iter x num_samples number of rectified images will be produced

  • --adaptive_control (True/False)

    adaptive control strength as introduced in paper, currently only works for batch size of 1. We tend to use fixed control strength as default.

  • --padding_bbox

    padding controls the size of masks around the hand

  • --seed

    set seed to maintain reproducibility