dghs-imgutils

imgutils

A convenient and user-friendly anime-style image data processing library that integrates various advanced anime-style image processing models.

Installation

You can simply install it with pip command line from the official PyPI site.

pip install dghs-imgutils

If your operating environment includes a available GPU, you can use the following installation command to achieve higher performance:

pip install dghs-imgutils[gpu]

For more information about installation, you can refer to Installation.

Supported or Developing Features

• Tachie(差分) Detection and Clustering
• Contrastive Character Image Pretraining
• Object Detection
• Edge Detection / Lineart Generation
• Monochrome Image Detection
• Image Rating
• Truncated Image Check
• Image Tagging
• Character Extraction

imgutils also includes many other features besides that. For detailed descriptions and examples, please refer to the official documentation. Here, we won't go into each of them individually.

Tachie(差分) Detection and Clustering

For the dataset, we need to filter the differences between the tachie(差分). As shown in the following picture

We can use lpips_clustering to cluster such situations as shown below

from imgutils.metrics import lpips_clustering

images = [f'lpips/{i}.jpg' for i in range(1, 10)]
print(images)
# ['lpips/1.jpg', 'lpips/2.jpg', 'lpips/3.jpg', 'lpips/4.jpg', 'lpips/5.jpg', 'lpips/6.jpg', 'lpips/7.jpg', 'lpips/8.jpg', 'lpips/9.jpg']
print(lpips_clustering(images))  # -1 means noises, the same as that in sklearn
# [0, 0, 0, 1, 1, -1, -1, -1, -1]

Contrastive Character Image Pretraining

We can use imgutils to extract features from anime character images (containing only a single character), calculate the visual dissimilarity between two characters, and determine whether two images depict the same character. We can also perform clustering operations based on this metric, as shown below

from imgutils.metrics import ccip_difference, ccip_clustering

# same character
print(ccip_difference('ccip/1.jpg', 'ccip/2.jpg'))  # 0.16583099961280823

# different characters
print(ccip_difference('ccip/1.jpg', 'ccip/6.jpg'))  # 0.42947039008140564
print(ccip_difference('ccip/1.jpg', 'ccip/7.jpg'))  # 0.4037521779537201
print(ccip_difference('ccip/2.jpg', 'ccip/6.jpg'))  # 0.4371533691883087
print(ccip_difference('ccip/2.jpg', 'ccip/7.jpg'))  # 0.40748104453086853
print(ccip_difference('ccip/6.jpg', 'ccip/7.jpg'))  # 0.392294704914093

images = [f'ccip/{i}.jpg' for i in range(1, 13)]
print(images)
# ['ccip/1.jpg', 'ccip/2.jpg', 'ccip/3.jpg', 'ccip/4.jpg', 'ccip/5.jpg', 'ccip/6.jpg', 'ccip/7.jpg', 'ccip/8.jpg', 'ccip/9.jpg', 'ccip/10.jpg', 'ccip/11.jpg', 'ccip/12.jpg']
print(ccip_clustering(images, min_samples=2))  # few images, min_sample should not be too large
# [0, 0, 0, 3, 3, 3, 1, 1, 1, 1, 2, 2]

For more usage, please refer to official documentation of CCIP.

Object Detection

Currently, object detection is supported for anime heads and person, as shown below

• Face Detection

• Head Detection

• Person Detection

Based on practical tests, head detection currently has a very stable performance and can be used for automation tasks. However, person detection is still being further iterated and will focus on enhancing detection capabilities for artistic illustrations in the future.

Edge Detection / Lineart Generation

Anime images can be converted to line drawings using the model provided by patrickvonplaten/controlnet_aux, as shown below.

It is worth noting that the lineart model may consume more computational resources, while canny is the fastest but has average effect. Therefore, lineart_anime may be the most balanced choice in most cases.

Monochrome Image Detection

When filtering the crawled images, we need to remove monochrome images. However, monochrome images are often not simply composed of grayscale colors and may still contain colors, as shown by the first two rows of six images in the figure below

We can use is_monochrome to determine whether an image is monochrome, as shown below:

from imgutils.validate import is_monochrome

print(is_monochrome('mono/1.jpg'))  # monochrome images
# True
print(is_monochrome('mono/2.jpg'))
# True
print(is_monochrome('mono/3.jpg'))
# True
print(is_monochrome('mono/4.jpg'))
# True
print(is_monochrome('mono/5.jpg'))
# True
print(is_monochrome('mono/6.jpg'))
# True
print(is_monochrome('colored/7.jpg'))  # colored images
# False
print(is_monochrome('colored/8.jpg'))
# False
print(is_monochrome('colored/9.jpg'))
# False
print(is_monochrome('colored/10.jpg'))
# False
print(is_monochrome('colored/11.jpg'))
# False
print(is_monochrome('colored/12.jpg'))
# False

For more details, please refer to the official documentation .

Image Rating

We can use imgutils to perform a rough and quick determination of the rating (safe, r15, or r18) of anime images, as shown below.

from imgutils.validate import anime_rating

print(anime_rating('rating/safe/1.jpg'))  # ('safe', 0.9999998807907104)
print(anime_rating('rating/safe/2.jpg'))  # ('safe', 0.9924363493919373)
print(anime_rating('rating/safe/3.jpg'))  # ('safe', 0.996969997882843)
print(anime_rating('rating/safe/4.jpg'))  # ('safe', 0.9966891407966614)
print(anime_rating('rating/r15/5.jpg'))  # ('r15', 0.9996721744537354)
print(anime_rating('rating/r15/6.jpg'))  # ('r15', 0.9998563528060913)
print(anime_rating('rating/r15/7.jpg'))  # ('r15', 0.9827859997749329)
print(anime_rating('rating/r15/8.jpg'))  # ('r15', 0.8339558839797974)
print(anime_rating('rating/r18/9.jpg'))  # ('r18', 0.9997004270553589)
print(anime_rating('rating/r18/10.jpg'))  # ('r18', 0.9997699856758118)
print(anime_rating('rating/r18/11.jpg'))  # ('r18', 0.9999485015869141)
print(anime_rating('rating/r18/12.jpg'))  # ('r18', 0.9994290471076965)

Truncated Image Check

The following code can be used to detect incomplete image files (such as images interrupted during the download process):

from imgutils.validate import is_truncated_file

if __name__ == '__main__':
    filename = 'test_jpg.jpg'
    if is_truncated_file(filename):
        print('This image is truncated, you\'d better '
              'remove this shit from your dataset.')
    else:
        print('This image is okay!')

Image Tagging

The imgutils library integrates various anime-style image tagging models, allowing for results similar to the following:

The ratings, features, and characters in the image can be detected, like this:

import os
from imgutils.tagging import get_wd14_tags

rating, features, chars = get_wd14_tags('skadi.jpg')
print(rating)
# {'general': 0.0011444687843322754, 'sensitive': 0.8876402974128723, 'questionable': 0.106781005859375, 'explicit': 0.000277101993560791}
print(features)
# {'1girl': 0.997527003288269, 'solo': 0.9797663688659668, 'long_hair': 0.9905703663825989, 'breasts': 0.9761719703674316,
#  'looking_at_viewer': 0.8981098532676697, 'bangs': 0.8810765743255615, 'large_breasts': 0.9498510360717773,
#  'shirt': 0.8377365469932556, 'red_eyes': 0.945058286190033, 'gloves': 0.9457170367240906, 'navel': 0.969594419002533,
#  'holding': 0.7881088852882385, 'hair_between_eyes': 0.7687551379203796, 'very_long_hair': 0.9301245212554932,
#  'standing': 0.6703325510025024, 'white_hair': 0.5292627811431885, 'short_sleeves': 0.8677047491073608,
#  'grey_hair': 0.5859264731407166, 'thighs': 0.9536856412887573, 'cowboy_shot': 0.8056888580322266,
#  'sweat': 0.8394746780395508, 'outdoors': 0.9473626613616943, 'parted_lips': 0.8986269235610962,
#  'sky': 0.9385137557983398, 'shorts': 0.8408567905426025, 'alternate_costume': 0.4245271384716034,
#  'day': 0.931140661239624, 'black_gloves': 0.8830795884132385, 'midriff': 0.7279844284057617,
#  'artist_name': 0.5333830714225769, 'cloud': 0.64717698097229, 'stomach': 0.9516432285308838,
#  'blue_sky': 0.9655293226242065, 'crop_top': 0.9485014081001282, 'black_shirt': 0.7366660833358765,
#  'short_shorts': 0.7161656618118286, 'ass_visible_through_thighs': 0.5858667492866516,
#  'black_shorts': 0.6186309456825256, 'thigh_gap': 0.41193312406539917, 'no_headwear': 0.467605859041214,
#  'low-tied_long_hair': 0.36282333731651306, 'sportswear': 0.3756745457649231, 'motion_blur': 0.5091936588287354,
#  'baseball_bat': 0.951993465423584, 'baseball': 0.5634750723838806, 'holding_baseball_bat': 0.8232709169387817}
print(chars)
# {'skadi_(arknights)': 0.9869340658187866}

rating, features, chars = get_wd14_tags('hutao.jpg')
print(rating)
# {'general': 0.49491602182388306, 'sensitive': 0.5193622708320618, 'questionable': 0.003406703472137451,
#  'explicit': 0.0007208287715911865}
print(features)
# {'1girl': 0.9798132181167603, 'solo': 0.8046203851699829, 'long_hair': 0.7596215009689331,
#  'looking_at_viewer': 0.7620116472244263, 'blush': 0.46084529161453247, 'smile': 0.48454540967941284,
#  'bangs': 0.5152207016944885, 'skirt': 0.8023070096969604, 'brown_hair': 0.8653596639633179,
#  'hair_ornament': 0.7201820611953735, 'red_eyes': 0.7816740870475769, 'long_sleeves': 0.697688639163971,
#  'twintails': 0.8974947333335876, 'school_uniform': 0.7491052746772766, 'jacket': 0.5015512704849243,
#  'flower': 0.6401398181915283, 'ahoge': 0.43420469760894775, 'pleated_skirt': 0.4528769850730896,
#  'outdoors': 0.5730487704277039, 'tongue': 0.6739872694015503, 'hair_flower': 0.5545973181724548,
#  'tongue_out': 0.6946243047714233, 'bag': 0.5487751364707947, 'symbol-shaped_pupils': 0.7439308166503906,
#  'blazer': 0.4186026453971863, 'backpack': 0.47378358244895935, ':p': 0.4690653085708618, 'ghost': 0.7565015554428101}
print(chars)
# {'hu_tao_(genshin_impact)': 0.9262397289276123, 'boo_tao_(genshin_impact)': 0.942080020904541}

We currently integrate the following tagging models:

• Deepdanbooru model, but not recommended for production use.
• wd14-v2 model, inspired by SmilingWolf/wd-v1-4-tags.

In addition, if you need to convert the dict-formatted data mentioned above into the text format required for image training and tagging, you can also use the tags_to_text function (see the link here) for formatting, as shown below:

from imgutils.tagging import tags_to_text

# a group of tags
tags = {
    'panty_pull': 0.6826801300048828,
    'panties': 0.958938717842102,
    'drinking_glass': 0.9340789318084717,
    'areola_slip': 0.41196826100349426,
    '1girl': 0.9988248348236084
}

print(tags_to_text(tags))
# '1girl, panties, drinking_glass, panty_pull, areola_slip'
print(tags_to_text(tags, use_spaces=True))
# '1girl, panties, drinking glass, panty pull, areola slip'
print(tags_to_text(tags, include_score=True))
# '(1girl:0.999), (panties:0.959), (drinking_glass:0.934), (panty_pull:0.683), (areola_slip:0.412)'

Character Extraction

When we need to extract the character parts from anime images, we can use the segment-rgba-with-isnetis function for extraction and obtain an RGBA format image (with the background part being transparent), just like the example shown below.

from imgutils.segment import segment_rgba_with_isnetis

mask_, image_ = segment_rgba_with_isnetis('hutao.png')
image_.save('hutao_seg.png')

mask_, image_ = segment_rgba_with_isnetis('skadi.jpg')
image_.save('skadi_seg.png')

This model can be found at https://huggingface.co/skytnt/anime-seg .