procslib 0.4.1

Multi-purpose processing library for downstream use

procslib

Multi-purpose processing library for downstream use. Project generated with copier-uv

Installation

pip install procslib

With uv:

uv tool install procslib

Usage

procslib allows inference with many models using a single interface, by sumbitting different keys, similar to Huggingface AutoModels:

from procslib import get_model_keys, get_model

# get available keys
print(get_model_keys())

# create a model
model = get_model("twitter_logfav")

# do inference
iamge_paths_list = ["path/to/image1.jpg", "path/to/image2.jpg"]
res_df = model.infer_many(image_paths_list)

Supported Models:

the models can be retrieved by calling get_model(key) where key is one of the following:

• note that Q-Align models requires transformers==4.36.1, which is incompatible with siglip_aesthetic.

key	field	description	backbone
twitter_logfav	anime	log(predicted twitter favorite count)	convnext v2 base
weakm_v2	anime	previous version of numerical aesthetics score	convnext v2 base
siglip_aesthetic	general	an Clip Aesthetics alternative that uses siglip backbone and with better performance on anime discus0434/aesthetic-predictor-v2-5	siglip (vit) + mlp
pixiv_compound_score	anime	numerical aesthetics score based on pixiv bookmarks and other metrics	convnext v2 tiny
cv2_metrics	general	many useful image related metrics, such as noise, exposure, edge count, etc	(Not a model)
complexity_ic9600	general	a model that analyzes the "complexity" of images tinglyfeng/IC9600	ICNet (resnet18)
rtmpose	general	analyzes the presence of body parts of images mmpose/projects/rtmpose at main	RTMDet
depth	general	using MiDaS 3.0 to analyze the "depthness" of images and returns a numerical metric Intel/dpt-hybrid-midas · Hugging Face	MiDaS
q_align_quality	general	image quality assessment using Q-Align model (rough/distorted images = lower score; refined images = higher score) Q-Future/Q-Align	VLM
q_align_aesthetics	general	image aesthetics assessment using Q-Align model (warn: has a western, or "midjourney-like" taste for higher qualities) Q-Future/Q-Align	VLM
laion_watermark	general	a very fast watermark detection model that detects if there's text on the image. (works 80% of the time but could be inaccurate) LAION-AI/LAION-5B-WatermarkDetection	EfficientNet B3
clip_aesthetic	general	(WIP) caches clip embeddings, calculates similarities with a given list of prompts, then outputs aesthetics scores by supplying a list of MLP models. Very fast when embeddings are cached. troph-team/pixai-aesthetic	clip (vit) + mlp

Development

for how to navigate the project repo (generate changelogs, release versions, etc) see the project template documentation:

make setup  # only once

<write code>
make format  # to auto-format the code

<write tests>
make test  # to run the test suite

make check  # to check if everything is OK

<commit your changes>

make changelog  # to update the changelog
<edit changelog if needed>

make release version=x.y.z  # to release a new version (find the exact version number to use from CHANGELOG.md)

Documentation

To view (and live edit) the documentations, run:

make docs host=0.0.0.0