batch-image-quality-rating

Simple python script to rate image quality and aesthetics using one-align

Requirements

Hardware

The unquantized version of this model takes about 16GB of VRAM, luckily, JIT quantization doesn't seem to cause much quality loss.

• 3090 - 4090: Can run full unquantized
• 3080+ : Can run in 8-bit
• 2070+ : Can run in 4-bit

Torch version

This torch version worked for me with an NVIDIA card. I had issues with the latest version.

pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 xformers --index-url https://download.pytorch.org/whl/cu121 --upgrade

Transformers version

Transformers version 4.36.1 is required.

I recommend setting up a venv

Dependencies

• torch
• transformers==4.36.1
• pil

Usage

python inference.py [--db path] [--image path] [--dir path] [--r] [--q4 | --q8]

Arguments

• --db {path} : Path to an SQLite database file | Optional - To store results
• --image {path} : Path to a single image
• --dir {path} : Path to a folder of images
• --r : Recursively search for images in the directory provided | Optional
• --q4 : Load model with 4-bit quantization (~4GB + PyTorch) | Optional
• --q8 : Load model with 8-bit quantization (~8GB + PyTorch) | Optional

Behavior

Prints ratings to console as they are generated

Stores ratings in SQLite database if path provided

• Will create new database if one does not exist in that location

Credits

Powered by https://github.com/Q-Future/Q-Align