shinka 0.1.2

A high-quality image upscaling package using a genetically-evolved algorithm.

SHINKA Upscaler

This repository provides a Python API for high-quality image upscaling using a specialized, genetically-evolved algorithm. The upscaler is designed for both grayscale and color images, and can be used as a drop-in replacement for traditional interpolation-based upscaling methods.

Features

• Upscales images by a specified factor (e.g., 2x, 4x)
• Supports input as file path, PIL Image, or NumPy array
• Outputs as PIL Image or NumPy array
• Optionally saves the upscaled image to disk

Benchmark Results

SHINKA demonstrates state-of-the-art perceptual quality on real-world images. In a benchmark on 1,000 images from the takara-ai/image_captions dataset, SHINKA achieved the highest SSIM (Structural Similarity Index) among common upscaling methods:

Method	SSIM	PSNR (dB)
SHINKA	0.9459	35.32
Lanczos	0.9432	36.07
Bicubic	0.9376	35.54
Bilinear	0.9206	34.16
Nearest	0.9145	32.01

How to Use

Installation

Install from PyPI:

pip install shinka

Basic Usage

from shinka import upscale

# Upscale an image from file and save the result
result = upscale("input.jpg", scale=2, save_path="upscaled.png")

# Upscale a NumPy array and get a PIL Image
import numpy as np
from PIL import Image
arr = np.array(Image.open("input.jpg"))
result = upscale(arr, scale=2, output_type="PIL")

# Upscale and get a NumPy array
result = upscale("input.jpg", scale=2, output_type="np")

Command-Line Example

A minimal test script is provided:

# test_shinka.py
from shinka import upscale

upscale("DSC00183.jpeg", scale=4, save_path="test_evolved_upscaled.png")
print("Saved test_evolved_upscaled.png")

How the Algorithm Was Created

The upscaling algorithm was developed using a genetic optimization process. A population of image processing pipelines was evolved over multiple generations to maximize perceptual quality (measured by SSIM and sharpness) on a diverse set of images. The best-performing pipeline was selected and implemented in shinka for efficient, non-ML upscaling.

License

This project is licensed under the MIT License. See the LICENSE file for details.