unfake

unfake

Improve AI-generated pixel art through scale detection, color quantization, and smart downscaling. Features Rust acceleration for critical operations, achieving a 10-20% speedup over the original JavaScript implementation.

Based on the excellent work by:

• Eugeniy Smirnov (jenissimo/unfake.js) - Original JavaScript implementation
• Igor Bezkrovnyi (ibezkrovnyi/image-quantization) - Image quantization algorithms

Examples

Click each image to view the original, processed result, and results of two different naïve fixed-size-nearest-neighbor methods.

Images taken from examples for AI pixel art models like Pixel Art XL, FLUX.1-Dev Pixel LoRA, and FLUX.1-Kontext Pixel LoRA.

Features

• Automatic Scale Detection: Detects the inherent scale of pixel art using both runs-based and edge-aware methods
• Advanced Color Quantization: Wu color quantization algorithm with Rust acceleration
• Smart Downscaling: Multiple methods including dominant color, median, mode, and content-adaptive
• Image Cleanup: Alpha binarization, morphological operations, and jaggy edge removal
• Transparent Background: Automatic background removal using edge-based flood fill
• Grid Snapping: Automatic alignment to pixel grid for clean results
• Flexible API: Both synchronous and asynchronous interfaces
• Fast: Process a 1-megapixel image in as fast as half a second.

Upcoming

• Vectorization

Installation

From PyPI (recommended)

pip install unfake

From Source

# Clone the repository
git clone https://github.com/painebenjamin/unfake.py.git
cd unfake

# Install with pip (includes Rust compilation)
pip install .

# Or for development
pip install -e .

Requirements

• Python 3.8+
• Rust toolchain (for building from source)
• OpenCV Python bindings
• Pillow
• NumPy

Usage

Command Line

# Basic usage with auto-detection
unfake input.png

# Specify output file
unfake input.png -o output.png

# Control color palette size
unfake input.png -c 16                    # Maximum 16 colors
unfake input.png --auto-colors            # Auto-detect optimal color count

# Force specific scale
unfake input.png --scale 4                # Force 4x downscaling

# Choose downscaling method
unfake input.png -m dominant              # Dominant color (default, best for pixel art)
unfake input.png -m median                # Median color
unfake input.png -m content-adaptive      # High quality but slower

# Enable cleanup operations
unfake input.png --cleanup morph,jaggy    # Morphological + jaggy edge cleanup

# Use fixed color palette
unfake input.png --palette palette.txt    # File with hex colors, one per line

# Adjust processing parameters
unfake input.png --alpha-threshold 200    # Higher threshold for alpha binarization
unfake input.png --threshold 0.1          # Dominant color threshold (0.0-1.0)
unfake input.png --no-snap                # Disable grid snapping

# Make background transparent
unfake input.png --transparent-background # Flood-fill background from edges

# Verbose output
unfake input.png -v                       # Show detailed processing info

Python API

import unfake

# Basic processing with defaults
result = unfake.process_image_sync(
    "input.png",
    max_colors=32,                       # Maximum colors in output
    detect_method="auto",                # Scale detection: "auto", "runs", "edge"
    downscale_method="dominant",         # Method: "dominant", "median", "mode", "mean", "content-adaptive"
    cleanup={"morph": False, "jaggy": False},
    snap_grid=True                       # Align to pixel grid
)

# Access results
processed_image = result['image']        # PIL Image
palette = result['palette']              # List of hex colors
manifest = result['manifest']            # Processing metadata

# Auto-detect optimal colors
result = unfake.process_image_sync(
    "input.png",
    max_colors=None,                     # Auto-detect
    auto_color_detect=True
)

# Use fixed palette
fixed_colors = ['#000000', '#ffffff', '#ff0000', '#00ff00', '#0000ff']
result = unfake.process_image_sync(
    "input.png",
    fixed_palette=fixed_colors
)

# Make background transparent
result = unfake.process_image_sync(
    "input.png",
    transparent_background=True          # Flood-fill background from edges
)

Asynchronous API

import asyncio
import unfake

async def process_image_async():
    result = await unfake.process_image(
        "input.png",
        max_colors=16,
        detect_method="runs",
        downscale_method="median",
        cleanup={"morph": True, "jaggy": False},
        snap_grid=True,
        transparent_background=True          # Optional: make background transparent
    )
    result["image"].save("output.png")

asyncio.run(process_image_async())

Processing Options

Scale Detection Methods

• auto (default): Tries runs-based first, falls back to edge-aware
• runs: Analyzes color run lengths (fast, works well for clean pixel art)
• edge: Uses edge detection (slower but handles anti-aliased images)

Downscaling Methods

• dominant (default): Uses most frequent color in each block (best for pixel art)
• median: Median color value (good for photos)
• mode: Most common color (similar to dominant)
• mean: Average color (can create new colors)
• content-adaptive: Advanced algorithm based on Kopf & Lischinski 2011

Cleanup Options

• morph: Morphological operations to remove noise
• jaggy: Removes isolated diagonal pixels

Transparent Background

When enabled (transparent_background=True or --transparent-background), the tool automatically removes the background by flood-filling from the image edges. This feature:

• Uses edge-based flood fill to detect background regions
• Converts background pixels to transparent
• Uses a configurable tolerance for color matching (default: 10)
• Works with both RGB and RGBA images
• Includes Rust acceleration for better performance

Performance

Example processing times for a 1024x1024 image on a high-end Intel desktop CPU using defaults:

• Pure Python: ~71 seconds
• With Rust Acceleration: ~700 milliseconds (about 100x speedup!)

Algorithm Details

Scale Detection

The tool uses two methods to detect the inherent scale of pixel art:

• Runs-based: Analyzes horizontal and vertical color runs to find the GCD
• Edge-aware: Uses Sobel edge detection to find regular grid patterns

Color Quantization

Implements the Wu color quantization algorithm (1992) which:

• Builds a 3D color histogram
• Recursively subdivides color space
• Minimizes variance within each partition
• Produces high-quality palettes

Downscaling

The dominant color method:

• Divides image into scale×scale blocks
• Finds most frequent color in each block
• Falls back to mean if no color is dominant
• Preserves original palette colors

Credits

This Python/Rust implementation is based on:

• unfake.js by Eugeniy Smirnov - The original JavaScript implementation that inspired this project
• image-quantization by Igor Bezkrovnyi - TypeScript implementation of various color quantization algorithms

Additional references:

• Wu, Xiaolin. "Efficient Statistical Computations for Optimal Color Quantization" (1992)
• Kopf, Johannes and Dani Lischinski. "Depixelizing Pixel Art" (2011)

License

MIT License