KennethJAllen/proper-pixel-art

What it solves

It converts noisy, high-resolution images that look like pixel art (often produced by AI generative models or low-quality web uploads) into clean, true-resolution pixel art assets. It solves the problem where standard downsampling fails because these images often have non-uniform grids and random artifacts.

How it works

The tool uses a multi-step computer vision pipeline to recover the original pixel grid:

1. Preprocessing: Trims edges and handles transparency.
2. Edge Detection: Uses Canny edge detection and morphological closing to find the boundaries of the "pixels."
3. Grid Recovery: Employs a probabilistic Hough transform to find vertical and horizontal lines, clusters them, and snaps them to gradient projection profile peaks to establish a pixel mesh.
4. Grid Validation: Estimates pixel width based on line gaps and validates it using reconstruction error (color variance within cells).
5. Color Processing: Either quantizes the image to a fixed palette or preserves original colors using a dominant color method.
6. Finalization: Assigns a single representative color to each cell in the mesh to create a true-resolution image.

For videos and GIFs, the tool samples a few frames to determine a single shared grid and palette, ensuring the animation remains consistent without flickering.

Who it’s for

• Game developers and artists who need to clean up AI-generated pixel art.
• Users wanting to convert low-quality web screenshots of pixel art into usable assets.
• Creators working with pixel-art style animations (GIFs/MP4s).

Highlights

• Animation Support: Processes videos and GIFs with frame-to-frame consistency.
• Automated Grid Detection: Automatically detects the pixel width and mesh of the source image.
• Flexible Color Control: Offers both dominant color preservation and fixed-palette quantization.
• Multiple Interfaces: Available as a CLI, Python API, and a web interface via Hugging Face Spaces.