Perfect Pixel

Auto detect and Get perfect Pixel art

Standard scaling often fails to sample AI-generated pixel art due to inconsistent sizes and non-square grids.

This tool automatically detects the optimal grid and delivers perfectly aligned, pixel-perfect results.

Features

• Automatically detect grid size from pixel style images.
• Refines AI generated pixel style image to perfectly aligned grids.
• Easy to integrate into your own workflow.

Try the Web Demo

Installation

Perfect Pixel provides two implementations of the same core algorithm. The Lighweight Backend is designed in case you can't or don't want to use cv2. You can choose the one that best fits your environment:

Feature	OpenCV Backend (perfect_pixel.py)	Lightweight Backend (perfect_pixel_no_cv2.py)
Dependencies	opencv-python, numpy	numpy

You can install Perfect Pixel via pip. It is recommended to install the OpenCV version for better performance.

# Recommended: Fast version with OpenCV support
pip install perfect-pixel[opencv]

# Numpy version: Lightweight (NumPy only)
pip install perfect-pixel

Usage

Step 1: Get pixel style image

First you need extra tools to get a pixel styled image. The recommanded size is between 512 to 1024.

You can use Stable Diffusion with any Pixel Style Lora, or you can use ChatGPT or Gemini to generate one.

For example, I used ChatGPT to transfer an image into pixel style.

prompt: Convert the input image into a TRUE perler bead pixel pattern designed for physical bead crafting, not digital illustration. Canvas size must be exactly 32×32 pixels OR 16×16 pixels, where each pixel represents exactly one perler bead. Use extremely large, chunky pixels with very few active pixels overall. Simplicity is critical. Only keep the main subject. Remove the entire background. For human characters, make sure the face is flat and no shadow. The subject must be centered with clear empty bead rows around all edges to allow easy mounting on a bead board. Add a clean, continuous dark outline around the subject so the silhouette is clearly readable when made with beads. Use a very limited solid color palette (maximum 6–8 colors total). No gradients, no shading, no lighting, no dithering, no texture. No anti-aliasing or smoothing — every pixel must be a perfect square bead aligned to the grid. The output image should be pixel-perfect, each grid only contains one color. Background must be pure solid white.

The image is in pixel style but the grids are distorted. Also we don't know the number of grids.

Step 2: Use Perfect Pixel to refine your image

import cv2
from perfect_pixel import get_perfect_pixel

bgr = cv2.imread("images/avatar.png", cv2.IMREAD_COLOR)
rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)

w, h, out = get_perfect_pixel(rgb)

Also see example.py.

python example.py

The grid size is automatically detected, and the image is refined.

Try integrate it into your own projects!

API Reference

Args	Description
image	RGB Image (H * W * 3)
sample_method	"center", "median" or "majority"
grid_size	Manually set grid size (grid_w, grid_h) to override auto-detection
min_size	Minimum pixel size to consider valid
peak_width	Minimum peak width for peak detection.
refine_intensity	Intensity for grid line refinement. Recommended range is [0, 0.5]. Given original estimated grid line at x, the refinement will search in [x * (1 - refine_intensity), x * (1 + refine_intensity)].
fix_square	Whether to enforce output to be square when detected image is almost square.
debug	Whether to show debug plots.

Returns	Description
refined_w	Width of the refined image
refined_h	Height of the refined image
scaled_image	Refined Image(W * H * 3)

Algorithm

The whole algorithm mainly contains 3 steps:

1. Detect grid size from FFT magnitude of the original image and generate grids.
2. Detect edges using Sobel and refine the grids by aligning them to edges.
3. Use the grids to sample the original image and to get the scaled image.

Star History

Thanks so much!