When AI draws pixel art but forgets the grid

What it does

AI image generators are prolific but sloppy pixel artists: they output images that look blocky yet sit on distorted, non-square grids with inconsistent cell sizes. Perfect Pixel ingests these faux-pixel images, recovers the actual underlying lattice dimensions, aligns the grid to real content edges, and resamples everything into honest, uniform squares. You get back a properly quantized image and its exact grid dimensions.

The interesting bit

Instead of asking you to count blocks manually, the tool treats grid detection as a signal-processing problem. It analyzes the image’s FFT magnitude spectrum to estimate the repeating pattern, then uses Sobel edge detection to nudge those grid lines into place. It is essentially a computer-vision spellchecker for AI pixel art.

Key highlights

• Auto-detects grid dimensions via frequency-domain analysis rather than metadata or guesswork.
• Refines estimated grid lines against Sobel edges to correct the subtle warping common in generated images.
• Offers both an OpenCV backend and a pure-NumPy backend for environments where OpenCV is unwelcome.
• Supports multiple sampling strategies—center, median, and majority—for reconstructing the final color of each cell.
• Provides a ComfyUI node for dropping directly into Stable Diffusion workflows.

Caveats

• The tool expects an input that is already pixel-style and roughly 512–1024 pixels; it will not turn a smooth photograph into pixel art.
• The README does not characterize failure modes for heavily anti-aliased or non-grid-like inputs, so robustness on messy outliers is unclear.

Verdict

Grab this if you are feeding ChatGPT or Stable Diffusion output into game sprites, perler-bead patterns, or any workflow that demands a mathematically clean grid. If you are drawing pixel art by hand on a known canvas, you are already done.