Why your AI pixel art looks blurry.

Real pixel art is built one cell at a time. Each cell holds a single colour, every edge is a hard step between two cells, and the whole image lives on a tidy palette of maybe four to thirty-two colours. The grid is explicit, the edges are intentional, and the file size is tiny because the information is.

The image a generator hands you usually looks like pixel art at thumbnail size, then falls apart the moment you scale it, tile it, or open it in an editor. The cells have soft gradients between them. The edges that should be hard are anti-aliased into ramps. The palette balloons to hundreds or thousands of colours. It is a 1024 x 1024 raster wearing a pixel-art costume.

You cannot polish that image into a real source by scaling, sharpening, or quantising the big version. The information you want — one colour per cell, on a known grid — has to be reconstructed by sampling the implied grid underneath.

Before you can fix it, you have to see it. The same three symptoms show up on almost every generated image trying to look like pixel art, and once you know them you cannot un-see them.

The generator was prompted to make pixel art, so it almost always picks an implied cell size when it composes the image. Your job is to guess it. Common values are 16, 24, 32, 48, and 64. Start with 32 for character art, 16 for items and icons, 64 for scenes.

To verify the guess, overlay a grid of that size on the original and look at the cell boundaries. If most of them land where the colour blocks visually change, you have the right size. If the grid slices through what should be a single cell, try the next size up or down. The right grid is the one where almost every cell holds one dominant colour.

Once the grid is known, downscale the large image to that exact grid using nearest-neighbor sampling. Not bilinear, not Lanczos, not any smoothing filter. Nearest-neighbor reads a single pixel from each cell and keeps it; smoothing filters average across cells and reintroduce the soft edges you are trying to delete.

The output of that step is a tiny image — often 16 x 16 or 32 x 32 — that already feels more like pixel art than the big source did. Optionally, snap each pixel to a small palette of N colours to clean up any remaining drift. The small image, not the big one, is the new source of truth. Anything you need later — a 256-pixel preview, a 512-pixel store asset, a tiled background — comes from scaling this small source by whole numbers.

Reconstruct, don't rescue

Discord avatars and forum profile images are the easy case. A 32 x 32 reconstructed source scaled by nearest-neighbor to 128 or 256 looks crisp at every size the platform renders it at, including the tiny corner of a notification badge.

Game sprites are the next case. Whether you are dropping the reconstructed asset into Godot, Unity, or a custom engine, the small source is what the engine should hold. Set the texture filter to nearest, scale by camera or transform at integer factors, and the sprite stays sharp across any device pixel ratio. Watch-face motifs follow the same rule, with the added constraint that the visible safe area should hold the important pixels.

Reconstruct, don't rescue

The Pixelpond Engine's Upload view is built for exactly this step. Drop the generated image, pick the canvas size that matches the implied grid, and the engine reconstructs a clean source onto an explicit grid you can keep editing. The file never leaves your machine — see converting an image to pixel art for the feature in full.

The same recipe works without the engine — see the SplitPanels below for Aseprite, ImageMagick, and shader options — but the engine collapses the three manual steps into one drop-and-pick, then lets you keep working on the result instead of round-tripping through another tool.

"Reconstruct, don't rescue" is the studio's whole way of working in four words: let the machine hand you the fast, rough version, then rebuild it clean instead of polishing the mess. Move fast to get the source, slow down to make it yours.

Once you have a clean 32 x 32 source, the question becomes how to display it at every size the destination needs — the avatar, the in-game sprite, the watch-face preview, the icon — without smoothing creeping back in. Scaling pixel art without blur covers the second half of the recipe: integer multiples, renderer settings, and the four places blur tends to sneak in.

And if the reconstruction left stray specks or a faint halo around the edges, cleaning up a messy AI sprite clears them in about a minute before you scale anything.