Pixel art, made with AI
Most AI “pixel art” is a smooth illustration wearing a costume: soft edges, hundreds of colours, a grid that does not line up. This site is about the other approach — prompting, downscaling, and palette work that produces art you can actually drop into a project.
The directory
Everything we publish stacks onto one of these. Start wherever your current project is stuck — the pieces are meant to be read in any order.
Prompting for real pixel art. Why "pixel art style" on its own returns a smooth illustration with fake pixels, and which prompt structures push a model toward genuine blocky output instead.
Resolution and downscaling. Generating large and shrinking down is the standard route, but nearest-neighbour, box, and bicubic all give you very different results at 64x64.
Colour cleanup and re-indexing. AI output usually carries hundreds of near-identical colours. Re-indexing to a fixed palette is what makes the difference between "pixel-ish" and pixel art.
Isometric scenes and tilesets. Tiles have to line up. We look at what AI tools can and cannot do for seamless tiles, isometric blocks, and consistent lighting across a scene.
Character sprites and frames. Character work is where AI struggles most: the same character has to survive across poses, directions, and animation frames without drifting.
Licensing what you generate. Before art goes into a game or a store page, you need to know what the tool’s terms allow. We summarise the questions worth asking of any generator.
How this site is edited
There are no scores here, no leaderboards, and no numbers we cannot show our working for. Just what we tried, what came out, and whether it was worth it.
Every technique on this site gets run through an actual tool first. If a workflow only works on a cherry-picked prompt, we say so rather than presenting it as a recipe.
For some jobs — tight animation frames, tiny UI icons, precise tilesets — hand-drawing is still faster and better. We would rather lose the click than waste your afternoon.
Some outbound links to tools are affiliate links, which may earn us a commission at no cost to you. They never decide what we cover or what we recommend.
The workflow we keep coming back to
Nearly every good result follows the same climb. Skip a step and you can usually tell which one from the finished image.
Decide the finished canvas size first, then choose whether to generate at that size or generate large and reduce. The choice changes everything downstream.
Reduce with a method that keeps hard edges, and check the result at 100% zoom. If the pixels are not square and aligned, nothing after this will fix it.
Quantise down to a deliberate palette so shading reads as bands, not gradients. This is the step most AI pixel art skips, and the one people notice.
Fix stray pixels, tidy the outline, and correct anything the model invented. A few minutes in an editor is usually the difference between usable and not.