How to Make Pixel Art with Gemini — and Fix the Blurry Grid (Free)
Gemini can generate genuinely good pixel art — but it paints the style, not the grid. Zoom into any output and every “pixel” is a soft block roughly 8–14 screen pixels wide, with feathered edges and hundreds of stray colors. This guide gives you three prompts that produce the best raw output, shows what the model actually draws, and walks the free 30-second fix: recovering the true grid with the Pixelmade Pixel Art Converter. Our test corgi went from a 1024×1024 blur to clean 96×96 pixel art, 16 colors flat.
The whole article in one image: what Gemini gives you, and what the free recovery turns it into.
Copy-paste prompts that work
These three are tested phrasings, not theory — the first one produced every screenshot in this article. Paste them into Gemini (the current image model is the one the community nicknamed “Nano Banana”; see Google’s image generation docs) and swap the subject.
1. Character sprite (the one we used):
Pixel art sprite of a corgi wearing a red bandana, 16 colors max, large visible square pixels, flat plain background, no anti-aliasing, no gradients, crisp 1-pixel outlines
2. 16-bit portrait (SNES era):
16-bit SNES-style pixel art portrait of a tabby cat, three-quarter view, 24 colors max, large visible square pixels, flat colors, no shading, plain solid background, crisp 1-pixel outlines
3. Retro landscape (Game Boy era):
Game Boy style pixel art of a lighthouse on a cliff at dusk, 4-color green palette only, large visible square pixels, chunky dithering, no anti-aliasing, flat plain background
Why each clause matters
- The subject comes first (“sprite of a corgi wearing a red bandana”). Front-load it — everything after is style constraint, and Gemini weights early tokens heavily.
- A color budget (“16 colors max”). Gemini will not actually obey it — our corgi file contained 528 unique shades — but the budget pushes the model toward flat, cel-style fills instead of painterly lighting, which is what you want. The real 16 gets enforced later, by the converter’s color cap.
- “Large visible square pixels” is the single highest-leverage phrase. Drop it and you get a smooth digital illustration with no pixel structure at all.
- “Flat plain background” does two jobs: it stops the background from eating your palette (even so, sky blue took 3,639 of our corgi’s 9,216 cells — about 39%), and it gives grid recovery clean, high-contrast block edges to measure.
- “No anti-aliasing / no gradients / crisp 1-pixel outlines” reduces edge feathering. The model still feathers — that is baked into how it renders — but noticeably less.
- Era styling (“SNES”, “Game Boy”) anchors palette size and proportions to a recognizable hardware generation, which keeps the composition honest — a Game Boy scene in 4 greens reads as pixel art even before you fix the grid.
Prompt #1’s raw output. At full-page zoom it looks like finished pixel art. It is not — yet.
Zoom in: it’s pixel-art style, not pixel art
Open the raw file and the numbers give it away. The PNG is 1024×1024 — about a megapixel — when the artwork it depicts is a 96×96 design. The prompt asked for 16 colors; the file contains 528. Each “pixel” the model drew is really a soft block averaging 10.7 screen pixels across, and no two blocks are exactly the same size.
We call these blocks mixels: the model’s imperfect rendering of one intended pixel. Gemini does not paint on a grid — it renders a picture that depicts a grid, the way it would depict wood grain or fur.
The same file at 8× zoom. Every block edge is a gradient. This is why AI pixel art can’t be used as a sprite, indexed to a palette, or scaled cleanly.
Mixels are why naive fixes fail. Downscaling by a guessed factor resamples across block boundaries and smears neighboring cells together; upscaling just enlarges the blur. (Real pixel art scales losslessly with nearest-neighbor resampling — but only once there is an actual grid to snap to.) For the longer version of why, see our mixels explainer.
Recover the true grid (30 seconds)
The Pixelmade Pixel Art Converter does the fix with deterministic grid detection — it measures where the block boundaries actually fall, no AI guessing, same input → same output. Free, no signup, and images are processed in memory.
- Open the converter and drop in your Gemini PNG (up to 16 MB / 4096 px).
- Wait a second. The tool auto-detects what you gave it: an AI pixel-art-style image triggers grid recovery; an ordinary photo triggers photo mode instead. No settings needed for the first pass.
- Check the result banner — it reports the measured grid, e.g. recovered 96×96 true pixels for our corgi. If the dimensions look right, the hard part is done.
- Optional cleanup: tap a color cap chip — 16 enforces the palette budget the prompt asked for. Our 528 stray shades collapsed to exactly 16 in one click.
- Download — a crisp PNG (scaled up for screens) or a chart BMP at exactly 1 pixel per cell.
After recovery: a true 96×96 bitmap. Every cell is exactly one color, every edge exactly one pixel.
The same face crop as before, after recovery. The feathering is gone because each cell was measured and re-read, not resampled.
One honest footnote on the palette: the 16-color chart still contains three near-identical sky blues (3,639, 1,695 and 1,453 cells) — the color cap keeps close shades apart rather than silently merging your art. If that bothers you, set the cap lower and they collapse.
The final 16-color palette with per-color counts — useful for sprites, and exactly what you need for craft charts.
Troubleshooting: when Gemini fights you
| What you got | Why it happens | The fix |
|---|---|---|
| Smooth illustration, no big pixels | The style anchor got lost | Re-add “large visible square pixels” and “no anti-aliasing” |
| Gradients and soft shading everywhere | The model defaults to painterly light | Add “flat colors, no shading, no gradients” |
| A photo with a vaguely pixelated filter | Subject overpowered the style clauses | Regenerate with both fixes above, then re-run recovery |
| Tiny text or fine detail is garbled | The model drew those regions on a finer grid | Regenerate with the detail drawn larger (add “large visible square pixels”), then re-run recovery |
| The converter says no reliable grid was found | The image is genuinely smooth — there is no latent grid to recover | Use photo mode instead: pixelate at a grid you choose (8–256) |
That last row is deliberate: the converter refuses to fake a recovery on an image that was never pixel-art-styled. A tool that “finds” a grid in everything is making one up.
What to do with your recovered pixel art
- Game sprites and assets. A true 96×96 indexed image drops straight into an engine or editor — and now scales losslessly with nearest-neighbor.
- Avatars and emotes. Crisp at 32 px or 512 px, because scaling a real grid is exact.
- Craft charts. The chart BMP is 1 pixel per cell with a counted palette — which is literally a perler-bead layout or a cross-stitch chart. (Recovery-only tools like Spritecook’s pixel grid detector stop at the image; the counted palette is what makes it a usable chart.)
- A physical kit. If you want the corgi on your wall without doing the crafting math: Pixelmade prints pixel art as made-to-order diamond painting kits — minimum order of one, produced per order, under your brand if you sell. See how it works →
FAQ
Can it convert a regular photo too? Yes. The converter auto-detects what you upload. A photo (no latent grid) switches it to photo mode: choose a grid from 8–256 cells and 2–64 colors, and get gridded pixel art with a counted palette, exported as PNG or chart BMP.
Does Gemini make real pixel art? No — it makes pixel-art style. The output depicts pixels as soft blocks (typically 8–14 screen px each) on a roughly 1-megapixel canvas. Recovery measures that latent grid and rebuilds the image at true size, one color per cell.
What does “recovered 96×96 true pixels” mean? It means grid detection measured 96 columns and 96 rows of intended pixels in your image and rebuilt it at exactly that size — 9,216 cells, each a single flat color. The banner reports whatever your image’s real grid is.
Is the Pixelmade Pixel Art Converter free? Yes — free, no signup, no watermark. Uploads are capped at 16 MB / 4096 px and processed in memory; the conversion is deterministic and takes seconds.
Why did my “16 colors max” prompt come back with 528 colors? Anti-aliasing and soft shading: the model blends every block edge, creating hundreds of in-between shades. That’s normal. Recover the grid first, then apply the 16-color cap — the cap works cleanly once each cell is a single read color.
PNG or chart BMP — which download do I want? PNG for screens: it’s the recovered art scaled up crisply for sharing or game use. Chart BMP for making things: exactly 1 pixel per cell, so it opens in any craft or chart tool at the design’s true dimensions.