Pixel Art Converter turns AI illustrations into game-ready sprites

If you already have a character, item or scene from an image model but cannot drop it straight into a game, Pixel Art Converter handles that next step. It does not generate an illustration from nothing. It takes an existing PNG and helps normalize its size, colors, transparency and animation frames into a more usable game asset. The walkthrough below follows the desktop app, first for one character portrait and then for an animated sheet. Builds are available for Windows, macOS and Linux; check the latest GitHub release for the current files.

1. Download and open the app

Windows: Download PixelArtConverter-windows-x64.zip from the latest release. The README currently lists a package of about 363 MB. Right-click the ZIP, choose Extract All, open the extracted folder and run PixelArtConverter.exe. On first launch, SmartScreen may say the publisher is unrecognized because the executable is not signed yet. If you have confirmed that the file came from the project’s official GitHub Releases page, follow the README and choose More info → Run anyway. If you are uncomfortable bypassing that prompt, you can run from source or wait for a signed build.

To check the download, open PowerShell in the folder containing the ZIP and run Get-FileHash .\PixelArtConverter-windows-x64.zip -Algorithm SHA256. Compare the result with the Windows SHA-256 value in the README. A checksum can detect a damaged or mismatched download; it does not replace downloading from the official release page.

macOS: The README currently lists a DMG for Apple Silicon. Open it and drag the app to Applications. Because the app is not yet signed, Gatekeeper may block the first launch; the project documents opening it via right-click → Open or allowing it under System Settings → Privacy & Security. Linux: Download and extract the x64 tar.gz, open a terminal in that folder, run chmod +x PixelArtConverter, then ./PixelArtConverter. If Qt libraries are missing, the README lists the packages for Debian/Ubuntu.

2. Prepare a PNG that will convert cleanly

For a first try, use a PNG with a clear subject and a simple background rather than a busy scene. A pixel canvas has a strict budget: when a 512×512 illustration is reduced to 64×64, every object shares those limited cells, so tiny text, hair strands and fine textures are likely to merge. A flat background is also easier to remove than a gradient or a cluttered scene. Keep the original image so you can compare the silhouette and colors after conversion.

Drop the PNG into the app or use its file picker. For a single character or item, leave Sprite Sheet off; that option is for an image already arranged in animation rows and columns. First confirm that the preview has not cropped the subject, then set the output dimensions.

3. Choose sprite dimensions and a palette

Use Preset / W × H for the final sprite dimensions, not the dimensions of the source image. A 64×64 canvas is a useful first setting for a character with a few recognizable accessories. Try 16×16 or 32×32 for a very simple icon, and a larger canvas when the source has details that need to survive. The README describes output sizes from roughly 16×16 to 256×256. Turn on Lock to preserve the width-to-height ratio while editing. For a square portrait, enter the same value for width and height.

Next choose Palette. Auto K-means derives a set of colors from the image; in this mode, set Colors to the desired count (the README lists 4–64). For a consistent art direction across several characters, choose a fixed palette such as PICO-8, Sweetie 16, Endesga 32 or Game Boy. A fixed palette determines its own color set. Palette choice changes the overall feel of shadows, skin and materials, not just the number of colors. Export one version with a palette, compare another, and avoid reducing colors too aggressively on the first pass.

4. Remove the background and inspect the preview

To export transparency, enable Remove Background. AI images often use a flat backdrop; start with Edge-color (auto), which keys out the background color at the edge. It needs no model download and works offline. With a solid green, blue or other single-color background, make sure the subject’s edges are not a similar color; otherwise hair, leaves or glass-like details may disappear with the backdrop.

For a complex background or colors close to the subject, try one of the rembg models. On Windows and Linux, the first selection downloads about 170 MB from GitHub and caches it in your user directory; all models together are about 700 MB. The app shows progress and lets you cancel. The README says that if a model cannot find a subject, the app keeps the original image and warns you rather than silently deleting it. Still inspect fine edges—ears, hair tips and weapon points—because automatic background removal is not a guarantee of a perfect cutout.

Use Preview Scale to inspect at 1×, 2×, 4× or 8×, and turn on the grid to check cells individually. Zoom changes how large the preview appears; it does not add pixels. If the sprite looks washed out, adjust Bright, Contr or Sat a little at a time. These adjustments happen before color quantization, so large changes can push colors too far. Dithering applies an 8×8 Bayer ordered pattern to gradients; flat icons are often cleaner with it off, while sky or metal gradients are worth comparing both ways. Selective Outline adds a palette-matched, one-pixel outline based on nearby material colors. On a tiny sprite, check that the outline has not closed gaps or swallowed an eye or narrow detail.

5. A complete single-portrait example

1. Choose a PNG with the full character visible and a clear difference between the subject and backdrop.
2. Drop it into the app. Set W × H to 64×64 and turn on Lock for a square portrait.
3. Try Sweetie 16 or Game Boy for a strong retro look. To retain more of the source colors, use Auto K-means and set Colors to 24 or 32.
4. If the image has a flat background, enable Remove Background → Edge-color (auto). Inspect the outline before trying a rembg model.
5. Zoom to 4× and turn on the grid. Check the face, held item and silhouette; adjust contrast, dithering or outline only where it helps.
6. When it looks right, click Save Pixel Art PNG and choose a destination. Keep both the illustration and first export so you can compare or revert.

The converter uses edge-preserving smoothing, perceptual color clustering and majority-vote downscaling to reduce blur and muddy blended colors. It does not decide where the light source should be, how a curve should look, or how many colors a material deserves. If a nose, badge or weapon turns into a blob, try a larger sprite, simplify the source image or clean it up in a pixel editor such as Aseprite or Krita instead of stacking more filters.

6. Make animation frames and a sprite sheet

For separate PNG frames, name them in playback order—for example walk_01.png, walk_02.png and walk_03.png—then drop them together or multi-select them in the file picker. The app sorts filenames and lays the frames out horizontally. Zero-pad the numbers so walk_10 does not sort ahead of walk_2. All frames share one palette, which helps prevent the character’s colors from flickering during playback.

If you already have a grid image, tick Sprite Sheet, enter its actual Cols × Rows, then import the PNG. For a source with four columns and two rows, enter 4×2. A wrong row or column count will cut through the frames. The app splits, processes and recombines the grid. The suggested output name includes the frame count and per-frame dimensions, which helps a game engine slice it consistently.

The current README lists PNG sheet output. GIF export, frame-by-frame playback scrubbing and automatic grid detection remain on the roadmap, so use another animation app if you need a finished GIF or a timeline inside the converter.

7. Use the prompt helper when starting from an AI image

AI Prompt Samples is a prompt builder, not an image-generation service inside the app. Choose single-image or animation-sheet mode, then select ChatGPT/DALL·E 3 or Gemini/Imagen as the target. Set the subject, view and a solid background, then choose an action preset such as walk, run, idle, jump, sit or sword swing. Animation mode also lets you choose a frame count and describe each frame. Click Copy Prompt, paste it into the corresponding ChatGPT or Gemini service, generate the image there, and bring the returned PNG back into the converter.

For animation, keep the Subject description and camera view consistent across frames, and use the same flat background so removal and sheet assembly are easier. The README points out that the cast-spell preset describes the pose only; it does not draw a spell effect, so add a fireball or sword trail separately. The Gemini prompt template emphasizes iconic details, thick outlines, gold trim and complete accessories because the same prompt may produce plainer art there. The external model still creates the image, so consistency depends on its output and may require another generation or manual correction.

Final checks before export

At 1×, make sure the character reads clearly; at 4× or 8×, inspect isolated pixels, transparent edges, palette consistency and frame-to-frame color shifts. Import the PNG into the target game engine once to confirm its dimensions, alpha channel and slicing settings. The repository uses the MIT license, while bundled Lospec palettes remain subject to their creators’ terms. See the project README and latest release for the complete reference.