Abstract
We present a simple technique for upscaling small pixel-arts. First, an input image is upscaled on the basis of the saliency, i.e., the local majority of each color. The corner of blocks is rounded and jaggy noise on color boundaries is smoothed. Using our method preserves every color in the original image, i.e., smoothes images only geometrically. The upscaling procedure is reversible, i.e., the original pixel-art is easily obtained from the upscaled output. We show with experiments that our method does not produce the unnatural patterns that appear when using the EPX method and the smoothness of edges is also superior.
Introduction
Upscaling pixel art is a classic challenge in computer graphics. Simple nearest-neighbor interpolation leaves images looking "blocky" or "jaggy." While advanced vectorization methods (like the well-known Kopf-Lischinski algorithm) produce smooth results, they often interpolate colors, creating gradients and thousands of new shades.
This creates two primary issues: irreversibility and data overhead. Once new colors are added and shapes are smoothed, it is difficult to return to the original low-resolution "source of truth." Furthermore, storing high-resolution smoothed versions or complex vector files is less efficient than storing a tiny pixel art file and upscaling it on the fly. This research introduces a geometric smoothing technique that eliminates "jaggies" without introducing new colors or losing the ability to perfectly reconstruct the original image.
Methods
The core of the approach is a three-stage geometric transformation that expands each input pixel into a $7 \times 7$ block of pixels.
- Saliency-Based Diagonal Connection: When two pixels of the same color are diagonally adjacent, the algorithm must decide whether to "connect" them. If two different colors compete for a diagonal connection, the tie is broken by color saliency. Saliency is defined as the inverse of a color's frequency within a local window (e.g., $7 \times 7$). Rare colors, such as a single-pixel eye, are considered high-saliency and prioritized so they do not disappear during smoothing.
- Corner Rounding: The algorithm identifies the sharp corners of the expanded blocks and applies a rounding logic to remove the "staircase" effect.
- Edge Smoothing: A final pass further refines the boundaries between color regions to ensure lines look fluid rather than faceted.
Reversibility: Because the center pixel of every $7 \times 7$ upscaled block remains exactly the same color as the original input pixel, the original artwork can be recovered perfectly by simply downsampling the image.
Key Results
The method was evaluated on various classic pixel art sprites, including character sprites, UI elements, and game items. The results were compared qualitatively against established algorithms like EPX (Eric’s Pixel Expansion) and hq4x.
- Visual Quality: Compared to EPX, the proposed method produces much smoother, more natural curves. It avoids the "blurry" look and unnatural artifacts that hq4x sometimes generates in complex areas.
- Detail Preservation: The saliency logic successfully preserves thin lines and small dots that other smoothing algorithms might accidentally merge or erase.
- Color Integrity: Unlike vectorization methods, the output contains only the colors present in the input, maintaining the specific aesthetic intent of the pixel artist.
Conclusion
By focusing entirely on the geometry of boundaries rather than color interpolation, this method maintains the crisp look of the original palette while removing the technical limitations of low-resolution grids. Future work may involve making the $7 \times 7$ window adaptive to different art styles or refining the connectivity logic to better match human perception in complex edge cases.
Summary updated