How the ai image upscaler works
AyeCalc uses the Swin2SR lightweight 2× super-resolution model. The browser downloads its model weights and runtime when you start enhancement. A worker processes overlapping image tiles, crops the context around each tile, and stitches the enhanced pixels into one image. The original alpha channel is resized separately so transparent PNG and WebP inputs can retain transparency. The output is encoded as PNG.
Output width = input width × 2; output height = input height × 2Doubling both dimensions produces four times as many pixels. More pixels do not guarantee more accurate detail; the model estimates visual patterns.
Enlarge an 800 × 600 photo to 1600 × 1200
An 800 × 600 source contains 480,000 pixels. At 2×, the export measures 1600 × 1200 and contains 1,920,000 pixels. At a chosen print density of 300 pixels per inch, those dimensions correspond to about 5.33 × 4 inches. Inspect faces, lettering, and edges before using the result; the model cannot verify or recover information that the original did not capture.
Using the processed image
Keep the exported dimensions, format, transparency, and compression level appropriate for where the image will be used.
Original: 800 × 600 px
Enhanced: 1600 × 1200 px
Download: PNG with transparency
Metadata: removed from the new fileUsing the result accurately
Start with the best original you have. This tool is intended for small photos and supports still JPEG, PNG, and WebP files up to 15 MB. The input limit is 1 megapixel on typical desktops and 0.5 megapixels on smaller or lower-memory devices, with each dimension between 8 and 4,096 pixels. Use the image resizer first for larger files. Compare the result at full size, then compress or convert it if the PNG is too large for your website.
AI enhancement may invent textures, alter faces or text, introduce artifacts, or leave blur unchanged. It is not a way to recover reliable evidence or read illegible documents. Large photos can take several minutes. This version provides 2× upscaling only, requires WebAssembly, Web Workers, OffscreenCanvas, and a first-use network download, and does not support animation or HEIC input. Export removes EXIF, other embedded metadata, and Content Credentials; keep the original file.
References: Swin2SR model and Transformers.js usage · Swin2SR research paper · Original model and Apache 2.0 license.