Image Upscaling

AI upscaling enlarges images while reconstructing missing details using neural networks. Unlike regular enlargement, which only interpolates between existing pixels and leads to blurriness, the AI model recognizes patterns like edges, textures, and structures, generating appropriate new pixels.

Formidex uses the ESRGAN-Slim model, running directly in your browser via TensorFlow.js. It offers a good balance between quality and speed at a model size of only about 5 MB.

The ESRGAN model was trained on millions of image pairs: low-resolution and high-resolution versions of the same image. This taught it which fine details belong to which coarse structures. During application, it analyzes your image block by block and adds appropriate details.

For large images, Formidex uses tiling: the image is divided into overlapping tiles, each processed individually, then seamlessly reassembled. This allows processing high-resolution images without overwhelming memory.

Start with 2x enlargement. It delivers the most reliable results with minimal artifacts. The 4x option is for cases where you need significantly more resolution but may produce visible patterns on some images.

Images with clear structures benefit most: architecture, landscapes, and portraits are upscaled particularly well. Heavily compressed images or very small originals (under 100 pixels) can't be perfectly enlarged even with AI.

On mobile devices, only 2x is available since 4x processing would require too much memory. For best results, use a desktop browser with sufficient RAM.

Enhance old photos: family photos from early digital cameras often had only 1-2 megapixels. Upscaling can bring them to a size suitable for prints.

Enlarge thumbnails and small images: profile pictures, app icons, or preview images can be upscaled for larger displays without looking pixelated.

Prepare images for print: when an image is fine for screen but too small for print, AI upscaling can add the missing resolution.