AI Image Super-Resolution
Upscale a small image 2× or 4× — sharper than a regular resize. ESRGAN runs in your browser via TensorFlow.js; a high-quality Lanczos fallback ships too. Tuned for retro game sprites and old photos. No upload.
How it works · which algorithm to pick
Lanczos (default) — a classic high-quality sinc-based resampler. Sharp results on photographs and most images. No model download, instant on small inputs. Recommended starting point.
ESRGAN AI — a real super-resolution neural network via UpscalerJS + TensorFlow.js. Best for photographs at 2×. First click triggers a one-time ~15 MB model + tfjs download — cached for subsequent runs. Slow on mobile.
Nearest-neighbor — no interpolation, just block-replicate pixels. Perfect for pixel art / retro sprites where you want chunky edges preserved exactly.
Bicubic — smoother than Lanczos, softer edges. Good if Lanczos looks too crisp.
Post-sharpen is an unsharp mask applied after the upscale — useful with ESRGAN at 2× if the result feels slightly soft.