# Image Upscaler

> Enlarge images 2×, 3× or 4× with classic Lanczos/Mitchell filters and sharpening (no AI). Enlarges a small image by 2, 3 or 4 times (up to 16 megapixels) with high-quality classic resampling filters, optional noise reduction before and unsharp-mask sharpening after. A side-by-side crop compares plain pixel enlargement with the result. This is classic interpolation, not an AI upscaler: it cannot invent detail that is not in the original.

Part of usefulweb.site (https://usefulweb.site/). Results are deterministic calculations from the inputs and the listed assumptions; the API returns exactly what the web page shows.

## Use it
- Human interface: https://imageupscaler.usefulweb.site/ (interactive tool at https://imageupscaler.usefulweb.site/tool)
- JSON API: POST https://imageupscaler.usefulweb.site/api/v1/calculate with a JSON object of inputs (schema: https://imageupscaler.usefulweb.site/api/v1/tool)
- OpenAPI: https://imageupscaler.usefulweb.site/openapi.json
- MCP (Streamable HTTP): https://imageupscaler.usefulweb.site/mcp - tool `image_upscaler`

## Inputs
- file (file, required): Image - JPG, PNG, WebP, AVIF, GIF, TIFF or HEIC, max 50 MB. The enlarged result may be at most 16 megapixels.
- factor (select): Scale factor
- filter (select): Resampling filter
- sharpen (select): Sharpening after enlarging
- denoise (select): Noise reduction before enlarging
- format (select): Output format
- quality (integer): Quality - For JPG and WebP.

## Method
- Classic (non-AI) upscaling: every new pixel is a weighted average of nearby original pixels. Lanczos, Mitchell and Catmull-Rom are separable windowed filters; "EWA Lanczos Sharp" uses a cylindrical (elliptical weighted average) filter that produces fewer staircase artefacts on diagonal edges.
- Optional noise reduction runs before enlarging (so noise is not magnified): despeckle, enhance or a 3×3 median. Optional unsharp-mask sharpening runs after enlarging to counter the softness that interpolation causes.
- The image is turned upright from its EXIF orientation, resized in linear light (to avoid darkened edges), converted back to sRGB and encoded. The output is limited to 16 megapixels.
- What AI upscalers do differently: models such as Real-ESRGAN or SwinIR are neural networks trained on millions of image pairs. They hallucinate plausible texture (hair, foliage, text strokes) that is not in the original. That looks sharper, but it can invent wrong details, alter faces and change text. Classic filters never add information; they only interpolate it, so results look softer but stay faithful.

## Limits
- Enlarging cannot recover detail that was never captured. For prints, prefer the original high-resolution file if you have it.