Skip to content

Image Upscaler

Enlarges 2× or 4× with a directional interpolator that keeps diagonal edges sharp, with no model to download.

This is an interpolator, not a generative model. Nothing is downloaded, nothing is uploaded, and it will not invent a face that was four pixels wide — what it does is enlarge without staircasing the diagonals, which is the part ordinary bicubic gets wrong.

⚡ Local, with an optional server assistupscalesuper-resolutionaienlarge

Drop an image to enlarge

No model is downloaded and nothing is uploaded — see the note below for what actually runs.

How to use it

  1. Say what kind of picture it is

    Illustration and screenshot go through DCCI, which continues a 45° edge instead of stair-stepping it. Photograph does not, because DCCI has no measured advantage on grain and claiming otherwise would be the easy thing to do.

  2. Pick the scale

    DCCI works in doublings, so 2× and 4× run it alone and 1.5× or 3× run it and then trim with the resampling filter. The panel lists the exact chain before you start.

  3. Check the output megapixels

    A 4× enlargement is sixteen times the pixels. Above about 40 megapixels the page warns you, because that is where it becomes several seconds and a lot of memory.

Common questions

Can it recover a face from a blurry CCTV frame?

No, and neither can anything else honestly. Detail that was not photographed can only be invented, and inventing it needs a trained prior — a model. There is no model here.

Why is the sharpen slider greyed out?

Because DCCI ran. The unsharp mask exists to rescue edges a separable filter softened; DCCI already preserved them, and sharpening on top only adds halos.

How is DCCI better than plain bicubic?

It chooses its interpolation axis per pixel from the local gradient instead of always working along the rows and columns. Measured against separable cubic in this repo’s own suite: better on diagonal edges and strokes, within a fifth of a decibel on noisy texture and axis-aligned strokes, where it has no advantage.

Feed it from

Next

These accept what this tool produces, so the result can go straight in.