AI upscaler

4x upscaling that never sees a server

The 7 MB NERVE 4x model was trained on real-world blur, noise and JPEG damage, so it sharpens everyday photos instead of amplifying their artifacts. Tiles are processed with overlap and blended, so detail survives the seams — up to roughly 4 MP per upscale on a laptop GPU, since 4x means 16x the pixels.

Upscale a photoFree · no account · runs on your GPU

Built for real photos

NERVE learned on degraded real-world images, so compression blocks and sensor noise are cleaned up rather than enlarged — though like any generative upscaler it can redraw fine detail and letterforms instead of preserving them exactly.

Tiled, so big images work

The image is processed in overlapping 256px tiles with feathered seams instead of one giant tensor. Laptop GPUs and WASM fallback both cope.

Result is a new layer

The 2x or 4x render lands on its own layer. Compare against the original, denoise before or after, then export up to PNG, WebP, AVIF, or JPEG.

  1. 01

    Open a photo

    Anything from a thumbnail to a 4 MP shot. Upscaling multiplies pixels 16x, so the output has to fit the browser canvas limit — resize larger images first.

  2. 02

    Run the upscale

    The model is a ~7 MB one-time download. Progress streams live per tile.

  3. 03

    Export at full size

    Canvas budgets are enforced before export so you get a clear error instead of a frozen tab.

FAQ

Questions, answered plainly

How is this different from cloud upscalers?
Cloud upscalers meter you per image and keep a copy of your photo. Here the weights download once and every upscale after that is free, private, and offline-capable.
My upscaled text looks smeared. What do I do?
NERVE reconstructs detail rather than preserving it exactly, so tiny letterforms can change shape. If text still smears, the source resolution is likely too low.
How large can the input be?
The editor opens files up to 50 MB and 16384 px per side. Upscaling multiplies pixel count by 16, and the result must stay within the browser canvas limit (about 67 MP), so a 4x upscale accepts up to about 4 MP of input. Larger images are refused with a message; resize them first from the Canvas menu.

Same engine, more jobs