Photo Restoration makes blurry, low-resolution or flat photos clearer along two routes. “Local fast enhance” does edge-preserving denoising plus automatic levels and white balance plus USM sharpening, finishes in seconds, leaves the resolution alone, and comes in mild, standard and strong. “AI upscaling” uses the lightweight Real-ESRGAN model (realesr-general-x4v3, about 4.9 MB of ONNX) to run single-threaded tiled inference in the browser through onnxruntime-web, enlarging the image 4× and then outputting at the 2× or 4× you chose. Processing happens locally in the browser and photos are never uploaded.

An honest note: the local compute is limited — AI upscaling clearly improves blur and low resolution, but its effect on deep restoration tasks such as heavy scratches, damage and face reconstruction is limited.

1. 01Upload one imagesingle mode; the original dimensions are shown next to the file name, and once there is a result it shows “→ new dimensions” with a toggle between “Original” and “Processed”.
2. 02Local fast enhancechoose mild, standard or strong and click “Enhance now”; the result appears within seconds and is labelled “local fast enhance”.
3. 03AI upscalingchoose a 2× or 4× output factor and click “Start AI upscaling”; the first run shows “loading the inference engine”, then “downloading the upscaling model (about 4.9 MB, first time only)”, then “initialising the model”, and finally “AI upscaling (tile n/N)”, with a “Cancel” available during inference.
4. 04Download“Download PNG” or “Download JPG” at quality 0.92, named <original name>_enhanced.png or .jpg.

Supported input

JPG / PNG / WebP / GIF / BMP / AVIF / HEIC, one at a time; output is PNG or JPG.

Upscaling input cap

an image with a longest side above 800 px is scaled to 800 before the 4× enlargement (single-threaded WASM inference grows in memory and time with the square of the area), so a large image ends up at most about 3200 px on the longest side after upscaling.

Time

inference runs in 160 px tiles with a 16 px overlap, taking tens of seconds to a few minutes per image depending on the CPU, and can be cancelled at any point.

Transparent PNG

after AI upscaling, transparent areas are filled in as opaque.

Model source

the first run downloads from HuggingFace mirrors; if both mirrors fail the page reports a model download failure and suggests the local fast enhance instead; the onnxruntime wasm runtime is loaded from jsDelivr.

What it does not do

it does not repair scratches, damage or faded patches, does not do face-specific restoration and does not process in batches; for colour work without enlargement, use “Auto Enhance”.

The photo is processed and inferred entirely inside the browser and is never uploaded to any server; the first AI upscaling run downloads the model and wasm runtime files (about 4.9 MB of model) from third-party CDNs (HuggingFace mirrors and jsDelivr), and those requests contain none of your images and are reused from the browser cache afterwards.