CanDoYa

Free AI Image Upscaler

Runs entirely in your browser - no upload, no sign-up.

Choose a small image to enlarge with local AI.

Share this tool

What does an AI image upscaler do?

An AI image upscaler increases image resolution while predicting detail between the original pixels. This tool runs the Swin2SR super-resolution model on your device, produces an image exactly twice the width and height, and exports PNG or JPEG. Your source and result are never uploaded to CanDoYa.

How to use

  1. 1Choose a small source image. Drop in one JPEG, PNG, or WebP. Each side must be between 16 and 512 pixels, and the total source area cannot exceed 262,144 pixels.
  2. 2Run 2x super-resolution. Click Upscale image 2x. The worker downloads Swin2SR on the first run, then uses WebGPU when available or the WebAssembly CPU fallback.
  3. 3Compare and export. Drag the before-and-after control across the result. Choose lossless PNG or a smaller 92-quality JPEG, then download the exactly 2x image.

Who it's for

This is learned super-resolution, not ordinary canvas interpolation. The Apache-2.0 Swin2SR model uses a Swin Transformer V2 architecture trained for 2x image super-resolution. It reconstructs edges and textures from local image patterns. That can look sharper than stretching pixels, but it cannot recover factual detail that was never present.

The AI runs in a dedicated browser worker through Transformers.js 4.2.0. It prefers roughly 31 MB of FP16 ONNX weights on WebGPU and falls back to roughly 21 MB of q8 weights with WebAssembly on the CPU. Runtime files download separately and browsers may cache them. The tool caps each source side at 512 pixels, source area at 262,144 pixels, output area at 1,048,576 pixels, and estimated source/output tensor buffers at 20 MB.

FAQ

Is my image uploaded to a server?

No. Your selected image is decoded, passed to a worker, upscaled, compared, and encoded inside your browser. The worker fetches the public Transformers.js runtime and Swin2SR model from external hosts, but your image is not sent with those requests. CanDoYa does not receive or store it.

Is this AI image upscaler free?

Yes. It has no account, credits, watermark, or download fee. Your own device performs the inference. The first use downloads model and runtime files, so it still needs an internet connection and can use several dozen megabytes of data.

What image limits apply?

The tool accepts one JPEG, PNG, or WebP up to 12 MB. Width and height must each be 16 to 512 pixels, source area is capped at 262,144 pixels, and output area at 1,048,576 pixels. These strict limits reduce crashes and long stalls on memory-constrained browsers.

Does this really use AI or just resize the canvas?

It uses the 12.1-million-parameter Swin2SR image-to-image model through Transformers.js. The browser canvas only previews and encodes the neural network's RGB output. It never substitutes browser scaling for model inference, so a failed model produces an error rather than a silently interpolated result.

Does the image upscaler work without WebGPU?

Yes, if the browser supports WebAssembly and module workers. The worker first checks for a usable WebGPU adapter and loads FP16 weights when one is available. If loading or inference fails, it retries with a q8 model on the WebAssembly CPU backend, which is compatible but slower.

Why is the image limited to 2x?

The selected Swin2SR checkpoint is trained specifically for classical 2x super-resolution. Repeating it to claim 4x would compound invented detail and artifacts, while ordinary 4x interpolation would not be genuine AI upscaling. A fixed 2x mode keeps the model behavior and output dimensions honest.

Will AI upscaling restore missing details perfectly?

No. Super-resolution predicts pixels that fit patterns learned by the model; it cannot know an unreadable word, exact face, or texture that is absent from the source. It often improves edges and perceived sharpness, but it can also invent details or emphasize compression artifacts. Always inspect important results.

Should I download PNG or JPEG?

Choose PNG for text, line art, screenshots, or another lossless editing step. Choose JPEG for photos when a smaller file matters; this tool uses a high 0.92 quality setting. JPEG introduces lossy compression, so PNG is safer when you plan to edit or encode the image again.