AI image upscaler

Using the AI image upscaler

The tool is built around the decision you actually need to make: is the new detail useful, or merely convincing at a small preview size? Every control keeps that answer visible.

Add images and manage the queue

Drop JPG, PNG, WebP or AVIF files on the large target. You can also paste an image or use Choose images. Each file appears in the queue with its dimensions and byte size. Select a row to inspect it. Add images keeps the current queue, while Clear queue releases every source and result from the page.

Choose the scale and output

2× doubles each edge and creates four times as many pixels. 4× creates sixteen times as many. The selected image panel shows the exact result before processing. PNG preserves transparency. JPG and WebP reveal a quality slider because both formats trade some image information for a smaller file.

Run a batch and understand progress

Upscale runs every waiting item in order. The first use fetches a 4.9 MB model and prepares it in a worker. Tile progress then shows how much of the current image is complete. Cancel stops the active job and keeps the source queue intact.

Compare the real result

Drag the ↔ handle over the image to move the split. You can also focus it and use the arrow keys. Hold to see original removes the result while you hold it. Fit shows the whole picture. 100% maps one result pixel to one screen pixel, which is where false edges, waxy skin and broken lettering become clear. The plus and minus buttons cover the space between those views. Full screen gives the comparison more room.

Download only the version you checked

The selected image panel reports the encoded result size after processing. Download result saves that item in your chosen format. When at least two results are ready, Download all as ZIP appears below it. Source files, result pixels and previews are held in memory. They are not saved to your floi account.

What the model changes

Ordinary enlargement interpolates between known pixels. AI super-resolution also predicts the texture likely to sit between them. That can make hair, foliage and material edges look more natural. It can also invent a stroke in small text or a feature on a distant face.

This tool uses the compact Real-ESRGAN general 4× model. Its authors describe the small model as faster and lighter, with weaker denoising and deblurring than larger alternatives. The tradeoff suits an in-browser tool. It is not the right choice for forensic work or for a damaged photograph that needs manual restoration.

Why the 36 megapixel limit exists

A compressed file hides how much memory its pixels require. A 3000×3000 image at 4× would become 12,000×12,000. The final RGBA bitmap alone would occupy about 576 MB. Model tensors and the source need more memory beside it. The tool caps output at 36 megapixels and 12,000 pixels per edge, then explains the rejected size before starting.

If a source is too large, choose 2× or use the image resizer first. If you only need fewer bytes rather than more pixels, use the image compressor. Upscaling is the most expensive option, so it should answer a real size problem.

What it will not do

It will not recover evidence. The model predicts plausible detail. It does not reveal information that the source failed to capture.

It will not keep metadata. The browser decodes the file to pixels and writes a new one. EXIF, GPS and camera settings are not copied into the download.

It will not beat every desktop model. A large GPU application can offer several specialist models and process bigger files. This tool favours a private, inspectable workflow that opens without an install or account.