Upscale Image
Enlarge a small or blurry image 2× or 4× with a super-resolution model that runs on your device, not our servers. Sharper edges, JPEG artefacts cleaned up, full-resolution PNG or JPEG out — and your picture never leaves the browser.
Upscale
Cleans JPEG artefacts and noise while keeping detail — the right default.
About the model
Real-ESRGAN "general x4v3" (SRVGGNetCompact, BSD-3-Clause), 4.9 MB, fp32. Runs in a Web Worker with ONNX Runtime on your CPU, in 256-pixel tiles; the 2× option runs the 4× model and box-filters the result down, which is sharper than a native 2× would be. The three styles are the official weights blended with the "denoise" weights at 0, 0.5 and 1. Each style is cached on your device after its first download.
What this tool does
It takes a small, soft or compressed image and produces a larger one — twice or four times the width and height — with edges reconstructed and compression artefacts removed, using a super-resolution neural network (Real-ESRGAN general x4v3, BSD-3-Clause). The model is downloaded from ai-pix.co.uk once, cached, and run by ONNX Runtime in a Web Worker on your own CPU. There is no upload, no queue and no per-image charge because no server does any work.
Like the background remover, this is one of the two AI-Pix tools that genuinely earn the "AI" in the name. The compressor, converter and resizer are plain image encoding and we don't call them AI.
When it is the right tool
- Web images you only have in small — a product photo from an old listing, a logo saved from a website, a thumbnail that has to go on a poster or a slide.
- Old scans and phone photos from years ago — 640×480 or 1024×768 originals that look muddy on a modern screen. Use the Smooth style for grainy ones.
- Marketplace listings — Amazon and eBay want at least 1000 px on the long side; a 2× pass on a 600 px photo gets you there cleanly.
- Screenshots and UI captures at 4× for documentation; Smooth keeps flat colours flat.
- Anything confidential — unreleased products, client assets, family photos. Nothing leaves the device.
How it works under the hood
- Your browser reads the pixels and, if the image is larger than the limit for the chosen factor, reduces it first.
- The image is cut into 256-pixel tiles with a 12-pixel overlap so the network sees real neighbours at every seam. Tiles keep memory flat on phones.
- Each tile goes to a Web Worker where ONNX Runtime (WebAssembly, multi-threaded where the browser allows) runs the 4× model; the output tile is pasted into the result canvas, box-filtered down by two if you asked for 2×.
- If the source had transparency, the alpha channel is resampled and re-applied. The result is encoded as PNG or JPEG on your device and offered as a download.
Limits and honesty
- Super-resolution invents detail. It is excellent for edges, textures and cleaning up compression; it cannot make unreadable text readable or recover a face that was never captured. Do not use it for anything forensic.
- On very noisy or heavily compressed inputs the Sharp style can look over-processed ("plastic" skin, etched grass). Switch to Balanced or Smooth.
- Inputs are capped at 2.1 MP (2×) and 1.05 MP (4×) to bound run time; larger images are reduced first and the page says so. Output is up to 8 MP (2×) or 16 MP (4×).
- It is CPU-bound. A 1 MP image takes roughly 15–40 s on a laptop, longer on a phone; the page shows tile progress and you can cancel.
- Old browsers without WebAssembly SIMD will be much slower. The page tells you how many threads it is using.
Frequently asked questions
Is my image uploaded to upscale it?
No. Services like Upscale.media, Let's Enhance, Bigjpg or Topaz's web tools send your file to a GPU server and meter you by credits. Here a 4.9 MB neural network is downloaded from this site once, stored in your browser's cache, and runs on your own CPU in a Web Worker. Turn your Wi-Fi off after it loads and it still works. That is also why there are no credits, no watermark and no account.
What does "AI upscaling" actually do that a normal resize doesn't?
A normal resize (bicubic, Lanczos — what the resizer tool and every image editor do) interpolates between existing pixels, so an enlargement is always softer than the original. A super-resolution model has learned what edges, textures and compression artefacts look like at low and high resolution, so it reconstructs plausible sharp detail and removes JPEG blocking at the same time. The detail is inferred, not recovered: text that is unreadable stays unreadable, and a face that is a blur stays a blur (just a cleaner one).
Which model is it and how good is it?
Real-ESRGAN "general x4v3", a compact network released under the BSD-3-Clause licence by the Real-ESRGAN authors (Wang et al., Tencent ARC). It is the fast, general-purpose model of that family — the one chosen for speed in most desktop apps — and it is well suited to photos, product shots, web images and old scans. Large-scale models used by paid cloud services are heavier and slightly better on faces and fine text; they cannot run in a browser tab.
Why 2× and 4×? Can I choose another size?
The model is trained for 4×. The 2× option runs the same model and then halves the result with a box filter, which gives a cleaner 2× than a native 2× network would. For an arbitrary final size, upscale here, then use the resizer to hit the exact dimensions.
What do the Sharp / Balanced / Smooth styles mean?
The authors published two sets of weights: the base model and a "wdn" (with denoise) model, and recommend blending them depending on how noisy the input is. Sharp is the base model, Smooth is the denoise model, Balanced is the 50/50 blend — the upstream default. If a compressed photo looks over-processed on Sharp, use Balanced; for grainy scans or tiny web images use Smooth.
Why is there a size limit, and how long does it take?
Inference runs on your CPU at roughly 1–3 seconds per 256-pixel tile on a laptop (longer on phones), so a 1-megapixel image at 2× takes around 15–40 seconds and the page shows a tile counter. Inputs are capped at 2.1 MP for 2× and 1.05 MP for 4× — about 1920×1080 and 1024×1024 — and anything larger is reduced first. Upscaling is for small images; if you are starting from a 12 MP photo you do not need it.
Does it keep transparency?
Yes. For PNGs with an alpha channel the colour channels go through the model and the alpha channel is resampled separately and re-applied, so logos and cut-outs (for example from the background remover) stay transparent. Save as PNG to keep it; JPEG has no transparency.
Does it work on my phone?
Yes on recent phones. The model download is small (4.9 MB); processing is slower than on a laptop — expect 30–90 seconds for a 1 MP image at 2×. Keep the tab in the foreground; browsers pause background tabs.