You can upscale an image online without sending it anywhere. The super resolution model downloads into your browser, about one megabyte on the first run, and then enlarges the photo on your own graphics chip. Nothing is uploaded, there is no account, and there is no per image limit: choose 2x, 3x or 4x and download.
What you get is a rebuilt edge, not new information. The model sharpens lines and cleans up compression noise as it scales, so a soft phone photo looks crisper, while a blurred detail stays blurred. A 4x pass turns a 1000 x 750 photo into 4000 x 3000 pixels.
If your photos must not leave your device, this is the route to take: everything happens locally, and the only download is the model itself. If you would rather let a larger cloud model do the work and accept an upload, Aihangsoft also has an AI Image Enhancer, and the two paths are compared below.
The problem with most online upscalers is not the algorithm, it is the round trip: you hand a photo to a server you cannot audit, then trust what comes back. This page covers the other arrangement, where a compact neural network is downloaded into the page and the photo never leaves the machine it was already on.
AI Image Upscaler
Enlarge a photo 2x, 3x or 4x with an AI super resolution model that runs in your browser. Sharpening, noise reduction and a before and after slider. Nothing is uploaded.
What 2x, 3x and 4x actually do
The factor multiplies both edges of the image, so the pixel count grows with the square of the number you pick. That is why 4x is not twice the work of 2x, it is four times the pixels of it.
| Factor | A 1000 x 750 photo becomes | Pixels vs the original | What it is for |
|---|---|---|---|
| 2x | 2000 x 1500 | 4 times | The default. Fastest, lightest on memory, the best balance for most photos |
| 3x | 3000 x 2250 | 9 times | When 2x is not enough room for a crop or a larger screen |
| 4x | 4000 x 3000 | 16 times | The largest file and the heaviest run. Best for print, where output size matters most |
The tool's own example: a 4x pass turns a 1000 x 750 photo into 4000 x 3000 pixels. What changes is not only size. The network rebuilds edge contrast as it scales, so lines, lettering and hard edges gain definition, and flat areas such as sky or a studio backdrop stay clean instead of picking up noise.
Going higher does not add proportionally more of that. Past a point the extra pixels are smooth rather than detailed, which is why the interface defaults to 2x and labels it the fastest and safest choice.
How to upscale an image without uploading it
- Add your image. Drop a JPG, PNG, WebP or AVIF file into the tool, or click to browse. You can select several at once, and the queue holds up to 20 images and runs them one after another.
- Pick a factor. Choose 2x, 3x or 4x. Start at 2x. The factor sets the output dimensions and is the single biggest influence on how long the run takes.
- Tune sharpening and noise reduction. Sharpening adds edge contrast on top of the AI result and starts at 15 percent; leave it lower on portraits, where it can exaggerate skin texture. Noise reduction smooths grain and JPEG speckle before the image is enlarged and starts at 10 percent; keep it low if you want to preserve film grain or fabric texture.
- Choose an output format. PNG is the default and is lossless. Pick JPEG or WebP for a much smaller file, or stay on PNG when the image has transparency.
- Compare and download. Drag the divider to compare the original and the result at the same output size. Save a single image, or export the whole batch once several results are ready.
One detail makes that slider useful. Both sides are rendered at the output dimensions, and the original is scaled up with a bicubic filter first, so what you compare is the difference in detail rather than the difference in scale.
This runs on your own device
The tool is a browser front end for an ESRGAN style super resolution model and the TensorFlow.js runtime that executes it, and both are fetched into the page instead of running on a server. Your image is decoded from local storage into page memory, scaled on the graphics chip through WebGL, and encoded back into a file, all inside the tab.
Because the heavy lifting is local, three things follow. There is no account and no sign in. There is no credit balance and no daily cap, so the twentieth image of the day is treated exactly like the first. And there is no upload step, so a photo of a document, a client mock up or a family picture is not sitting on someone else's disk while you wait for the result.
What gets downloaded
The first run at a given factor downloads the model, about one megabyte, and compiles it for your graphics chip. The model used here is the slim variant and it is cached per factor, so switching from 2x to 4x loads a second model while reusing the first. Later images at the same factor start immediately.
How to check that your photo is not uploaded
Open your browser developer tools, go to the network panel, clear it, and run an upscale. You will see the model and library files arriving from a CDN on the first run, and you will not see any request that carries your image. That is the concrete difference between an on device tool and a converter site, and it takes about ten seconds to confirm for yourself.
Local upscaler or Cloud AI enhancer
Aihangsoft offers both, and they are genuinely different tools rather than a free and a paid version of the same thing. The local upscaler keeps the file on your machine. The AI Image Enhancer uploads it to a server, where a much larger generative model runs, and it meters your runs by a daily allowance and by model tier instead of leaving them unlimited. The guide to enhancing photo quality online covers that side in detail.
| AI Image Upscaler (this page) | AI Image Enhancer (Cloud AI) | |
|---|---|---|
| Where your photo goes | Stays on your device | Uploaded to the server for processing |
| Model | A compact ESRGAN model downloaded into the page, about 1 MB | A large generative model in the cloud, in three tiers |
| Cost and limits | Free, unlimited, no account | Runs are metered by a daily allowance and by model tier |
| Factors | 2x, 3x, 4x | 1x, 2x, 3x, 4x |
| Output ceiling | 8192 px long edge and 24 megapixels, set by browser memory | 6144 px, or 10240 px on the stronger tiers |
| Input formats | JPG, PNG, WebP, AVIF | PNG, JPG, WebP, TIFF, BMP, HEIC, up to 10 MB each |
| Result retention | Nothing is stored anywhere | Result deleted from the server after 30 minutes |
Choose the local upscaler when the photo should not be uploaded, when you want to run a large batch without watching a counter, or when you have a decent GPU and the work is routine. Choose the cloud enhancer when the source is badly degraded and you want the strongest reconstruction available, when you need a format the browser tool does not read such as TIFF or HEIC, or when the machine in front of you is weak.
A reasonable default is to try the local tool first. It costs nothing, it starts in seconds, and on a mildly soft photo it is usually enough. Move to the cloud when you can see that the local result is not sufficient, rather than reaching for it out of habit.
AI Image Enhancer
The other route: a larger cloud model, 1x to 4x, and three model tiers. It uploads your photo for processing, so use it when the file is not sensitive.
What this upscaler cannot do
These limits are worth reading before you invest time in a bad source.
- It rebuilds edges, it does not recover facts. The model reconstructs plausible edge contrast, so a blurred word, number plate or serial number comes out looking sharper without becoming the actual characters. If the goal is to read text from an image, use an OCR tool such as Image to Text, which extracts the characters that are really there.
- Higher factors are slower and heavier, not proportionally better. Push the number up and the file gets bigger and the run gets longer, while the amount of genuine detail added changes much less. 2x is the safest general choice.
- Speed depends on your graphics chip. The same photo can take noticeably longer on a phone or an older laptop than on a desktop with a discrete GPU. There is no server to fall back on, so the hardware in front of you sets the pace.
- Output is re-encoded. You choose PNG, JPEG or WebP, and your original is left untouched on disk, but the file you download is a new copy rather than the one you started with.
- Transparency needs a format that supports it. The alpha channel is upscaled along with the colours, so cut outs keep a clean edge, but if you export that result as JPEG the transparent area comes out black. Pick PNG or WebP for anything with a transparent background.
- The first run downloads the model. About one megabyte, compiled for your GPU. That is the price of doing the work locally, and later images at the same factor skip it.
- A browser has a memory ceiling. A run that would exceed 8192 pixels on the long edge or 24 megapixels overall is refused, and the tool suggests 2x or a smaller source instead. Very large photos may need to be resized or cropped first.
- There is no manual repair. Sharpening and noise reduction are global sliders. There is no brush for one bad corner and no way to mask a face, so a targeted fix still belongs in a photo editor.