Upscale Image
Upscale one JPG, PNG or WebP image up to 50 MB by 2× or 4× in your browser, without uploading it. Output is limited to 16 megapixels.
How to use
- Choose one image: Drop a JPG, PNG or WebP file onto the box or click to pick it. One image at a time, up to 50 MB. Once it is read, the page shows the original size, the output size and an estimated peak memory figure.
- Set the factor and the output format: Pick 2× or 4× under Upscale factor, then PNG, JPG or WebP under Output format — it starts on whatever your original was. For JPG and WebP a Quality slider (60–100, default 92) appears; PNG is written losslessly and ignores it.
- Run it and check the result: Click Start AI upscale. The first run downloads about 0.9 MB of model files, then a progress bar follows the tile-by-tile rebuild. Cancel requests a stop during loading or inference; inference stops at the next tile. The final export cannot currently be cancelled, and your image and settings survive a cancel either way. Zoom the downloaded file to 100% before you use it.
How to upscale image files with AI
The enlargement is done by an ESRGAN Slim super-resolution model running on TensorFlow.js inside your own browser. Your photo is never sent to a server: the model weights are fetched from this site (about 0.9 MB for the factor you picked), the picture is decoded locally, and the finished file is written straight to your downloads. Closing the tab clears the current image and result from this page; downloaded files and cached model files remain.
Neither TensorFlow.js nor the weights are loaded until you press Start AI upscale, so simply opening the page and choosing a file costs no model download at all. On the first run the ~0.9 MB arrives once; after that the same browser normally serves it from its cache.
Inside the run, the image is rebuilt in 64-pixel tiles with a 4-pixel overlap so neighboring tiles blend instead of showing seams, and the page gives up the main thread between tiles — which is why the progress bar keeps moving and the Cancel button stays responsive. Cancel requests a stop during loading or inference; inference stops at the next tile. The final export cannot currently be cancelled, so a click during the last stage still ends in a finished file. TensorFlow.js picks the backend: WebGL where the browser allows it, otherwise the CPU backend, and the page says so when it falls back.
2× or 4×: which factor to choose
2× → input up to 4 MP; square example: 2000×2000 4× → input up to 1 MP; square example: 1000×1000 Output cap: 16 MP; neither side has a separate 4000-pixel limit.
There are exactly two factors, 2× and 4×, and both multiply the width and the height — 4× therefore produces 16 times the pixels, not 4 times. For avatars, product shots and anything already a reasonable size, 2× is the sensible default: it finishes sooner and the invented detail is less visible. Keep 4× for genuinely small originals, such as a thumbnail or a logo pulled off an old web page.
The cap is on the total pixel count of the output, 16 megapixels, not on either side on its own: the page multiplies width × height × factor² and compares that one number. The square figures above are just the easiest way to picture it. A long, thin 3000×1000 photo is 3 megapixels, so it passes at 2× even though one side is wider than 2000, while a 1500×1500 photo at 4× works out to 36 megapixels and is refused. When a combination goes over, the estimate line turns into a warning and the run stops before it starts rather than freezing your browser halfway — switch to 2× or shrink the original first. Because the estimate appears as soon as the file is read, you can see the figure (and the peak-memory estimate next to it) before committing to anything.
There is no 3×, no 8× and no custom target size. If you need a specific pixel width, upscale to the nearest factor above it and resize down afterwards, which also tidies up any softness the model left behind.
How sharp the result really is
This is not lossless enlargement, and no AI upscaler is. The model predicts plausible texture and edges from the patterns it was trained on, so what it adds is a convincing guess rather than detail recovered from your file. That distinction matters for anything used as evidence — a license plate, a serial number, small print on a scanned document — where a sharp-looking invented character is worse than a blurry honest one.
The upstream measurement on this same model, dated 2026-08-22, puts a number on the trade-off. A 1200×800 photo was reduced to 600×400 and then taken back to 1200×800 two ways. Against the untouched original, the AI result scored 19.86 dB PSNR and the plain bicubic resize scored 20.33 dB, so pixel-for-pixel fidelity was slightly worse with AI. Perceived sharpness went the other way: Laplacian variance was 377 for the AI result against 245 for bicubic, with the original at 16104. Sharper to look at, marginally less faithful — the normal bargain for ESRGAN-type models, on one test image.
In practice that means the tool is a good fit for making a small picture usable on a bigger screen, a print at viewing distance, or a listing photo that looks mushy next to its neighbors. It is a poor fit when the pixels themselves are the point. Either way, open the download and zoom to 100% before you publish it.
What the enlarged file keeps and what it loses
Dimensions are exact: a 640×480 photo at 2× comes out 1280×960, and at 4× it comes out 2560×1920. Nothing is cropped and the aspect ratio is untouched, so the enlarged file drops straight into a layout built for the original shape.
Transparency survives if the output format can carry it. With a PNG or WebP source and PNG or WebP selected, the original alpha channel is scaled to the new size and reapplied, so a transparent logo stays transparent. Choosing JPG instead fills every transparent area with white, because the JPEG format has no alpha channel at all.
EXIF metadata does not survive. The result is re-encoded from the model's output, so the capture date, camera model, lens settings, GPS coordinates, XMP and the embedded color profile are all absent from the file you download. That is handy when you want a clean image to publish, and something to plan around when the shot data matters — keep the original alongside it.
The pixel count increases 4-fold at 2× and 16-fold at 4×. File size depends on the source, output format and quality setting; compare the downloaded file with the original. The download is named after your file with the factor appended, for example photo-upscaled-2x.png, and the extension follows the output format you chose.
A 4× PNG can easily be several megabytes, so if the result is too heavy to email or upload, run it through Compress Image afterwards.
What this tool does not do
It is one super-resolution model and nothing else. It does not repair old photographs — no scratch removal, no colorizing, no face reconstruction — and there is no separate anime or illustration mode: photos and drawings go through the same ESRGAN Slim weights. There are no denoise settings either, and no batch queue, so images are handled one at a time.
It also cannot be hurried. Where WebGL is unavailable the CPU backend does the work, and the upstream run in a WebGL-less container took about 158 seconds for a 600×400 image at 2× and about 338 seconds for 800×600 at 2×. The page warns you when it falls back, and Cancel gets you out of a loading or inference stage at the next tile, but there is no server to hand the job to — this tool has no upload path and therefore no remote fallback. On a device that struggles, a smaller original at 2× is the way through.
Animation is flattened rather than upscaled. GIF files are rejected outright, and an animated WebP is read as a single frame, so what you download is one still picture. The accepted inputs are JPG, PNG and WebP; AVIF, HEIC, SVG and anything else is refused with a message naming those three, so convert first if your file is one of them.
There is also no target file size here: you choose a factor, not a number of kilobytes. If what you actually need is a file under a stated limit, upscale first and set the size afterwards.
Need a specific KB limit? Use Compress Image to Size after upscaling.
Frequently asked questions
- Can you really upscale an image without losing quality?
- Not in the strict sense — the extra pixels are predicted, not recovered. On the upstream 2026-08-22 test (a 1200×800 photo reduced to 600×400 and enlarged back), the AI result measured 19.86 dB PSNR against 20.33 dB for a plain bicubic resize, while its sharpness measure was 377 versus 245. It looks crisper than ordinary interpolation; it is not a more faithful copy of the original.
- How large an image can I upscale here?
- The output is capped at 16 megapixels in total, so 2× accepts an input of up to 4 megapixels and 4× accepts up to 1 megapixel. For square images that works out to 2000×2000 and 1000×1000 respectively, but those are only examples — neither side has its own 4000-pixel limit, so a 3000×1000 photo at 2× is accepted while a 1500×1500 photo at 4× (36 megapixels) is refused before the run starts. The file itself can be up to 50 MB.
- Is my photo uploaded to a server?
- No. The AI model runs inside your browser, and only the model files themselves are fetched over the network — about 0.9 MB, from this site. Your image bytes are never transmitted, which is why there is also no account to create and nothing to delete afterwards.
- Why is it taking minutes instead of seconds?
- Your browser is probably on the CPU backend because WebGL is unavailable; check whether the WebGL warning is showing on the page. The upstream test in a WebGL-less container took about 158 seconds for a 600×400 image at 2× and about 338 seconds for 800×600 at 2×. Cancel stops the rebuild at the next tile, and 2× on a smaller original is far quicker.
- Will a transparent PNG stay transparent?
- Yes, as long as you keep PNG or WebP as the output format — the original transparency is scaled up and reapplied to the result. Switching the output to JPG turns every transparent area white, because JPEG cannot store transparency.
- Does the upscaled file keep EXIF and GPS data?
- No. The image is re-encoded from the model's output, so capture date, camera and lens details, GPS coordinates, XMP and the ICC color profile are not carried over. Keep your original if you need that information.
- Can it fix a blurry old scan, remove scratches or add color?
- No — this tool only enlarges. Restoration, colorizing and face repair are different models and are not part of it, so a damaged photo comes back larger with the same damage, rendered more sharply. There is no anime-specific mode either; every image uses the same ESRGAN Slim weights.
- Can I upscale HEIC or animated WebP images?
- A HEIC file has to be converted first: the three formats this tool reads are JPG, PNG and WebP, and anything else — HEIC, GIF, AVIF, SVG — is refused with a message naming those three. On an iPhone, Settings › Camera › Formats › Most Compatible makes the camera save JPG. An animated WebP is accepted, but the model reads one frame, so the download is a single still picture.
- Can I upscale multiple images at once?
- No. This tool processes one image per run, up to 50 MB. After a result, start over and pick the next file; the model is already in memory, so the following runs skip the 0.9 MB download. Ten product photos means ten runs.