How to use Image Upscaler
- 1Choose up to ten PNG, JPG or WebP images.
- 2Pick 2×, 4× or Fit to size, and check the mode and model download size.
- 3Click Upscale. Keep the page open while each image is processed.
- 4Inspect the slider and detail loupe, then download individual images or a ZIP.
What an image upscaler actually changes
An image upscaler increases the number of pixels in a picture. Ordinary resizing calculates extra pixels from nearby colours. A learned restoration model also uses patterns learned from other images to suggest sharper edges, smoother lines and plausible texture. Those suggestions can look better, especially when a small illustration has to fill a larger space. They are not a recovery of hidden original information.
Start with the clearest source you can find. An original camera file usually contains more useful detail than a screenshot of the same picture. A repeatedly forwarded JPEG may contain compression damage that no enlargement can fully undo. Keep the original beside every result. A larger file, a sharper outline or a higher pixel count does not by itself mean the image is more accurate.
Choose 2×, 4× or an exact size
Choose 2× for a modest enlargement, a mobile device or your first comparison. A 1,000 × 750 image becomes 2,000 × 1,500, with four times as many pixels. Choose 4× only when the destination needs it: the same source becomes 4,000 × 3,000, with sixteen times as many pixels. More output pixels require more memory and a longer processing time.
Fit to size keeps the original proportions. Enter a width of 1,920 pixels and leave the height empty for a full-HD-width image. If both dimensions are entered, the result fits inside that box without stretching or cropping. A square photo fitted to a portrait box stays square. The tool selects the necessary enlargement level and then resizes to the exact destination dimensions. Targets beyond 4× are rejected rather than silently repeated.
The A4 preset uses a 2,480 × 3,508 pixel box at 300 DPI. A photo with another aspect ratio will not fill the whole sheet without borders or cropping. PNG and JPG output receive density metadata; WebP receives the same pixel dimensions. Check the print application's physical dimensions before printing. Changing DPI alone cannot supply missing image detail.
Pick a mode, then inspect the result
Auto makes a simple guess from flat areas and colour variation. It is convenient for a mixed batch, but a product photo on white can resemble a drawing. At 4×, Photo uses a compact model for natural images and Drawing uses a separate model for bold outlines and flat colour. At 2×, every mode runs the compact photo model at 4× and averages the result down to 2×; on CPU this was three to seven times faster than a native 2× model in our tests. Try the other mode when a result becomes too smooth or too strongly outlined.
Clean up JPEG blocks applies a mild edge-aware filter at compression-block boundaries before enlargement. It only affects JPEG input. Turn it on for visible square patches in a heavily compressed photo; leave it off for a clean original or fine lettering. This is a small cleanup step, not a way to reverse all previous compression. Always compare the result with the switch off as well.
Preserve fine detail selects the slower Swin2SR model for either type of image. It is useful when structure matters more than speed. On CPU it takes about two to three times as long at 2×, and at 4× it runs two enlargement passes, which can take more than ten times as long. Try a small crop first, especially when your browser falls back to CPU processing.
What happened in our eight-image reference test
We tested four model candidates against bicubic resizing on eight paired fixtures. The originals were reduced to 128 pixels wide; the small photo and damaged-JPEG fixtures also received JPEG compression. Outputs were enlarged 4× and saved as PNG. The photo and portrait sources are public test photographs; the screenshot, logo, drawing and transparent graphic are controlled fixtures. The small natural-photo case is a proxy for a small shared picture, not a measured phone-camera capture.
The table below was measured on 14 November 2026 using the Python CPU reference pipeline with two execution threads, 128-pixel tiles, overlap and blending. These are processing measurements, excluding download and model initialization. They are not WebGPU, browser-WASM or phone timings. File sizes use 1,024 bytes per KB. Different browser encoders may produce different sizes.
| Image | Input | Bicubic PNG | Enhanced PNG | Bicubic time | Enhance time | What changed |
|---|---|---|---|---|---|---|
| Small natural photo | 3.5 KB | 162.0 KB | 207.7 KB | 0.007 s | 2.10 s | Sharper-looking fur, but lower reference fidelity than bicubic. |
| Portrait | 30.8 KB | 279.4 KB | 355.3 KB | 0.005 s | 2.98 s | Edges sharpened; skin and suit texture changed. |
| Product on white | 15.5 KB | 142.3 KB | 170.2 KB | 0.007 s | 2.25 s | Cup outline improved; inspect fine surface details. |
| Tiny screenshot text | 5.8 KB | 58.9 KB | 83.8 KB | 0.005 s | 10.81 s | Failed reliable letter recovery; some shapes were invented. |
| Logo | 3.9 KB | 42.0 KB | 68.1 KB | 0.004 s | 4.34 s | Cleaner boundaries than bicubic. |
| Drawing | 8.2 KB | 87.9 KB | 139.7 KB | 0.005 s | 5.65 s | Cleaner eyes and outlines; fine strokes changed. |
| Damaged JPEG | 1.7 KB | 159.8 KB | 210.8 KB | 0.004 s | 0.86 s | Reduced blockiness; fine texture remained limited. |
| Transparent graphic | 3.4 KB | 42.0 KB | 91.6 KB | 0.006 s | 7.74 s | Alpha retained separately; soft edges still need review. |
We compared the results with the larger originals using PSNR and SSIM as well as visual inspection. A separate lightweight restoration model preserved the most reference detail overall, but its two-pass 4× run was much slower in this CPU test. The compact photo option was the fastest learned model. Drawing mode gave crisp boundaries, although it sometimes changed thin strokes. There was no single winner on every image. The severely reduced screenshot failed reliable text recovery, which is why we do not promise to reveal unreadable words.
Faces, product details and small text
For faces, inspect eyes, teeth, hairlines and skin at 100%. An attractive change can still alter a person's appearance. Avoid repeated enlargement: feeding an enhanced result back into the model may reinforce invented texture. If a face already looks natural at 2×, increasing it to 4× just for a bigger number is rarely useful.
For product photos, check logos, stitching, labels and material texture. A sharper-looking surface can suggest a pattern that was not present. Use the original for factual product details. For screenshots, exporting the source document or taking a new screenshot at a larger display scale usually works better than restoring tiny letters. If a character was reduced to a few pixels, an upscaler cannot reliably determine which character it was. Verify every number and word before reuse.
The comparison slider shows both versions at the same display size. The detail loupe lets you move horizontally and vertically with keyboard controls, or point at the preview. At 100%, one output pixel occupies one CSS pixel; 400% helps expose edge changes and tile boundaries. The original side is enlarged for comparison. Your browser's preview interpolation is not the measured bicubic baseline in our reference table.
Keep transparent edges and choose an output format
PNG is the safest first choice for logos, screenshots and transparent graphics. It is lossless and does not use the quality slider. WebP supports transparency with a quality setting and can produce a smaller web image. JPG is useful for opaque photographs; transparent areas become white. If you need another background colour, keep a transparent PNG and place it on the desired background later.
The tool enlarges colour and transparency separately. This prevents the restoration model from inventing opacity. Completely transparent pixels receive nearby edge colour before processing, reducing dark fringes around cutouts. Very soft shadows and semi-transparent edges still deserve inspection against both light and dark backgrounds. Orientation stored by the camera is applied during decoding. Original EXIF and location metadata are not copied into the result.
Private processing, batch downloads and practical limits
Your images are decoded and processed on your device in a background worker. Image content is not sent to an upscaling server. The first click on Upscale downloads the selected model and its runtime; choosing files or opening this page does not trigger a model download. The model size is shown above the button. Cached models are reused while browser storage permits, but private browsing, storage pressure or clearing site data may require another download.
WebGPU is attempted first. If it is unavailable or processing fails, the tool attempts WASM SIMD. A fallback can be considerably slower. In Chrome, Edge and Firefox, WASM spreads the work over several CPU threads, which made it about three times faster on a 4-core computer in our tests; Safari uses one thread. The progress estimate learns from completed tiles and can change during a run. Cancel stops the worker, including an in-progress download or inference. It does not delete already completed results.
Add up to ten PNG, JPG or WebP images. They are processed sequentially and can be saved individually or together as a ZIP. The memory guard may recommend 2× or a smaller target before starting. It is a conservative estimate, not a guarantee against every browser's memory limit. A narrow viewport is not a measure of a phone's processing speed. Animated images, SVG, HEIC and HDR workflows are outside this tool's supported input set.
For the next step, crop the image to remove unused space, compress it to 200 KB for a size-limited form, or remove its background for a clean cutout. Crop before enlargement when possible; compress after you have checked the enhanced master copy.
Frequently asked questions
Is it free, and do I need an account?
Yes. There is no account, watermark or payment step. Processing limits come from your browser and available memory.
Can it make every blurry image sharp?
No. It can suggest clearer edges and texture, but severe blur, missing detail and unreadable text cannot be reliably recovered.
Why is my enlarged image much bigger in bytes?
2× creates four times as many pixels and 4× creates sixteen times as many. PNG preserves those pixels losslessly. Try WebP or JPG after checking the result.
Does A4 mean the image is guaranteed to print sharply?
No. A4 at 300 DPI describes the target pixel box and print density. The quality still depends on the original, and the tool preserves the image proportions.
Will my images leave this device?
The upscaler sends no image data to a server. It downloads model files, then processes your images in a browser worker. The report form is a separate, optional site feature.
Can I close the page while it is running?
No. Keep the page open until you download the result. Closing or reloading stops processing and removes unsaved results from the page.



