AI Image Upscaler — Free 2× Photo Enlarger Online
Double the resolution of any photo while sharpening detail. Real-ESRGAN runs entirely in your browser to enlarge low-res images without the soft, blurry look of standard scaling.
Drop your JPG / PNG / WebP file hereTap to select a file
Supports JPG, PNG, WebP, up to 15MB
What to do next
Related tools
Image Resizer
Upload an image and resize it by target dimensions or scale percentage with high-quality interpolation.
imageSharpen Image
Sharpen any photo using an unsharp-mask convolution — choose Light, Medium, or Strong preset to recover crispness in soft, slightly-out-of-focus, or downscaled images. Real PNG output processed in your browser.
imageCompress Image
Reduce image file size while preserving visual quality.
imageResize Image
Resize images to exact dimensions or percentages.
imageJPG to PNG
Convert JPG images to PNG format.
imageImage to WebP
Convert images to the modern WebP format.
imageAbout AI Image Upscaler 2×
Most "image upscalers" are bicubic resizers in disguise. They make pixels bigger; they cannot make detail richer. The AI Image Upscaler is different — it runs a Real-ESRGAN super-resolution model trained to predict what extra detail should look like at 2× resolution. Edges sharpen, textures recover, and the result genuinely looks better than the source instead of just larger. The whole pipeline runs in your browser tab and the model runs efficiently for every image.
Real-ESRGAN — the engine the tool ships — is one of the strongest open-source super-resolution networks for general photos and anime art. It is particularly strong on illustrations, line work, screenshots, and product photography; on photographic subjects with fine textures (skin pores, fabric weaves, foliage) it produces visibly sharper edges with subtle micro-detail that a Lanczos resize cannot recreate. Where it struggles, like every model in this family, is on noisy or compressed source material — JPEG block artefacts get amplified along with the real detail, so heavily compressed inputs benefit from a clean-up pass first.
Memory is the headline constraint. Upscaling holds both the input and the 4× output in working memory at the same time, so a 4 MP source ends up needing 16 MP of output buffers plus the model weights and ONNX intermediate tensors. The 15 MB upload cap exists because larger inputs reliably crash the tab on mid-range mobile browsers. For very large source files, resize down to ~3000 px on the long edge first with Resize Image; the upscaled output is essentially indistinguishable from running on the full-resolution original at the resolutions humans actually look at images.
The model file is hosted on the same R2 bucket as the Tesseract OCR engine and the @imgly background-removal weights. On first use it downloads (~25-60 MB depending on the variant), then sits in your browser cache so subsequent images are fast. The original on your disk is never modified, and a quick check of the Network tab in DevTools while the tool runs confirms that no requests carry your photo.
How it works
- 1Drop a JPG, PNG or WebP photo onto the upload area. Files up to 15 MB are accepted; larger files should be downsized first with Resize Image.
- 2Pick the variant: Anime/Illustration is faster and tuned for crisp lines; General Photo is slightly slower with better natural-image fidelity.
- 3On first use, the chosen Real-ESRGAN model downloads from the Favtoo CDN and is cached in your browser. The download only happens once per model variant.
- 4The photo is tiled into overlapping 256×256 patches so the working memory stays bounded on mobile, then each patch passes through the super-resolution network.
- 5The upscaled patches are stitched back together with feathered overlaps so tile borders are invisible. The output is a 2× PNG.
- 6Inspect the result side-by-side with the original. Run the output through Compress Image afterwards if you need a smaller file for web upload.
Common use cases
- Restore a low-resolution Instagram-size photo so it can be printed on an A4 poster without going visibly soft
- Bring an old social-media profile picture up to a usable size for a fresh LinkedIn or business-card crop
- Sharpen a screenshot of dense text or a chart so the typography is legible when projected on a meeting-room display
- Upscale anime fan art or illustrations for a higher-resolution print run
- Enlarge a product photo for an e-commerce listing without the soft, blurry look of a basic resize
- Recover detail in a downsampled compressed JPEG before feeding it into a video edit at a larger render size
FAQ
How is this different from a normal resize?
A normal resize uses bicubic or Lanczos interpolation, which makes pixels bigger but cannot recover detail. Real-ESRGAN is a neural network trained to predict what the missing detail should look like, so the result actually looks sharper at 2× — not just bigger.
How long does it take?
A 1024×1024 photo finishes in 8–20 seconds on a modern laptop after the model is cached. The first run includes a one-time download of the Real-ESRGAN model.
Why is the input limit only 15 MB?
Upscaling holds both the input and the 4× output in memory at the same time, so the working set climbs quickly. 15 MB keeps mobile browsers stable.
Will my photo be uploaded?
In your browser. The Real-ESRGAN model is downloaded once and cached, then every subsequent photo is processed in the page itself.
Does it work on screenshots, anime art and old photos?
It handles photographs and screenshots well. Line-art and anime see the biggest visual gains because the network was trained on similar content. Heavily damaged or noisy scans should be cleaned up before upscaling for the best result.
Is it really better than just resizing in a desktop image editor?
On flat material (solid colour blocks, simple gradients) the difference is minimal — bicubic and Lanczos already do a good job. The gap opens up on edges and fine detail. A neural super-resolution network has been trained to predict the high-frequency content that an interpolation method literally cannot produce. The result is sharper line edges, recovered texture in cloth and skin, and crisper text. Side-by-side at 2× the difference is usually obvious within a few seconds.
Can I upscale by 4× or 8×?
The model variant currently shipped is 2×. For 4× you can run the tool twice (2× → 2×) and the result is comparable to a true 4× model on most subjects, with slightly more compounded artefacts. A native 4× model variant adds another model file to the download budget; we will add it as a second option once it is uploaded to the CDN.
Why is there a 15 MB input limit when the BG remover allows 25 MB?
Background removal predicts a per-pixel mask the same size as the input — a single channel of extra working memory. Upscaling produces an output four times the input area at higher bit depth, so the working set climbs much faster. 15 MB keeps the tab stable on mid-range mobile devices.
Will it sharpen blur or motion blur?
Not really. Super-resolution recovers detail that was lost to downscaling, but motion blur and out-of-focus blur are different physical processes — the underlying high-frequency information was never captured. The model can mildly de-blur a slightly soft photo, but it cannot rescue heavy blur. For that case, use Sharpen Image as a complementary pass.
Will my image be uploaded?
No. The model bytes come from our CDN on first run; after that, your image is read into memory inside the tab, processed by the Real-ESRGAN ONNX model running locally on your CPU/GPU via ONNX Runtime Web, and the result is offered as a download. The Network tab in DevTools confirms — no requests carry image data.
Why does the output have a different file extension than the input?
The output is always exported as PNG so the per-pixel detail is preserved without JPEG compression artefacts. If you need a JPEG for web upload, run the PNG through Image to WebP or a JPEG converter afterwards — applying lossy compression after the upscale is far better than applying it before.
Does it work on text and screenshots?
Yes, and screenshots are actually one of its strongest use cases. Text rendered to a bitmap, charts, code screenshots, and UI mockups all upscale cleanly because the model has seen plenty of similar content during training and the source has well-defined edges to lock onto.