Upscale images 2× or 4× with AI or classic resampling
This free image upscaler enlarges photos, screenshots, artwork and small web images to two or four times their width and height. Choose AI · ESRGAN for photos, where a neural network redraws edges cleanly instead of simply stretching pixels, or Classic · Lanczos-3 for a fast, faithful enlargement that works well on any phone.
A before/after slider with a 100% zoom lets you inspect the result pixel by pixel before you download it. Everything happens on your device: the AI model is downloaded once from a public CDN and runs in your browser with TensorFlow.js, so your image is never uploaded.
Be realistic about what upscaling can do. The AI model invents plausible detail based on what it learned from training photos — it makes edges crisper and removes jagged steps, but it cannot recover writing, faces or numbers that were never captured. Blurry text stays unreadable, just with smoother outlines.
How to use the Image Upscaler
- Drop an image onto the upload area, click to choose one, or paste it with Ctrl+V.
- Choose 2× or 4×. The note below shows the exact output size and whether the image needs to be reduced first.
- Pick AI · ESRGAN (Balanced or Fast model) or Classic · Lanczos-3 with adjustable sharpening.
- Press Upscale image and follow the progress bar. You can cancel at any time.
- Drag the compare slider, switch to 100% to check details, then choose a format and press Download upscaled image.
Image Upscaler features
ESRGAN models via UpscalerJS
Uses the MIT-licensed ESRGAN Medium and ESRGAN Slim models from UpscalerJS, running on TensorFlow.js with GPU acceleration through WebGL when available.
Tiled processing
The image is processed in 128-pixel tiles with a 12-pixel overlap, so memory stays low and there are no visible seams.
Classic Lanczos-3 mode
A windowed-sinc resampler with premultiplied alpha and luminance-only unsharp masking — instant, no download, and kind to low-end phones.
Transparency kept
PNG transparency is preserved: in AI mode the colour goes through the model and the alpha channel is resampled with Lanczos-3.
Before/after comparison
A draggable split view and a 1:1 zoom show exactly what the upscaler added.
Private and free
No upload, no account, no watermark and no daily limit.
AI or Classic: which should I use?
| Image | Best method | Why |
|---|---|---|
| Photo of people, places or products | AI · ESRGAN Medium | Cleaner, sharper edges and fewer jaggies than ordinary resizing |
| Old phone or web photo, slow device | AI · ESRGAN Slim | A third of the download and several times faster |
| Logo, icon, pixel art or flat graphic | Classic · Lanczos-3 | Keeps colours and shapes exact; AI may add texture to flat areas |
| Screenshot with text | Classic, low sharpening | Text stays faithful; AI can distort letter shapes |
| Very large image (over 2 megapixels) | Classic | AI input is limited to 2 MP (2×) or 1 MP (4×) to keep processing time reasonable |
The model, the method and the limits
The AI mode uses ESRGAN Medium (2.8 MB) and ESRGAN Slim (0.9 MB) from the open-source UpscalerJS project, published under the MIT licence and trained on the DIV2K dataset of high-resolution photos. Despite the ESRGAN name, both are compact Residual Dense Networks (RDN), a design chosen to run quickly in a browser. TensorFlow.js (Apache-2.0) executes the network on your graphics processor through WebGL, or on the CPU if no GPU is available — which is much slower.
Lanczos-3 computes each new pixel from the six nearest original pixels in each direction, weighted by a windowed sinc function. It is the reference method for high-quality resizing: sharper than bilinear or bicubic, with very slight ringing next to hard edges. A mild unsharp mask on brightness only restores crispness without colour fringes.
The output is capped at 16 megapixels, a size every modern browser and phone can handle. AI input is capped at 2 megapixels for 2× and 1 megapixel for 4×; larger images are reduced first, and the note tells you the exact sizes before you start.
Tips for the best upscale
- Start from the original file, not a screenshot of it or a copy downloaded from a chat app — every recompression removes detail.
- Remove noise first with the Image Denoiser; the AI model treats grain as texture and enlarges it.
- Crop to the part you need with Crop Image before a 4× upscale — a smaller input upscales faster and can stay within the AI limit.
- For an exact print size, upscale 2× or 4× here, then set the final dimensions with the Image Resizer.
- Save as PNG for further editing; choose JPG or WebP for a smaller file to share.
Image Upscaler FAQ
How do I upscale an image without losing quality?
Open the image upscaler, choose 2× or 4×, select AI · ESRGAN and press Upscale image. The AI model produces much sharper results than ordinary resizing, though no method can add true detail that the original never contained.
Can an AI image upscaler make blurry text readable?
No. The model invents plausible detail; it cannot know which letters or numbers were there. It may make blurry text look smoother or even change letter shapes, so never rely on upscaled text for evidence or documents.
Is my image uploaded to a server?
No. TensorFlow.js runs the ESRGAN model inside your browser. Only the library and model files are downloaded from a public CDN; your image never leaves your device.
Why is AI upscaling slow on my phone?
The network does millions of calculations per tile. Phones with weaker graphics take longer, and browsers without WebGL fall back to the CPU. Use the Fast model, a 2× scale, or Classic mode, which is nearly instant.
What is the maximum image size?
The output can be up to 16 megapixels (for example 4,000 × 4,000 or 4,896 × 3,264). AI mode accepts up to 2 MP for 2× and 1 MP for 4×; larger images are reduced first.
Which licence do the AI models use?
ESRGAN Medium and ESRGAN Slim are published by UpscalerJS under the MIT licence, and TensorFlow.js is Apache-2.0, so they can be used freely, including for commercial images.