Image / AI enlargement
AI Image Upscaler: 2× or 4× in Your Browser
Enlarge PNG, JPEG and WebP images with Real-ESRGAN General x4v3. Compare the original and result with a slider, preserve transparency in PNG, or download JPEG with adjustable quality.
AI Image Upscaler: 2× or 4× in Your Browser: An AI model, Real-ESRGAN, runs in your browser and redraws your picture at four times the width and height, sharpening edges and textures instead of just blurring pixels larger. Large pictures are processed in small overlapping tiles that are joined without visible seams. Choose 2× for a smaller result. Nothing is uploaded. Runs 100% locally in your browser with zero server file uploads.
- Category
- Image tools
- Runs
- In your browser
- Cost
- Free · no sign-up
- Availability
- Ready to use
Runs entirely in your browser
AI upscaling invents plausible detail; it can't recover real detail, faces or text exactly.
Real-ESRGAN General x4v3 by Xintao Wang and contributors, Qualcomm AI Hub export. BSD-3-Clause. ONNX Runtime Web: MIT.
Real-ESRGAN research and export
Xintao Wang, Liangbin Xie, Chao Dong and Ying Shan. Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data. ICCV Workshops (ICCVW), 2021. Source: https://openaccess.thecvf.com/content/ICCV2021W/AIM/html/Wang_Real-ESRGAN_Training_Real-World_Blind_Super-Resolution_With_Pure_Synthetic_Data_ICCVW_2021_paper.html
This tool uses Real-ESRGAN General x4v3, the Qualcomm AI Hub float ONNX export with fixed 128 × 128 RGB input and 512 × 512 output. Repository: https://github.com/xinntao/Real-ESRGAN . Export catalogue: https://aihub.qualcomm.com/models/real_esrgan_general_x4v3 .
BSD-3-Clause and runtime attribution
Real-ESRGAN: copyright (c) 2021, Xintao Wang, BSD-3-Clause. Retain the copyright, conditions and disclaimer with redistribution; the copyright holder and contributors must not be presented as endorsing derived products. Bundled notices: /data/esrgan-x4v3/LICENSE-realesrgan.txt and /data/esrgan-x4v3/LICENSE-qualcomm.txt .
ONNX Runtime Web 1.30.0: copyright Microsoft Corporation, MIT. Its copyright and permission notice accompany the self-hosted runtime at /data/ort/LICENSE-onnxruntime.txt . The model and runtime are used locally in your browser.
How to use it
- Choose or drop an image up to 1 megapixel and 2,000 px on its longest side, under 20 MB.
- Select 2× or 4× and start. Follow the tile progress or cancel at any time.
- Compare before and after, then download PNG or choose JPEG quality and download.
Privacy & limitations
Images never leave the browser. The runtime and model load from this site's own origin only when you start. No external AI service is used.
Related tools
Frequently asked questions
Can it recover faces or text exactly?
AI upscaling invents plausible detail; it can't recover real detail, faces or text exactly. Check the result before using it where accuracy matters.
How long does upscaling take?
Time depends on your device and the image size. Large images may take several minutes. The tile count and remaining processing estimate appear after inference starts; loading, resampling and encoding take additional time. Cancel terminates the worker.
What happens to transparency?
PNG keeps alpha, enlarged with bicubic interpolation and recombined with the model's RGB output. JPEG flattens transparency onto white. The 2× resample uses premultiplied alpha to avoid colour fringes.
Why limit the input?
The input is limited to 1 megapixel, 2,000 px on the long side and 20 MB to keep browser memory use manageable. A 4× enlargement has sixteen times as many pixels. Resize larger images first.
Free tool · runs in your browser · no account required