AI Image Upscaler
Enlarge blurry photos and illustrations 2× or 4×, or keep the size and just sharpen.
Drag & drop or click to select imagesTap to select images
JPG, PNG, WebP, AVIF · multiple images are processed one at a timeUpscaled images
Open source and model licenses
How to Use
Use 'Photo' for photographs, 'Illustration' for manga and anime art, and 'Fast' on computers without a GPU. Choose 2× or 4×, or 1× to sharpen at the same size.
Drag & drop, click or paste with Ctrl/⌘+V and processing starts immediately. The model (5–64MB) downloads on the first run only and is cached in your browser.
Compare original and result up close, then save as PNG, JPG or WebP. Download several images as one ZIP or send them straight to the compressor or converter.
FAQ
Which AI models are used?
Three official Real-ESRGAN weights from Tencent ARC: 'Photo' is realesrgan-x4plus (64MB), 'Illustration' is x4plus-anime-6B (18MB) and 'Fast' is realesr-general-x4v3 (5MB). All are 4× models; 2× is produced by downscaling the 4× result.
How does it compare with Upscayl or Topaz?
It uses the same weights as Upscayl's default model (realesrgan-x4plus), so output quality is the same. The difference is that nothing is installed and files never leave your device. Topaz uses paid proprietary models and is not a direct comparison.
Do I need a GPU?
Browsers with WebGPU (recent Chrome and Edge, Safari 26+, Firefox 141+ on Windows) run on the GPU and finish in seconds to under a minute. Without a GPU it falls back to the CPU, where the photo and illustration models are very slow — use the 'Fast' model.
How long does it take?
On a laptop GPU, upscaling a 1000×1000 photo 4× takes about 30–60 seconds with the photo model and a few seconds with the fast model. The remaining time is shown while processing. The first run adds the model download.
How large can the image be?
Up to 50 million output pixels (16 million on mobile). That means sources up to about 2000×1500 for 4× and 4000×3000 for 2×. Resize larger images first or lower the scale.
Can I sharpen without changing the size?
Choose '1×' as the scale. The 4× result the AI redraws is gathered back to the original size, removing block noise and blur from repeated messenger re-saves and making lines crisp. It is not a restoration of the original — thin lines can merge and textures can change — so check it with the original/result zoom comparison.
Are transparent PNGs supported?
Yes. Transparency is preserved in PNG and WebP output. JPG has no transparency, so transparent areas are filled with white.
