- Does this upload my image to a server?
- No. All processing happens inside your browser tab using WebGL-accelerated TensorFlow.js. Your image is never sent anywhere. You can verify this by turning off Wi-Fi after the page loads — the upscaler continues to work because both the AI model and your image are local. No analytics or pixel data are collected.
- What is the difference between 2× and 4× upscaling?
- 2× doubles each dimension: a 512×512 image becomes 1024×1024 (4× the pixel count). 4× quadruples: 512×512 becomes 2048×2048 (16× the pixel count). 4× gives the best visible improvement but takes longer and produces a larger file. Use 2× for quick previews, line art, or screenshots where extreme enlargement is not needed.
- What does the noise reduction level do exactly?
- Before upscaling, a two-pass box blur (horizontal then vertical) smooths compression artefacts — the blocky JPEG noise common in low-quality scans and screenshots. Level 0 = off (preserves original texture). Level 1 = radius 1 px, ideal for clean digital art. Level 2 = radius 2 px, good for typical JPEG-saved fan art or manga scans. Level 3 = radius 3 px, best for heavily compressed or screentone-heavy pages where the AI would otherwise amplify the pattern noise. Higher levels trade a small amount of fine detail for cleaner edge upscaling.
- What does post-sharpening do, and should I use it?
- After ESRGAN upscales the image, an unsharp mask pass increases perceived edge contrast. Level 0 = off. Level 1 = subtle (safe for all anime/illustrations). Level 2 = medium, recommended for line art or pencil work. Level 3 = strong, best for very soft or blurry source images. Avoid sharpening on manga scans — it amplifies screentone dots into harsh patterns.
- Should I export as PNG or JPEG?
- PNG is lossless — ideal for images you plan to edit further or that have transparent backgrounds (note: transparency is flattened to white in this tool). JPEG at quality 90+ looks indistinguishable from PNG at a fraction of the file size, and is a better choice for sharing on social media or embedding in web pages. Quality 92 is a good default; go lower only if file size is a hard constraint.
- Why does the first run take longer?
- On the first upscale the browser compiles the ESRGAN model shaders for your GPU. Subsequent runs in the same session are faster because the compiled shaders are cached. The model weights (~880 KB per scale) are also cached by your browser after the first download, so reloading the page is fast.
- What image sizes are supported?
- There is no hard size limit, but very large images (over 2000×2000 px at 4× scale) can take several minutes and may cause the browser tab to run out of GPU memory on low-end devices. For best performance, upscale images under 1000×1000 px at 4×, or under 2000×2000 px at 2×. If processing fails, try reducing the source image first.