AI Image Upscaler

Nope, files aren’t uploaded to a third-party server; images are upscaled in the browser. This tool uses the Real-ESRGAN v3 script (around 5 MB cost on GPU) for upscaling photos, with Swin2SR as a backup. For illustrations, it uses Real-ESRGAN anime 6B (around 18 MB) and runs on WebGPU or WASM. For portrait images, it utilizes the GFPGAN script. Once the page loads, these models are cached in the browser, so you can upscale images even without an internet connection.

Photo: Uses Real-ESRGAN v3 (4× resolution) for photos, falling back to Swin2SR 2× if needed. Illustration: Uses Real-ESRGAN anime 6B for rendering still line art, badges, and comic pages on WebGPU or WASM. Graphic / logo: Simply uses Lanczos resampling and a tiny amount of unsharp masking. Auto-detect: Sends images with few colors / flat patches to the Illustration codepath. Restore faces is an additional pass with GFPGAN on photos only.

In short: yes. Standard upscaling will yield heavily blurred images, whereas an image upscaler will first reconstruct the image’s edges again. While still not being able to hold a candle to a corresponding desktop application (using even larger models and additional processing passes, e.g. for face restore or for noise stripping), it’s certainly much better than just upscaling.

To be honest, this won’t match Topaz or any desktop upscaler since those use AI to regenerate blurry sections of the image. Topaz Photo AI and Gigapixel run full-precision models on your GPU - denoise, sharpen, face recovery, and then upsample. However, it works well to upscale a 600-pixel landscape image to a 2,400-pixel version without blurring it. I’ve tested it extensively on various photos and illustrations, and it matches other paid image upscaling websites. It works best for upscaling illustrations or graphics without stretching or pixelating the image.

4× is best balance of quality and speed, followed by 2× (faster) and 3× (slightly faster than 4×, uses same 4× model, then Lanczos downsampled to 3×). 8× mode runs the 4× model followed by Lanczos upsampling to 8×. The illustration uses the 4× still-art model, then upsamples to 2×/3×/8× using Lanczos.

Noisy or low-light images, and portraits with Restore faces disabled. Small text and logos in illustration mode can create halos – switch to Graphic/logo mode. Tall and wide images larger than 8,192 pixels in either dimension will be pre-shrunk to fit the browser’s canvas, resulting in a soft and blurry image.

This depends on the infographic’s size. It pre-shrinks anything over 8,192 pixels. For best results, use an infographic smaller than 2,048 pixels.

For example, a 500×500 pixel image scaled up 4× would result in a 2,000×2,000 pixel image, about 6×6 inches at 300 DPI, which is suitable for low-quantity printing. Please note that this upscaler will not be as good as what you get with Topaz tools. I’ve tried to make this upscaler as good as possible, but please let me know how it performs for you and any ideas for improvement. Email me here: hello @ convertyard . com.

This setting runs an extra pass after photo upscaling, uses a small detector to find faces in the image, and then GFPGAN refines the eyes, skin, and mouth in each crop. It’s great for recent selfies, ID photos, and old, worn portraits. The only images where you’d disable this and let the image upscale as-is are already very sharp images, and this feature can make them look very plastic and fake. See also: Illustration/Graphic Modes. The first use of this feature downloads and caches a ~340MB file to the browser cache (which is then reused for subsequent uses of this feature—no images are uploaded).

For the best results, do a batch of no more than 5 for the smaller 2x to 4x upscales. If you’re upscaling 8x, it’s best to do one image at a time.