AI Image Upscaler

How the upscaler works

Try a Google search for “image upscaler,” and most of these tools will upload your photos to a server, cap you at a few images per day, or limit you to upscaling one or two before you need to sign up and pay a monthly fee to keep using them.

These tools range from $9 to $500 depending on the plan you select. Most of these require users to upload photos to a server for their software to use AI to re-encode these photos (more on how image upscaling works below).

ConvertYard’s image upscaler works locally in your browser, so files aren’t uploaded to an unknown server, there’s no need to sign up, and there’s no usage limit.

When you open or drop a file, three things happen:

1. The browser detects the image type, whether it’s a photograph, graphic, or illustration (that is, if you leave it on auto-detect). The tool bases the detection on color count and flat regions. But for the best results, you’ll have to specify what type of image it is.
2. A model will run in your browser, locally. How fast the process is depends on the image type, the memory available on your computer, and browser GPU support. One of these will load and cache:
   • Real-ESRGAN v3 (around 5 MB) — photos at 4× on GPU browsers (WebGPU).
   • Swin2SR — photos at 2×, and the 4× fallback when WebGPU isn’t available.
   • Real-ESRGAN anime 6B (around 18 MB) — for illustrations, line art, badges, or comic graphics.
   • Lanczos + unsharp mask — for logos, graphics, and anything else where you don’t want a neural net making decisions.
3. Optional second pass if the Restore faces option is turned on. The tool detects faces in the photo and runs GFPGAN (around 340 MB, cached after first use) over each detected face to clean up eyes, skin tone, and mouth.

Everything stays inside the browser. After you open the page and the tools are loaded, it will keep working even if you turn off the Wi-Fi connection — the models are saved in the browser’s cache.

Upscaling vs. enlarging: what’s the difference?

One of the biggest misconceptions is that upscaling and enlarging images are the same thing, but in reality these are different processes.

Image enlarging relies on pure math: the tool simply enlarges the pixels, spreads them over a larger grid, and fills the gaps by averaging “neighbors” (bilinear, bicubic, Lanczos). In layman’s terms, it stretches the image to meet size requirements; it doesn’t enhance anything. Soft edges stay soft, blurry details get blurrier, and small text gets bigger and fuzzier. This is exactly what happens when you drag a photo’s corner in a document, or enlarge an image to twice its size in a basic photo editor.

Here’s an example of an enlarged image using a photo editor.

Instead of recalculating and stretching pixels, upscaling uses an AI model trained on millions of sharp images. Instead of averaging, it “predicts” the missing detail and how it should look based on patterns it learned during training — like how an eyelash looks, a clean font edge, or the texture of a brick. The output includes detail that stays close to the original file.

And here’s an example of an upscaled image using ConvertYard.

One quick way to tell if an image is upscaled or enlarged is to zoom in around 4×. The enlarged photo will have noticeably blurry edges, while the upscaled photo will have sharper edges and some recovered texture.

For photographs and illustrations, this tool uses AI to upscale the image, meaning it will predict, enhance, and even add detail that wasn’t in the original.

However, for logos and graphics, it doesn’t use AI but a “Lanczos + unsharp mask” combination, otherwise called a “classic sharpening resize.” It works like a classic photo enlargement but adds a sharpening filter at the end to boost contrast. It’s called “classic” because this is the technique photographers and graphic artists have used to enlarge images for over 20 years.

Picking the right image type

I understand this tool has an auto-detect feature, and it gets it right most of the time. But for the best results, I suggest choosing the type of photo manually.

• Photo (AI): real-world photographs like landscapes, images of people, or places. Any photo you take on vacation falls into this category.
• Illustration (AI): graphic art like comics, anime stills, badges, stickers — anything with lines, text, or drawn art.
• Graphic / logo (Lanczos): UI mockups, QR codes, icons, anything with geometric lines or edges. This model skips AI because it’s prone to creating a halo effect on tight text.

I’ve tested these options extensively, and honestly, use either Photo or Illustration for any graphic that requires text, sharp lines, or logos. This will produce the best results for most non-photographed images.

So if using the Illustration mode produces a glow or halo around the lettering or logo, try the Graphic / logo option.

Choosing a scale

• 2× (fastest option): the best option for photos for websites. It basically doubles the image size, so a 900×300 pixel wide image becomes 1800×600 pixels — perfect for headers or hero images.
• 3×: this runs the 4× model, then uses Lanczos to downsize the image. It uses the 4× model to upscale the image to four times its original size, then uses Lanczos to downsize it to 3×.
• 4× (recommended option): the sweet spot for upscaling images. Best quality-to-speed ratio, and the model is specifically trained to upscale images 4×.
• 8× (slowest option): runs the 4× model once, then uses Lanczos to upsample the result to 8×. Because of the extra resize and larger output, it uses a lot of memory. Avoid using this unless you need the larger dimensions.

One thing to note is that the 4× result is limited to 8,192 pixels on the longest side. If you’re working with large photographs or images that would exceed this, the tool will shrink the photo or graphic first before upscaling, which softens and in some cases makes the upscaled image worse. For the best results, keep the source image under 2,048 pixels and use a PNG source.

If you’re working with a JPG file, convert it to PNG first before upscaling to prevent any loss of pixel data.

Best practices by image type

For photos

1. Start by using the 4× upscale and auto-detect.
2. Turn on the Enhance option if the source image is slightly soft — it adds contrast and sharpens the edges without touching the skin or sky elements of the image.
3. If you’re upscaling selfies, ID photos, or scanned portraits, turn on the Restore faces option. Skip this step on photos that are already sharp.
4. For photos with a lot of noise taken in low-light conditions, expect a softer result. The model smooths out the noise and can’t invent something that wasn’t there.

Illustrations and comics

1. Select Illustration mode manually if you have a graphic image with lots of flat color. The Auto option may work, but can miscategorize clear JPG files.
2. Avoid turning on the Enhance or Restore faces option. These options aren’t used in Illustration mode, so they won’t affect the output.
3. I mentioned earlier to use Illustration mode on graphic files, including logos or screenshots with lots of text, but if it produces halos or glows, try Graphic / logo instead.

Logos and UI

This is the last resort for any logo or images with lots of text. Choose this model only if the output from Illustration mode produces a glow or halo.

1. Use this on logos, graphic files, or images with lots of text.
2. If the source comes from a vector file, re-export it from the original file to a PNG, then try using this upscaler. However, if you have access to a graphic editor, export it to the size you need instead of using an upscaler.

Prints

• A 500×500 photograph upscaled at 4× gives you an image with a 2,000×2,000 pixel dimension, which is good enough for a 6×6 inch printout at 300 DPI. This is fine for small printouts (between 4R and 6R).
• For larger print requirements, start from a higher-resolution source with dimensions of at least 2,000×2,000 pixels, then upscale to 4×. If you need something larger, you’ll have to pay for a premium tool like Topaz for larger-format prints.

Batching

While this tool can run batches of files, I suggest processing one file at a time to prevent the browser from freezing because it runs out of memory. If you’re upscaling images 2× or 4×, you can do up to 5 images per batch, but for 8×, do one image at a time for smaller images below 800 pixels. If you try doing 20 images or more, you’ll see a warning.

Output format

I recommend leaving the output format at the default setting, which is to match the input, unless you have a good reason to switch. For the best results (and I’ve said this repeatedly), convert lossy formats like JPG into PNG before upscaling. Choose WebP if a smaller file size is a high priority.

What this tool is not

Take note that this is a browser upscaler, not software like Topaz Gigapixel. These desktop apps run full-precision models that can do multiple passes, have dedicated denoisers, and allow more memory usage than a browser upscaler.

A browser upscaler will provide comparable results at smaller output dimensions. I’ve tested Topaz, and the results I get are similar to what I got from this image upscaler. The only time I’d recommend using a dedicated image upscaling app is if you need a larger-format printout for the graphic and you don’t have access to the source vector file, or you don’t have an editor to open it. But if you only need an upscaled image for your website, the ConvertYard image upscaler is enough for the task. Try it out and let me know what you think.