AI Image Upscaler

Top, a 4×4 source becomes an 8×8 output at the same physical size; outlined cells came from the source, the rest were synthesised. Bottom left, Fast mode fills a gap by averaging: new = A + (B − A) × ½. Bottom right, the same edge rendered by bilinear versus a GAN — averaging produces a soft ramp, the model restores the step.

A 640 × 480 product shot has to be printed at 2× — 1,280 × 960 — without the fabric weave turning into mush.

1. The arithmetic:640 × 480 is 307,200 pixels. At 2× the tool writes a 1,280 × 960 canvas holding 1,228,800 pixels — four times as many, of which only a quarter were ever measured by the camera.
2. Fast mode:Multi-step bilinear interpolation fills each new pixel from its neighbours. Along an edge running from a 0.85 to a 0.15 brightness pixel the result is a ramp rather than a step, so lettering and woven texture read as soft.
3. Balanced or Ultra:APISR GRL (~7MB) or APISR RRDB (~18MB) runs locally through Transformers.js and ONNX Runtime Web. Large images are processed in 256px tiles, and the GAN predicts high-frequency texture instead of averaging it, so the edge comes back as an edge.
4. First run vs later runs:The model downloads once and is then cached. On a second upscale the header shows the CACHED pill and the download stage is skipped, so only inference time remains.
5. Output:The canvas is encoded as a PNG at quality 1.0. The result panel reports the output dimensions and the elapsed seconds, and the before/after slider lets you scrub between the original and the 2× result to check the edges yourself.

Why upscale in your browser instead of a cloud AI service?

Most AI upscalers work the same way: you upload the image, a GPU server somewhere runs the model, and you download the result. That round trip has three costs, and you pay all of them on every single image. This tool inverts the architecture. The APISR networks — roughly 7 MB for the Balanced model and 18 MB for Ultra — are downloaded once into your browser's cache and then run on your own hardware through WebGPU, falling back to WASM on the CPU. Your photo is decoded, cut into 256 px tiles, and reconstructed pixel by pixel without a single byte of image data leaving the machine.

Privacy is the most obvious win. Photographs of clients covered by an NDA, unreleased product shots, scans of family documents, and ID photos are exactly the files people are least comfortable handing to a stranger's upload queue — and exactly the files that most often need enlarging. Speed is the less obvious win: after the one-time model download, a 2× job on a 1,000 px source typically finishes in seconds, with no upload progress bar, no processing queue, and no download wait. Cost is the third: cloud upscalers meter their work in credits, so a batch of fifty product images becomes a subscription. Here the marginal cost of the fifty-first image is the same as the first — zero.

One honest caveat keeps the comparison fair. Server-side models can be larger than anything that fits in a browser, and on extreme enlargements of very noisy sources a premium desktop product can still extract more detail. If your job is a clean 2×–4× on reasonably decent input — which covers most e-commerce, web, and print-preparation work — local quality is competitive, and the overall workflow is faster because there is nothing to wait for.

Getting a clean result in five steps

(1) Start from the best source you have. Upscaling amplifies whatever the input contains — including compression artifacts — so a clean original, a PNG screenshot, or a lightly compressed JPEG will beat a heavily re-shared social media image every time. (2) Pick the mode that matches the content. Fast is instant and download-free, good for previews and screenshots where softness matters less than speed. Balanced (APISR GRL) shines on anime, illustrations, and product images with clean lines. Ultra (APISR RRDB) is tuned for real photographs — portraits, landscapes, textures. (3) Prefer 2× when it does the job. Every extra multiplier asks the model to invent more, so 2× is the safest, sharpest choice; go to 3× or 4× only when the target size genuinely demands it. (4) Judge the result at 100% zoom with the before/after slider, not at fit-to-screen, where everything looks sharp. (5) Download as PNG. PNG is lossless, so it avoids stacking a fresh generation of JPEG artifacts on top of the freshly reconstructed detail — convert to JPEG later, at the size you actually need, if the platform requires it.

Common mistakes — and how to keep expectations realistic

The most common failure is asking for too much. A 100 px thumbnail contains a fixed, tiny amount of information; a 4× upscale writes 16,000 output pixels for every 100 input pixels, and no algorithm can reconstruct a recognizable face or readable lettering from data that was never recorded. Treat 4× as a tool for modest gaps — turning a 800 px web asset into a 3,200 px print asset — not a time machine for lost resolution.

The second mistake is upscaling damaged sources. Heavy JPEG compression leaves ringing halos around edges; the model will faithfully sharpen those halos along with the edges. If the source is badly compressed, repair it first or accept a softer result. The third mistake is judging only by overall impression. Fine text, straight lines, and repeated patterns are where generative models wobble — check a screenshot's menu labels and a portrait's eyes at full zoom before shipping. Faces deserve particular care: the model reconstructs plausible skin texture, and plausible is not always identical to the person. Finally, resist chaining. Upscale, compress, then upscale again and each pass compounds error; do all resizing in one trip, from the original, once.

Frequently asked questions about the AI upscaler