AI Image Upscaler Guide: When to Use 2× vs 4× Super-Resolution

Enlarging an image used to mean blurry edges and visible compression artefacts. AI super-resolution changes that by predicting missing detail from learned patterns — rather than simply stretching pixels. Here is how the NSS AI Upscaler works and how to choose the right settings.

What is AI super-resolution?

Traditional upscaling (bicubic, Lanczos) interpolates between existing pixels. If you double an image's size, each new pixel gets a value somewhere between its neighbours. The result is smooth but soft — information that was never in the original cannot be invented.

AI super-resolution trains a neural network on millions of image pairs: a low-resolution version and its high-resolution original. The model learns to recognise patterns — text serifs, fabric weaves, hair strands, skin pores — and hallucinate plausible high-frequency detail when upscaling. The result has sharper edges, finer textures, and more apparent detail than interpolation alone.

Swin2SR: what the model actually does

NSS uses Swin2SR, a state-of-the-art super-resolution model based on the Swin Transformer architecture. It processes your image in overlapping 64×64 pixel tiles and predicts a 128×128 pixel output for each tile. The tiles are assembled, blended at their edges, and the final 2× upscaled image is exported as a lossless PNG.

The model is ~47 MB and is downloaded once and cached in your browser. All processing runs locally — your image is never uploaded.

2× vs 4×: which should you choose?

2× AI upscale uses the full Swin2SR model at every pixel. It produces the sharpest, most accurate result.

4× upscale applies Swin2SR for the first 2× step, then uses a high-quality bicubic pass for the second 2× step. This is faster than running AI inference twice and produces output close to full AI 4× quality for most content.

As a rule: use 2× unless you specifically need the larger output dimensions of 4×.

Processing time and image size

Processing time scales with the number of tiles. The upscaler splits your image into 512×512 pixel tiles, processes each one, and assembles the result.

Images larger than 2048px on the longest side are pre-downscaled to 2048px before tiling. The output is still upscaled from the original — it's the intermediate tile count that's reduced to keep processing time reasonable.

A real-time tile-by-tile progress indicator shows estimated time remaining once processing starts.

What types of images benefit most?

High benefit:

• Product photos from mobile cameras (compressed, slightly soft)
• Old or archival photos (scanned prints, vintage film)
• Screenshots of UI or web content (text, icons, logos)
• Game screenshots
• Any image that was resized down during sharing

Moderate benefit:

• DSLR photos at native resolution (these are already sharp; upscaling adds modest improvement)
• Headshots and portraits

Limited benefit:

• Images that are already blurry (motion blur, out of focus) — AI cannot recover detail that was never captured
• Heavily compressed JPEGs (AI upscaling preserves compression artefacts, potentially making them more visible)
• Very noisy images — the model may interpret noise as texture and sharpen it

Practical tips

Remove the background first, then upscale. If you plan to cut out a product against a transparent background, do the background removal first. Upscaling the cutout afterwards produces sharper edges on the subject. Upscaling before removal gives the segmentation model more pixels to work with — try both for critical work.

Export as PNG. The upscaler always outputs lossless PNG to preserve every detail the model added. If you need a smaller file, convert the PNG to WebP after upscaling.

4K inputs are automatically handled. Very large images are pre-downscaled to 2048px before tiling — this reduces processing from ~50 tiles to ~16 tiles. If you need the absolute maximum detail from a large original, use the 2× scale.

Use the Browser cache. The Swin2SR model (~47 MB) is cached in your browser after the first download. Subsequent upscale sessions skip the download entirely.