Free AI Image Upscaler — Swin2SR Transformer 4× Detail Enhancement

Swin2SR transformer 4× upscale with detail enhancement. Better than bicubic for photos. Self-hosted GPU, free.

AI Image Upscaler uses the Swin2SR transformer model (an attention-based super-resolution network published at ECCV 2022) to upscale images 4× while reconstructing plausible detail in the new pixels. Bicubic upscaling makes images larger but blurry; Swin2SR makes them larger AND adds sharpness/texture/edges that look natural — trained on millions of high-resolution photos to predict what each blurry pixel SHOULD look like.

Use it for: enlarging old photos for print, reviving small product images for an Etsy listing, sharpening low-resolution screenshots before posting, fixing pixelated profile pictures, upscaling generated AI images that came out blurry, or any case where you have a small image and need a bigger version that doesn't look like a magnified blob.

The model runs on a self-hosted GPU (RTX 3060) on the operator's home server. ~3-10 seconds per upscale for typical images. No upload to a third-party cloud service that may stamp watermarks or rate-limit your account. Output stays in your browser tab.

Key features

Swin2SR transformer model — Attention-based super-resolution architecture from ECCV 2022. State-of-the-art for general photo upscaling — significantly outperforms bicubic, ESRGAN, and RealESRGAN on standard benchmarks.
4× and 2× scaling — Model produces 4× output natively; we downsample to 2× if you choose (still better quality than browser-native bicubic 2×).
Detail synthesis, not just stretching — Edges sharpen, textures restore, faces gain pore detail (not perfect, but plausible). What was a blurry pixel becomes a plausible color + sharper boundary.
GPU-accelerated — Runs on a self-hosted RTX 3060 — typical 800×600 photo takes ~3 seconds to 4×-upscale. 1200×1200 takes ~8 seconds.
Self-hosted = free — No cloud API costs. The operator's GPU does the work. Free for all users. No signup, no quota.
Browser-side download — Output PNG arrives as a data URL — never written to disk on the server, never logged, never tracked.

How to use

Step 1: Upload image — Drop or click. JPG / PNG / WebP supported. Max ~6MB (about a 4000×3000 JPG). Smaller is faster.
Step 2: Pick scale — 4× for maximum detail recovery (recommended). 2× for moderate upscale (smaller output file, slightly faster).
Step 3: Click Upscale — Wait ~3-10s for the GPU to run inference. Original + upscaled shown side-by-side.
Step 4: Download — Output is a PNG (lossless). Save with your preferred filename.

When to use

Old family photos — scan at 300dpi, upscale 4× → 1200dpi-equivalent for big print
Etsy / eBay product photos — small thumbnail from a supplier upscaled to listing-quality
Profile picture rescue — tiny avatar (200×200) upscaled to 800×800 for high-DPI display
Game screenshot upscaling — 720p screenshot to 2880p for crisp blog post embed
AI-generated image fix — DALL-E / Midjourney output at 512×512 → upscale to 2048×2048 for print
Old artwork digitization — phone photo of a painting upscaled to fine-art resolution
Historical / archival photos — gentle upscale + sharpen without inventing detail (Swin2SR is conservative)

Frequently asked questions

How does Swin2SR compare to RealESRGAN or DALL-E HD?

Swin2SR is a transformer-based super-resolution model. RealESRGAN is GAN-based. In general: Swin2SR is more conservative (preserves real structure but adds less synthetic detail), RealESRGAN is more aggressive (sharper output but occasional 'hallucinated' features). DALL-E HD or Topaz Photo AI use larger more complex models with even more detail synthesis. Swin2SR is a good middle ground for everyday photo upscaling.

Is the upscale lossless?

Output is a lossless PNG. But the upscaling process is NOT lossless — the model invents detail. A 200×200 image upscaled to 800×800 is 16× more pixels, and the model fills the new 15/16 of pixels with predicted values. They're plausible, not the 'true' detail (which can't be recovered).

Why does the result look soft / lacks expected sharpness?

Swin2SR is conservative by design — it doesn't aggressively over-sharpen. If you want crisper output, run the result through a sharpening filter (or use Pickrack's image-compressor at high quality to apply a mild sharpen). For maximum sharpening at the cost of artifacts, try Real-ESRGAN externally.

Does it work for text / screenshots?

Yes but the model is trained on photos. Text in screenshots upscales OK but can show slight blur at edges. For pure text/diagram upscaling, vector formats (SVG) or raster-with-anti-alias-off scaling work better.

What about faces?

General Swin2SR handles faces decently. For face-specific upscaling (skin texture, eye detail, lip shape), specialized face models like GFPGAN or CodeFormer beat Swin2SR. Not currently in Pickrack — on the roadmap.

Why is my image pre-resized?

GPU VRAM is finite. Input images larger than ~1500×1500 are pre-resized to fit. After 4× upscale they'd produce 6000×6000+ output which is 100MB+ raw. If you need true ultra-resolution, split image into tiles externally and upscale each tile separately.

Why does the badge sometimes say 'GPU offline'?

The Swin2SR model runs on the operator's home PC GPU. If the PC is asleep / shut down / network down, the badge shows offline and the Upscale button is disabled. The PC is the operator's daily-use machine, so availability varies. Status refreshes every 60s.

Is my image uploaded anywhere?

Your image is uploaded to the operator's home PC via Pickrack's server, processed locally by the GPU, output returned as base64. The Pickrack server doesn't log image content or store the file. The operator's PC GPU sees the image briefly during inference but writes nothing to disk.

Related tools

Image Upscaler

Enlarge images 2×, 3×, or 4× with smooth stepped resampling. Browser-side, no upload, no signup. Best for moderate enlargements of clean sources.

Image Compressor

Compress JPG, PNG, WebP images with adjustable quality. Browser-side, no upload, real-time preview.

AI-Art QR Generator

Stable Diffusion + ControlNet creates a scannable QR code blended into a custom artwork. Self-hosted GPU.