PRPickrack

Image Upscaler

Enlarge an image 2×, 3×, or 4× with smooth stepped resampling. Browser-side, no upload, no signup.

This is a high-quality bicubic upscaler, not an AI super-resolution model

Output is smoother than the browser default but won't invent missing detail the way Real-ESRGAN or Topaz Gigapixel do. Best for moderate enlargements of clean source images. A WASM Real-ESRGAN version is on the roadmap.

Drop an image here or click to browse

PNG, JPG, WebP, AVIF — up to 30 MB

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

Image Upscaler enlarges an image while keeping it as smooth as possible. The naive approach (single big resize) produces blocky output. This tool uses stepped 1.5× resampling — enlarging in incremental hops with high-quality bicubic interpolation at each step. That gives smoother gradients and softer edges than a single 4× jump.

Important honesty: this is not an AI super-resolution model. It won't invent missing detail the way Real-ESRGAN, SwinIR, or Topaz Gigapixel AI do. Those tools use neural networks trained on millions of images to *hallucinate* plausible detail when enlarging. Pickrack's upscaler uses traditional bicubic — best for moderate enlargements (2-3×) of already-clean source images. A WASM Real-ESRGAN version is on the roadmap but not yet shipped (the model is ~30 MB to download).

Free, no signup, no upload. Output is PNG (lossless). Up to 30 MB input.

Key features

  • 2×, 3×, and 4× scale factors — Pick a multiplier. Live preview of output dimensions before you click Upscale.
  • Stepped 1.5× resampling — Internally enlarges in 1.5× hops with high-quality bicubic at each step — smoother than a single big jump.
  • Browser-side, no upload — Source image stays in your browser. No server roundtrip. Faster for small images than uploading to a cloud service.
  • PNG lossless output — No JPG re-encoding artifacts. PNG output preserves every pixel of the bicubic result.
  • 8192px output safety cap — Caps the longest output edge at 8192 pixels to prevent browser memory crashes on very large enlargements.

How to use

  1. Step 1: Upload a clean image — Drop a PNG, JPG, WebP, or AVIF up to 30 MB. The cleaner the source (no JPG compression artifacts, no noise), the better the upscale result.
  2. Step 2: Pick a scale factor — 2× doubles each dimension (4× pixel count). 3× triples (9× pixels). 4× quadruples (16× pixels). Larger multipliers progressively lose more detail.
  3. Step 3: Click Upscale — Processing time scales with source size and target multiplier. A 2 MP image at 2× takes ~1-2 seconds; 4× takes ~3-5 seconds.
  4. Step 4: Download PNG — Click 'Download PNG' to save the upscaled result. File size will be roughly 4-16× the source PNG (proportional to pixel count).

When to use

  • Print a small web image — enlarge a 800px web image to 1600px or 2400px for a small print (acceptable up to ~6 inches at 300 DPI)
  • Improve a tiny avatar — take a 200×200 user avatar to 600×600 for a larger display context
  • Old phone photo — older smartphone photos (pre-2015) often were 4-6 MP; upscale 2× for a 16-24 MP look that better fits modern screens
  • Logo enlargement — vector is always better, but if you only have a small PNG logo, upscale 2-3× to make it usable in a header
  • Stock photo licensing — some stock sites offer small versions; if you legally have a small one and need larger, upscale 2× for casual web use
  • Pre-AI prep — for AI super-resolution tools that need a minimum input size, upscale 2× first to meet that floor

Frequently asked questions

Will this upscaler add detail like Real-ESRGAN does?

No. This is bicubic upscaling — smooth, but it cannot invent detail. Real-ESRGAN, SwinIR, and Topaz Gigapixel use neural networks trained on millions of images to hallucinate plausible detail (extra hair strands, clearer eyes, sharper text). They're 10-100× more compute-expensive and require model files of 20-200 MB. A WASM Real-ESRGAN tool is on Pickrack's roadmap for when the model can be small enough.

Why 'stepped' instead of single-pass upscale?

Bicubic interpolation introduces some smoothing each time it runs. Stepping in 1.5× hops applies the smoothing across multiple passes, producing softer gradients and less ringing at sharp edges than a single 4× pass. The tradeoff is slightly more compute (about 2× slower than single-pass) — a few seconds vs. instant.

What sources work best?

Clean, low-noise images: digital screenshots, vector exports as PNG, professionally lit photos. The upscaler can't fix JPG compression artifacts or sensor noise; it'll smooth them but they remain visible.

What's the max output size?

8192 pixels on the longest edge, regardless of source × scale. So a 3000×2000 source at 4× would *want* 12000×8000 but gets capped to 8192×5461 to avoid browser memory crashes.

Why PNG output, not JPG?

Upscaled images are typically used for further editing or print. JPG re-encoding at this stage would discard quality you just paid compute time to preserve. If you need a smaller file, run the PNG through Pickrack's Image Compressor afterward.

Does this work offline?

Yes once the page loads. The entire upscale is JavaScript + Canvas API, both built into the browser. No network call after the page is cached.

Can I upscale to a specific pixel dimension instead of 2×/3×/4×?

Not directly in this tool — only the 3 multipliers are supported. For an exact dimension target, upscale to the nearest multiplier that overshoots, then use Pickrack's Image Resizer to scale down to the exact target. Two steps.

Related tools

Image Resizer

Resize JPG, PNG, WebP images to any dimensions. Browser-side, no upload, no quality loss for downscaling.

Image Compressor

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

Image Cropper

Crop images with preset aspect ratios (1:1, 4:3, 16:9, 9:16) or freeform. Browser-side, no upload, instant export to PNG or JPG.