Image Upscaler
Classical resampling blurs detail as it enlarges; AI super-resolution synthesises plausible detail that preserves apparent sharpness. Upload a low-resolution image and this tool offers both — Lanczos for a fast, faithful upscale or a neural model for a 2x/3x/4x result that looks like the original was shot at the larger size.
How to upscale an image
-
1
Upload
JPG, PNG, WebP; low-res photos and illustrations both supported.
-
2
Pick upscale factor
2x, 3x or 4x.
-
3
Pick method
Lanczos (fast, faithful) or AI (slower, invents plausible detail).
-
4
Download
Output at the enlarged resolution; JPG or PNG.
Classical resampling vs AI upscaling
Side-by-side
| Method | Speed | Detail recovery | Artefact risk |
|---|---|---|---|
| Nearest neighbour | Instant | None | Blocky |
| Bilinear | Instant | None | Blurry |
| Bicubic | Fast | Slight | Soft |
| Lanczos | Fast | Modest | Ringing at high contrast |
| AI (ESRGAN class) | Slow (seconds) | Significant | Invented detail, sometimes wrong |
AI upscalers do not “recover” lost information — they predict what plausible higher-resolution content would look like based on the patterns they learned during training. For photos of real objects the result is usually convincing; for text and faces, accuracy depends heavily on the source quality.
When to use which
- AI: photo scans at low DPI, old web images, small product shots you need at larger display size.
- Lanczos: screenshots and UI mockups where every pixel is intentional.
- Nearest neighbour: pixel art (Game Boy sprites, 16-bit games).
Limits
- No upscaler adds colour or corrects exposure — fix those first.
- Faces upscaled with generic models can shift features subtly; for portraits, use a face-aware upscaler.
- Upscaling a JPG with visible compression artefacts amplifies the artefacts. De-noise first.
- 4x on a 500 × 500 image gives 2000 × 2000 but no more real detail than the source contained.
File sizes
A 4x upscale produces 16x the pixel count. Expect file sizes to balloon proportionally — plan storage accordingly or apply WebP/AVIF compression on the result.
Frequently Asked Questions
For natural photos, usually yes. For precise technical images (schematics, UI screenshots, pixel art), Lanczos or nearest neighbour is better because AI can hallucinate details that were never there.
Generic upscalers can subtly shift features. For portraits, use a face-aware model (GFPGAN or similar) which is tuned to reconstruct facial detail more accurately. The tool offers a face mode in the AI path.
AI upscaling runs on a GPU server because browser inference is too slow. Images are processed and deleted immediately after the result is downloaded. Lanczos upscaling runs in your browser with no upload.
Lanczos yes, unlimited batch. AI upscaling is queued and limited by GPU throughput — up to 20 images per batch typically.