AI Upscaler vs Bicubic / Lanczos — When to Use Which
"AI upscaler" tools promise to enlarge images without losing detail. They sound magical. They are sometimes magical. They are also sometimes slower and worse than a 30-year-old algorithm called Lanczos. This article explains the actual difference, when each approach wins, and how to pick the right one for the image you have.
Tool referenced: paste-to-download.com/upscale — it picks the right method automatically based on image content, but you can override.
The Three Approaches
Bicubic
The default in most image editors since 2000. Computes new pixels from a 4×4 neighborhood of existing pixels using cubic curves. Fast (milliseconds per megapixel), simple, well-understood. Result: smooth but slightly blurry enlargement. No new information is invented — just smooth interpolation between existing pixels.
Lanczos
More sophisticated than bicubic. Computes new pixels from a wider neighborhood (typically 8×8) using sinc-based math. Slower than bicubic but still fast (still milliseconds per megapixel). Result: sharper than bicubic, especially for line art, logos, and screenshots. Best non-AI resampler for content that doesn't need invented detail.
AI Super-Resolution
A deep neural network trained on millions of image pairs (low-res → high-res). The network learns to invent plausible high-resolution detail from low-res input. Slower (seconds to tens of seconds per megapixel). Result: striking detail on photos — fine texture, eyelashes, fabric weave — that simply isn't present in the source. The detail is plausible, not real.
Common models in 2026: Swin2SR, Real-ESRGAN, SRGAN, Lightweight Super-Resolution variants.
Where AI Wins
AI upscalers shine on photographic content where the network has lots of training examples:
- Faces (especially headshots, portraits)
- Skin texture, hair, fabric
- Natural landscapes (foliage, water, clouds)
- Animals (fur, scales, feathers)
- Old or blurry photos restored to credible sharpness
- Low-resolution downloads that need to be enlarged for print
On these, AI fills in detail that simply wouldn't exist with bicubic or Lanczos. A 400 × 400 portrait can become a convincing 1600 × 1600 with AI; with bicubic the same image just gets bigger and softer.
Where AI Loses (or Goes Wrong)
Line art and logos
Vector-like content with hard edges and flat colors confuses many AI models. The network was trained on photos, so it may try to introduce "texture" that doesn't belong — creating subtle noise on what should be flat colors, or wavy edges on what should be straight lines.
Lanczos beats AI here every time. Or, better, convert the image to vector with /vectorize and forget about upscaling.
Screenshots with text
Text is sharp-edged, geometric. AI sometimes blurs character edges or hallucinates extra detail in serifs. Bicubic or Lanczos preserve text fidelity better.
Exception: AI does well on photo of a sign or document where the text is mixed with photographic content.
Charts, diagrams, UI mockups
Flat colors and geometric shapes. AI sometimes adds texture noise. Lanczos preserves the clean look.
Already-high-res images
If the source is already 2000+ px on the long edge, AI upscaling brings diminishing returns. The model fills in detail at the scale it was trained on (typically 4×), but doubling a 2000 px image to 4000 px rarely produces visible benefit for typical screen viewing.
The Speed Difference
On a 1000 × 1000 source upscaling 4×:
| Method | Time | Output |
|---|---|---|
| Bicubic | ~50 ms | 4000 × 4000 smooth |
| Lanczos | ~150 ms | 4000 × 4000 sharp |
| AI (Swin2SR) | 3-15 sec | 4000 × 4000 with invented detail |
For batch jobs (50+ images), the speed difference matters. AI on 100 images is 5-25 minutes; Lanczos is under 30 seconds.
How paste-to-download Picks for You
/upscale has three modes:
- Auto (default): analyzes the image, picks AI for photographic content or Lanczos for line art / text
- Photo: forces AI (Swin2SR)
- Design & text: forces Lanczos
The auto detector looks at edge characteristics, color count, and contrast distribution. If the image has many sharp edges and few colors, it's classified as design/text. Otherwise it goes through AI.
You can always override. If auto picks AI on a logo and the result looks wrong, switch to design mode and re-run.
Practical Decision Tree
Q: Is the image a photograph (faces, landscapes, products, animals)?
- Yes → AI upscale
- No → Lanczos
Q: Is the image already 2000+ px on the long edge?
- Yes → reconsider whether you actually need to upscale
- No → proceed with chosen method
Q: Does the source have lots of text or hard geometric edges?
- Yes → Lanczos, or convert text/vector via /vectorize
- No → AI is probably the right call
Q: Is the source very blurry or very small (under 500 × 500)?
- Yes → AI is your only hope of getting credible detail
- No → bicubic might actually suffice for modest enlargement
Q: Time pressure?
- Batch of 50+ → Lanczos (5 minutes total)
- Few images, quality matters most → AI
Real Examples From the /upscale Tool
Old family photo, 600 × 450
A scanned 35mm print from the 1990s, slightly blurry. AI upscale to 2400 × 1800 produces convincing skin detail, hair, fabric. Lanczos produces an enlarged blurry version of the source. AI wins.
Logo PNG, 300 × 300 with text
Company logo with company name in serif font. AI introduces subtle wavy distortions to the serif strokes. Lanczos keeps the edges crisp. Lanczos wins (or better: vectorize and stop upscaling).
Screenshot of a chart, 800 × 600
Flat colors, gridlines, axis labels. AI adds faint texture to the white background — visible at 200% zoom. Lanczos stays clean. Lanczos wins.
Pet photo, 1200 × 800
Dog photo, dense fur, good lighting. AI upscale to 4800 × 3200 produces highly detailed fur texture, sharp eye reflections, defined whiskers. Lanczos produces a soft enlargement. AI wins decisively.
Privacy Note
Most online upscalers upload your image to a GPU cluster. For private photos (family, products, sensitive content), this is a real consideration.
/upscale runs in your browser. The Swin2SR model loads as ONNX Web Runtime, runs as WebGPU or WebAssembly. Your image never leaves your device. The first run downloads the model (one-time ~30 MB), cached for future use.
The TL;DR
- Photos → AI upscale (/upscale photo mode)
- Logos / text / line art → Lanczos (/upscale design mode)
- When in doubt → Auto mode, let the tool decide
- Don't upscale past 2× from source unless the result is for print
- For text and vectors specifically, consider /vectorize instead of any raster upscale