Guide
Image Upscaling — AI vs Lanczos, When to Pick What
A practical guide to choosing between High Quality AI upscaling and Instant Lanczos — and what 2× and 4× actually mean for output resolution.
High Quality (Swin2SR AI) vs Instant (WebGL Lanczos)
High Quality (AI) uses Swin2SR — a transformer-based super-resolution model that reconstructs detail rather than just smoothly interpolating. Better at sharp edges, text, and texture preservation. Slower (~30 seconds for a typical 2× job) and downloads a ~47 MB model on first use.
Instant (Lanczos) uses a WebGL fragment shader implementing Lanczos-2 resampling. Sharp output, runs in under 1 second, no model download. Better choice for fast iteration, batch processing, or when you don't need reconstructed detail.
Target resolutions
The upscaler enforces minimum output sizes so a small input always becomes a useful output:
- 2× → at least 2048 px on the longest edge. Tiny inputs are pre-scaled with Lanczos to ~1024 px before AI 2×, so the model output hits the 2048 px floor.
- 4× → at least 4096 px (4K) on the longest edge. Inputs under 512 px run a 2-pass cascade (Swin2SR → Lanczos 2× → Swin2SR again) to maintain quality.
You can't accidentally upscale a 100×100 PNG into a 200×200 PNG and wonder why it doesn't look 4K — the tool always reaches the target floor.
File size implications
Upscaling produces larger files in two dimensions: pixel count (linear with scale) AND noise (AI models hallucinate detail that doesn't compress well). A 2× upscale of a 500 KB JPG often produces a 3-5 MB PNG. Re-encode as WebP at 0.85 quality to cut that roughly in half without visible degradation.
When to use which
- Product photography — AI. The detail reconstruction matters.
- Old photos / archival — AI. Worth the wait for legibility.
- Bulk thumbnails — Instant. AI overhead × 50 images = no fun.
- Screenshots of UI — Instant. UI has no "detail" to reconstruct; Lanczos is sharper anyway.
- Pixel art — Instant in nearest-neighbour mode (not AI — it would smear the pixels).