What “upscaling” actually means
Upscaling = adding pixels to an image so it can render at a larger physical size. Classic algorithms like bicubic and Lanczos blend existing pixels mathematically. AI upscalers go further: they were trained on millions of low-res / high-res pairs, so they can predict what high-res detail probably looked like in the original scene.
How AI upscalers work
- The model splits the input into overlapping tiles.
- Each tile passes through a deep convolutional network (e.g. ESRGAN, Real-ESRGAN, SRCNN).
- The network outputs a higher-resolution version of each tile, drawing on patterns it learned during training.
- Tiles are seamlessly stitched back together.
Strengths
- Crisp edges, even at 4×–8× enlargements.
- Restores plausible texture in fabric, hair, foliage.
- Cleans up JPEG artefacts as a side benefit.
Trade-offs to watch
- Hallucinated detail. The AI invents what could plausibly have been there — not what was actually photographed. Faces, text, and product shots often shift subtly.
- Smooth-skin effect. Pores and fine wrinkles get over-smoothed.
- Style bias. Models trained on stock photography over-sharpen artistic textures.
When to use AI vs classic up-scaling
- AI — small source, large output, no fidelity requirement (illustrations, web art, social).
- Bicubic / Lanczos — small enlargements (≤2×), evidence-grade fidelity, faces or branded content.
- Don’t up-scale at all — if the original is sharp and you just need print-density metadata, use the DPI Converter instead.
Print workflow that combines both
- Determine the required pixel dimensions with our Print Size Calculator.
- Up-scale the image to or slightly above that target with Upscale Image for Print.
- Inspect at 100% zoom; reject if hallucinated detail is unacceptable.
- Set the DPI metadata with the DPI Converter.