Upscaling glossary

Twenty terms you’ll meet in upscaler settings, reviews and papers, each in a sentence or two.

Artefact

Any visible error introduced by processing: halos, ringing, blocks, seams, waxy skin, repeated texture.

Bicubic interpolation

A classic resizing method that estimates new pixels from a 4 × 4 neighbourhood. Smooth, fast, adds no detail.

Creative upscaling

Upscaling with a generative model (usually diffusion) that adds plausible new detail. See creative vs faithful.

Degradation model

The set of damages (blur, noise, compression, downscaling) a model is trained to undo.

Denoise strength

In diffusion upscalers, how much noise is added before regeneration. Higher means more change.

DPI / PPI

Dots per inch (printer) vs pixels per inch (image on paper). Often used interchangeably. See the print guide.

Face restoration

A specialised model (e.g. GFPGAN, CodeFormer) that rebuilds facial detail. Powerful, and risky for identity.

Faithful upscaling

Upscaling that reconstructs what the pixels imply and avoids adding new content.

GAN

Generative adversarial network: a generator trained against a discriminator. Gives sharp, realistic texture.

Hallucination

Detail that the model generated and that wasn’t in the source. Desired in AI art, a problem in evidence.

Lanczos

A sharper classic resampling filter. Can cause slight ringing near edges.

LPIPS

A perceptual similarity metric based on deep network features (Zhang et al., 2018).

Megapixel (MP)

One million pixels. 4000 × 3000 px = 12 MP. Many services price or limit by output megapixels.

Nearest-neighbour

Copies the closest pixel. Blocky, but correct for pixel art at integer scales.

PSNR

Peak signal-to-noise ratio: pixel-level fidelity to a reference. Tends to reward blur.

Resemblance / structure

Controls in creative upscalers that keep the output close to the input’s shapes.

Ringing

Faint light/dark ripples beside sharp edges, typical of over-sharpening and some filters.

Scale factor

Multiplier per side: 4× of 1000 px wide is 4000 px wide (16× the pixels).

Super-resolution

Producing a higher-resolution image from a lower-resolution one. See explainer.

Tiling

Processing a large image in overlapping tiles to fit in memory. Too little overlap causes seams.