AI image editing glossary

Last reviewed 2026-09-17. Every competitor figure links to the vendor's own page.

AI image editing borrows vocabulary from computer vision, printing and web development at once, which is why tool documentation can be hard to read. These are the terms that appear most often, defined in plain language with the practical consequence noted.

Terms

Alpha channel

A fourth channel alongside red, green and blue that stores how opaque each pixel is, from fully transparent to fully solid. It is what makes a transparent background possible.

Artifact

Any visible defect introduced by processing rather than present in the original: compression blocking, halos around edges, smearing, or invented texture from an upscaler.

Batch processing

Running the same operation across many images in one job instead of one at a time. Essential for catalogue work, where per-image handling does not scale.

Bicubic interpolation

A traditional resizing method that estimates new pixels by averaging the sixteen nearest existing ones. Fast and predictable, but produces soft results when enlarging because it can only average what is already there.

Credit

A unit of processing in a usage-priced tool. One operation on one image consumes a set number of credits, so a credit allowance is a monthly processing budget.

Diffusion model

A generative model that starts from noise and removes it step by step toward a plausible image. Slower than a GAN but generally more stable, and now common in upscaling and inpainting.

Downsampling

Reducing an image's pixel dimensions. Always lossy in the sense that discarded detail cannot be recovered, though it rarely looks degraded because information is removed rather than invented.

EXIF

Metadata a camera writes into an image file: settings, timestamp and often GPS coordinates. The location data is the usual privacy concern when publishing photos.

Generation loss

Cumulative quality degradation from repeatedly saving in a lossy format. Each save re-applies compression to already-degraded data, which is why JPEGs deteriorate across rounds of editing.

GAN

Generative adversarial network. Two networks trained against each other, one generating images and one judging them. Long the standard for photo-realistic upscaling, increasingly replaced by diffusion models.

Inpainting

Filling a removed or damaged region with generated content that matches the surrounding image. The mechanism behind object removal and much of photo restoration.

Interpolation

Estimating new pixel values from existing neighbours when resizing. Distinct from AI upscaling, which predicts detail rather than averaging what is present.

Lossless compression

Reduces file size while allowing the original data to be reconstructed exactly. Used by PNG and available in WebP and AVIF.

Lossy compression

Reduces file size by permanently discarding information judged least visible. Used by JPEG and optionally by WebP and AVIF.

Mask

A binary selection marking each pixel as kept or discarded. Sufficient for hard-edged shapes, inadequate for hair or anything semi-transparent.

Matting

Estimating continuous, partial opacity for each pixel rather than a binary keep-or-discard decision. What separates a natural cutout from one that looks cut with scissors.

Megapixel

One million pixels. A 4000 by 3000 image is 12 megapixels. Output caps in image APIs are usually expressed in megapixels rather than dimensions.

Premultiplied alpha

A storage convention where colour channels are already scaled by opacity. Efficient for compositing, but produces a dark halo if software treats it as straight alpha.

Real-ESRGAN

A widely used open-source super-resolution model, notable for being trained on synthetically degraded images so it handles real-world compressed photos rather than only clean ones.

Segmentation

Classifying each pixel as belonging to a particular region, such as subject or background. The first step in background removal.

sRGB

The default colour space of the web. Narrower than Adobe RGB or Display P3, but the only one reliably interpreted correctly across browsers and platforms.

Super-resolution

Increasing an image's resolution by reconstructing plausible detail with a trained model, as opposed to averaging existing pixels. Commonly called AI upscaling.

Upscaling

Increasing pixel dimensions. Traditional upscaling interpolates and goes soft; AI upscaling predicts detail and stays sharp, at the cost of inventing rather than recovering information.

WebP

A web image format supporting both lossy and lossless compression plus an alpha channel, typically 25 to 35 percent smaller than JPEG or PNG at equivalent quality.