The AI Image Generation Glossary: 28 Terms, Explained Plainly

The vocabulary around AI image generation is a mix of research jargon, marketing language and terms borrowed from photography. This is a plain-language reference, grouped so related ideas sit together. It is written for people who want to use these tools well, not for people building them.

How the models work

Model. The trained system that produces images. A model is a large collection of numerical weights that encode patterns learned from training data. Different models produce recognisably different output from identical instructions.

Training. The process of adjusting a model's weights by showing it enormous numbers of examples. Training happens once, in advance, and is enormously expensive. It is not happening when you generate an image.

Inference. What happens when you actually use the model — it takes your input and produces an output. When a service says generation takes five seconds, five seconds is the inference time.

Diffusion. The dominant technique in image generation. The model starts from random noise and removes a little of it at a time, over dozens of steps, each step moving the image closer to the requested result.

Denoising steps. How many of those noise-removal passes the model makes. More steps generally means more detail, at the cost of time. Most services choose this for you.

Latent space. The compressed mathematical representation the model actually works in. Images are encoded into it before generation and decoded back into pixels afterwards. It is the reason generation is fast: the model manipulates a small representation rather than millions of pixels directly.

Multimodal. A model that accepts more than one kind of input — text and images together, rather than only one. Google's Gemini is a multimodal model, which is what allows it to reason about your photo and a theme description at the same time.

Transformer. The neural network architecture behind modern language models, now widely used in image models too. Its strength is handling long, structured instructions.

Controlling the output

Prompt. The instruction given to the model. On a theme-based service, the prompt is written and maintained by the service; you select a theme and the full prompt is applied on the server.

Negative prompt. An instruction describing what should not appear. Used to suppress recurring artefacts.

Seed. The number that initialises the random noise a generation starts from. Same seed plus same prompt plus same model gives the same image. Different seed gives a different image — this is the mechanism behind the variation you see between runs.

Guidance scale. How strictly the model follows the prompt versus how much freedom it takes. Set too high, images look over-saturated and rigid; too low, they drift off-brief.

Aspect ratio. The proportion of width to height — 1:1 square, 3:4 portrait, 16:9 landscape. Chosen before generation, and it changes how the model composes the frame, not just how much frame there is.

Resolution. The pixel dimensions of the output. Distinct from aspect ratio, which is the shape.

Working with a person in the image

Identity preservation. Keeping a specific person recognisable through a stylistic transformation. This is the hard constraint in portrait generation and the thing that distinguishes a portrait service from a general image generator.

Face embedding. The numerical representation of a specific face that a model uses to hold onto identity through the generation process.

Reference image. The photo you supply that the model draws identity from.

Style transfer. Applying the visual character of one image to the content of another. An older technique that generation has largely absorbed.

Inpainting. Regenerating a selected region of an image while leaving the rest untouched — used for fixing a hand or removing an object.

Outpainting. Extending an image beyond its original borders, with the model inventing what lies outside the frame.

Quality and post-processing

Upscaling. Increasing the resolution of a finished image. AI upscalers invent plausible detail rather than merely interpolating, which is why an upscaled image can look sharper than the original but differ from it in fine detail.

Artefact. Any unintended visual defect — a smeared texture, an extra finger, a section of gibberish text.

Hallucination. The model producing something confidently wrong: a nonsensical object, impossible anatomy, invented text. In image models this is usually visible; in language models it usually is not, which is why the term migrated.

Photorealism. How closely output resembles an actual photograph. Distinct from image quality — a stylised illustration can be excellent and deliberately non-photorealistic.

Things you may see mentioned elsewhere

LoRA. Low-Rank Adaptation, a technique for cheaply specialising an existing model on a narrow subject or style without retraining it from scratch. Common in self-hosted tools.

Checkpoint. A saved snapshot of a model's weights. Communities around open models share checkpoints the way photographers share presets.

Fine-tuning. Further training of an existing model on a specific dataset to specialise it.

API. The programmatic interface a service uses to send requests to a model. When a platform says it is powered by a particular model's API, it means it sends requests to that model rather than running its own.

Why the vocabulary is worth knowing

Most of these terms describe knobs that consumer services deliberately hide, and hiding them is usually the right call — a theme is easier to choose than a guidance scale. But knowing what is underneath explains the behaviour you observe: why two runs differ (the seed), why heavily stylised themes lose your face (the trade-off against identity preservation), and why text comes out as nonsense (the model is generating shapes that look like letters, not writing words).