AI Image Generation Glossary

AI image tools use a lot of jargon, and much of it comes from tools very different from IMAGEKA. This glossary defines each term in plain English and, where it matters, says whether and how it applies to IMAGEKA, which runs several open-source models behind a deliberately small set of controls and picks one automatically for each prompt.

How to use this glossary

Terms are in alphabetical order, grouped into letter ranges so you can jump around with the contents list or the letter index below. "See also" links point to related entries. Many terms (sampler, LoRA, ControlNet, inpainting) describe features of other tools or of running models yourself. They are included because you will meet them in tutorials and forums, and knowing that IMAGEKA doesn't offer them saves you looking for a setting that isn't there. For the bigger picture of how the parts fit together, read How AI image generators work.

A–C

Artifact

A visible flaw the model introduced: smeared textures, melted edges, garbled lettering, an extra limb, a pattern that breaks halfway across a surface. Artifacts are most common in small, detailed areas, where the model has the least room to work. Regenerating with a different seed or making the problem area larger in the frame usually helps more than adding words like "no artifacts". See also: hallucination, VAE.

Aspect ratio

The ratio of an image's width to its height, such as 1:1 (square), 4:3 (landscape) or 3:4 (portrait). In diffusion models it affects composition, not just shape: wide frames invite horizons and side-by-side subjects, tall frames invite full-length figures. IMAGEKA offers 1:1 (512, 768 or 1024 px), 4:3 (1024×768) and 3:4 (768×1024). For other ratios, generate the closest shape and crop. See also: resolution.

Attention (cross-attention)

The mechanism that lets each region of the image being generated "look at" the parts of the prompt relevant to it. When a colour bleeds from one object to another ("a red car next to a blue house" gives a purple car), the prompt's words have been attended to in the wrong places. Clear sentence structure reduces this, especially on models with a language-model text encoder. See also: transformer, text encoder.

Base model

A model trained from scratch on a large general dataset, as opposed to one that has been fine-tuned or merged afterwards. Some fast models are distilled from a larger base model but are released as general-purpose models in their own right. Specialist models, such as ones tuned for photorealism or anime, are usually fine-tunes of a base model. See also: fine-tuning, checkpoint.

Batch

Several images generated in one run from the same prompt, each with a different seed. In IMAGEKA, the Images setting (1, 2 or 4) is the batch size. Batches of 2 or 4 are the best way to judge whether a prompt change actually helped, because they show you the range of results instead of one random sample. See also: seed, variation.

Bias (dataset bias)

A tendency to produce some things by default because they dominated the training data: a certain age, gender or ethnicity for "a doctor", a Western kitchen for "a kitchen", a sunny day for "a street". If you want something else, say so explicitly. The model can't know what you meant by default. See also: training data.

C2PA / Content Credentials

An open technical standard from the Coalition for Content Provenance and Authenticity for attaching tamper-evident information to a file's metadata, recording how an image was made and edited. Some tools and cameras add these credentials automatically, but most social platforms and many editors strip metadata on upload. Don't rely on metadata to disclose that an image is AI-generated; label it yourself where disclosure matters. See also: watermark.

CFG (classifier-free guidance) / guidance scale

A technique that makes a model follow the prompt more strongly. At each step it compares a prediction made with your prompt to one made without it (or with the negative prompt), then exaggerates the difference. The guidance scale sets how much: on Stable Diffusion-family models, roughly 5–8 is typical, and very high values look over-saturated. Some distilled fast models run at a guidance of about 1, which effectively switches CFG off. That makes them faster, but it is also why negative prompts do little on them. IMAGEKA has no guidance control; each of its models runs at settings chosen for it, and only its fast general model largely ignores the negative prompt. See also: negative prompt, distillation.

Checkpoint

A saved file of a model's weights at a point in training. In community usage it often means any complete downloadable model, including fine-tuned ones. Swapping checkpoints changes the whole model. IMAGEKA has no manual checkpoint choice: it picks among several models automatically. See also: model routing, model weights, base model.

CLIP

Contrastive Language–Image Pre-training: a model trained on image–caption pairs to judge how well a piece of text matches a picture. Its text half is used as a text encoder in many image generators. It is good at overall style and gist but weak at word order, and it reads only 77 tokens. Many newer models pair CLIP with a larger language-model encoder such as T5. See also: T5, text encoder, token.

Conditioning

Any input that steers generation: the text prompt, and in some tools also a reference image, a pose skeleton or a depth map. In IMAGEKA the only conditioning is your text prompt (with the style preset and Enhance phrases being added text). See also: ControlNet, embedding.

ControlNet

An add-on network that conditions generation on a structural guide, such as an edge map, depth map or human pose, so the output follows a given layout exactly. It is popular in self-hosted workflows. IMAGEKA doesn't support it. Steer layout with position words ("in the lower left", "centred") and with the aspect ratio. See also: conditioning, img2img.

D–F

Denoising

The core operation of a diffusion model: estimating which part of a noisy image is noise and removing some of it. Generation is repeated denoising, starting from pure noise. Each pass is one inference step. See also: diffusion model, inference steps, noise.

Diffusion model

A type of generative model trained by adding noise to images and learning to reverse it. To create an image, it starts from random noise and removes it over a series of steps, guided by a prompt. Stable Diffusion and Flux are well-known open examples; several newer models use a closely related training method called flow matching. See also: flow matching, latent space.

Distillation

Training a smaller or faster "student" model to imitate a slower "teacher". In image generation it is mostly used to cut the number of steps. Some distilled models produce good images in a handful of steps, sometimes without classifier-free guidance. The trade-off is somewhat less fine detail and variety than a slow, non-distilled model. See also: CFG, Flux.

DiT (diffusion transformer)

A diffusion model whose denoising network is a transformer rather than the older U-Net design. Transformers scale well with size and handle long, detailed prompts well. Flux and Stable Diffusion 3 are transformer-based. See also: transformer, U-Net.

Embedding

A list of numbers that represents a piece of text (or an image) in a form a neural network can use. Words with similar meanings end up with similar embeddings. A text encoder turns your prompt into embeddings before generation starts. In Stable Diffusion communities, "embedding" also refers to textual inversion: a small file that teaches the model a new word for a concept. IMAGEKA doesn't use these. See also: text encoder, token.

Fine-tuning

Continuing to train an existing model on a smaller, focused dataset, such as one art style, product photos or a particular subject, so it specialises. Full fine-tunes produce a new checkpoint; lighter methods such as LoRA produce small add-on files. IMAGEKA doesn't let you add your own fine-tunes; to specialise the look, use the style presets and prompt wording. See also: LoRA, checkpoint.

Flow matching / rectified flow

A training method related to diffusion. Instead of learning to predict added noise, the model learns the direction that moves an image from noise towards a clean picture, and rectified flow encourages that path to be close to a straight line. Straighter paths need fewer steps, which helps fast models generate quickly. See also: diffusion model, distillation.

Flux (FLUX.1)

A family of transformer-based text-to-image models released from 2024 onwards. The family includes a fast distilled variant with open weights, alongside variants with non-commercial licences or available only through paid APIs, so check the licence of the specific variant if you run one yourself. See also: licence, DiT.

G–L

Guidance distillation

Training a model to produce, in a single prediction, the result that classifier-free guidance would otherwise need two predictions to compute. It halves the work per step. A distilled model may still accept a "guidance" number as an input, but it doesn't work the same way as classic CFG, and negative prompts lose most of their effect. See also: CFG, distillation.

Hallucination

Borrowed from language models: the generator invents plausible-looking details that don't exist or that you didn't ask for, such as fake text on a sign, an extra door, or a product feature the real product lacks. It matters most when an image is meant to depict something real. Check every detail before publishing. See also: artifact.

Hires fix

A workflow in some Stable Diffusion interfaces: generate a small image, upscale it, then run a partial second denoising pass to add detail at the larger size. It works around models that compose badly at sizes larger than they were trained on. IMAGEKA doesn't have this; generate at 1024 px for maximum detail. See also: upscaling, img2img.

Img2img (image-to-image)

Generation that starts from an existing image with some noise added, instead of from pure noise, so the result keeps the original's layout and colours. A "strength" or "denoise" setting controls how much changes. IMAGEKA is text-to-image only: you can't upload an image to transform. See also: inpainting, ControlNet.

Inference

Running a trained model to produce an output, as opposed to training it. When you press Generate, IMAGEKA sends your prompt and settings to an inference service that runs the chosen model on GPUs and returns the images. See also: inference steps.

Inference steps (sampling steps)

The number of denoising passes used to make one image. Non-distilled models typically use 20–50, and more steps mean more refinement up to a point. Distilled models are designed for only a few, and extra steps add little. IMAGEKA doesn't expose a step setting; each model runs at settings chosen for it. See also: denoising, sampler.

Inpainting

Regenerating only a masked part of an image, for example to fix a hand or swap an object, while leaving the rest untouched. IMAGEKA doesn't offer inpainting. For small fixes, use an image editor's retouching or generative-fill tools on the downloaded image. See also: outpainting, img2img.

Latent space

The compressed representation in which the model actually works. A VAE squeezes images into a much smaller grid of numbers: typically 8× smaller on each side, with 4 or 16 values per position depending on the design. Denoising happens there, and the result is decoded into pixels at the end. Working in latent space is what makes generation fast enough to run in seconds. See also: VAE, resolution.

Licence (model licence)

The legal terms under which model weights are released, which govern how the model and its outputs may be used. Licences range from permissive ones such as Apache 2.0 to responsible-use licences such as CreativeML OpenRAIL++-M and non-commercial ones. The models IMAGEKA uses are all under open licences that allow commercial use. The model licence is separate from whether an output is protected by copyright, and from trademark or likeness rights in what it depicts. See using AI images commercially. See also: open weights.

LoRA (low-rank adaptation)

A lightweight fine-tuning method that stores a small set of weight adjustments, often only tens of megabytes, to add a style, character or concept to a base model. LoRAs are loaded on top of the base model and can be mixed. IMAGEKA doesn't support LoRAs, so get styles from prompt wording and the style presets. See also: fine-tuning, style preset.

M–P

Model routing

Sending each request to whichever of several models suits it best. IMAGEKA does this automatically: it runs several open-source models (a fast general model plus specialists for photorealism, anime and painterly styles) and chooses one from your prompt and style preset. You can't pick a model manually, but the preset and words like "photograph", "anime" or "watercolour" steer the choice, and quoted text sends the request to the general model, which handles lettering best. If a specialist fails, IMAGEKA retries with the general model. See also: style preset, checkpoint.

Model weights / parameters

The billions of numbers learned during training that define what a model does. "12B parameters" means about twelve billion such numbers; current open image models range from under a billion to over ten billion parameters. More parameters generally means more capacity for detail and prompt understanding, at the cost of more computing power per image. See also: checkpoint, open weights.

Negative prompt

Text describing what the image should avoid. On models that use classifier-free guidance it replaces the "empty" prediction, so the image is pushed away from it, which works well on Stable Diffusion-family models. On distilled models running with guidance around 1, it has little effect. In IMAGEKA, the photorealism, anime and painterly models use the negative prompt (and IMAGEKA adds sensible defaults for them, such as bad hands, blurry and watermark); the fast general model largely ignores it. Either way, describing what you do want ("an empty street" rather than negating "people") is the most reliable approach. See also: CFG, prompt.

Noise

Random values, like television static, that a diffusion model starts from and gradually removes. The particular noise pattern, set by the seed, decides much of the final composition, which is why the same prompt gives different images. See also: seed, denoising.

NSFW

"Not safe for work": sexual or graphic content. IMAGEKA's prompt filter blocks requests for clearly prohibited content, and a safety checker blurs outputs it detects as explicit. Filters sometimes misfire on innocent prompts (medical or art-history subjects, for example). Rewording usually resolves this. See also: safety checker.

Open weights

A model whose trained weights are publicly downloadable, so anyone can run it on their own hardware. This is different from "open source" in the strict sense, which would also cover training code and data. All the models IMAGEKA uses are open-weight models under open licences that allow commercial use. See also: licence, model weights.

Outpainting

Extending an image beyond its original borders by generating new content that continues the scene, for example to turn a square image into a wide banner. IMAGEKA doesn't offer it. Choose the right aspect ratio before generating, or extend the image with an editor's generative tools. See also: inpainting, aspect ratio.

Prompt

The text description you give the model. On most current models, natural sentences that state the subject, setting, light and medium work better than lists of keywords, and the first part of the prompt carries the most weight. Anime-tuned models are the exception: they also respond well to short comma-separated tags. The prompt-writing guide covers structure in depth. See also: token, prompt adherence.

Prompt adherence

How faithfully a model follows everything a prompt asks for: every object, attribute, count and position. It is one of the main differences between models. Even strong models still drop details from long prompts and miscount larger numbers. See also: attention, T5.

Prompt enhancement

Automatically adding words to a prompt to improve the result. Some tools use a language model to rewrite your prompt. IMAGEKA's Enhance prompt button is simpler: it adds a few phrases based on keywords it finds (for example "detailed eyes" for portraits). Review what it adds. See also: style preset.

Prompt weighting

Syntax in some interfaces, such as (red scarf:1.4) or ((red scarf)), that boosts or reduces the influence of part of a prompt. It is a feature of particular Stable Diffusion front-ends, not of the models themselves. IMAGEKA has no weighting syntax: brackets and numbers are just extra characters to the model. To emphasise something, put it early in the prompt and describe it more specifically. See also: prompt, token.

R–S

Resolution

The pixel dimensions of the generated image. Higher resolution gives the model more latent space for detail, and models perform best near the sizes they were trained at, around one megapixel for most current models. IMAGEKA's specialist models look poor below 768 px, so 512×512 requests routed to them are generated at 768×768. IMAGEKA offers 512×512, 768×768, 1024×1024, 1024×768 and 768×1024. Use 512 for quick drafts and 1024-class sizes for final images. See also: aspect ratio, upscaling.

Safety checker

A classifier that inspects generated images and hides or blurs those it judges to be explicit. It is separate from the prompt filter, which blocks requests before generation. IMAGEKA uses both. A blurred image means the checker flagged it, sometimes wrongly, and regenerating or rewording usually gives a clean result. See also: NSFW.

Sampler

The algorithm that decides how to move from one denoising step to the next (names you may see include Euler, DPM++ and DDIM). On step-hungry models the choice affects sharpness and speed. On a distilled few-step model it matters much less. IMAGEKA doesn't expose it. See also: scheduler, inference steps.

Scheduler

The rule for how much noise is removed at each step, for example removing more in early steps and less in later ones. The terms "sampler" and "scheduler" are often used interchangeably in user interfaces. IMAGEKA uses settings chosen for each of its models and doesn't expose them. See also: sampler, noise.

Seed

The number that initialises the random noise a generation starts from. The same seed, prompt, model and settings reproduce the same image; a new seed gives a new composition. IMAGEKA uses a random seed for every image and has no seed control, so download images you like, because you can't regenerate them exactly. See also: noise, batch, variation.

Stable Diffusion

A family of open-weights latent diffusion models released by Stability AI from 2022 onwards, including SD 1.5, SDXL and SD 3. Much of the jargon in this glossary (negative prompts, LoRA, samplers, prompt weighting) grew up around it. Some habits from early versions, such as long quality-tag lists, add little on newer models. See also: Flux, U-Net.

Style preset

A saved snippet of style text added to your prompt. Each IMAGEKA preset appends a fixed phrase; Watercolor, for instance, adds "watercolor illustration, soft washes, paper texture, wet on wet technique". Photorealistic is selected by default. Presets also steer which of IMAGEKA's models is used (for example, Anime routes to the anime model). Choose Default when your prompt already names a style, so the two don't conflict. The art styles guide lists every preset. See also: prompt, LoRA.

T–Z

T5

A large language model from Google whose encoder half is used as the main text encoder in several recent image models. Because T5 was trained on text alone, it understands grammar, word order and relationships far better than CLIP, which is why full descriptive sentences work so well on models that use it and why short quoted words are often spelled correctly. See also: CLIP, text encoder.

Text encoder

The part of the system that converts your prompt into embeddings for the image model to use. Its quality limits how well the model can understand you. Many models use more than one, for example CLIP for overall gist and a larger language-model encoder such as T5 for detailed meaning. See also: embedding, T5, CLIP.

Token

The unit a text encoder reads: a whole common word, part of a rarer word, or a punctuation mark. Encoders have a maximum length (77 tokens for CLIP, a few hundred for a T5 encoder), and text beyond it is ignored. That is one reason to put the important parts of a prompt first. See also: text encoder, prompt.

Training data

The image–text pairs a model learned from. It defines what the model knows (nothing after the training cut-off), its default looks and biases, and its blind spots: things that were rare in the data come out wrong more often. Most model makers don't publish a full list of their training data. See also: bias.

Transformer

A neural network architecture built around attention, originally developed for language and now used for images too. In many newer models, both the text encoder and the main denoising network are transformers. See also: DiT, attention.

U-Net

The convolutional network design used as the denoiser in earlier diffusion models, including Stable Diffusion 1.5 and SDXL. It processes the image at several scales, shrinking it and then expanding it again, hence the "U". Many newer models replace it with a transformer, although U-Net models and their fine-tunes remain widely used. See also: DiT, Stable Diffusion.

Upscaling

Enlarging an image to more pixels. Simple resizing just makes it bigger and softer; AI upscalers invent plausible fine detail. IMAGEKA has no built-in upscaler, and its largest output is 1024 px on the long side. For print or large displays, run the downloaded image through a dedicated upscaling tool, and check faces and text afterwards, since upscalers can alter them. See also: resolution, hires fix.

VAE (variational autoencoder)

The network that compresses images into latent space and decodes finished latents back into pixels. Its quality affects fine detail, small text and colour accuracy. Newer 16-channel VAEs keep noticeably more detail than the 4-channel one in Stable Diffusion 1.5. You can't change the VAE in IMAGEKA. See also: latent space, artifact.

Variation

A new image that keeps a prompt's idea but differs in composition and detail. Some tools generate variations close to a chosen image; on IMAGEKA, every run with the same prompt is effectively a variation, because each uses a new seed. To change one aspect deliberately, edit one phrase of the prompt and generate a batch. See also: seed, batch.

Watermark

A mark identifying an image's origin. Visible watermarks are overlaid logos or text. Invisible watermarks are patterns hidden in the pixels that software can detect, and some AI services embed them. IMAGEKA doesn't add a visible watermark. Separately, models sometimes paint fake watermark-like smudges because they appeared in stock photos in the training data; describing the image as "a clean, unbranded photograph" can help. See also: C2PA / Content Credentials, training data.

Frequently asked questions