Generative AI Glossary

67 generative AI terms explained in plain language — diffusion, LLMs, prompts, seeds, and more.

ethics

AI Agent — An AI system that does not just answer — it acts: planning multi-step tasks, using tools (search, code, APIs), checking its results, and iterating until the…
AI Alignment — The challenge of making AI systems reliably pursue what humans actually intend — not just what we literally asked for. It spans everyday product behavior…
AI Bias — Systematic skews in AI outputs inherited from training data and design choices. Ask for a CEO and get mostly men in suits; ask for a beautiful person and get…
Copyright and AI — The evolving legal territory around AI and ownership: whether training on copyrighted work is fair use, and who owns AI-generated output. Current practical…
Deepfake — Synthetic media that convincingly shows a real person doing or saying something they never did — fake faces, cloned voices, fabricated events. The same…
Embedding — A numerical representation of content — text, image, or audio — as a list of numbers (a vector) that captures its meaning. Similar meanings produce nearby…
Fine-Tuning — Taking a pre-trained model and continuing its training on your own focused dataset so it specializes — in your brand's visual style, a specific character, or…
RAG (Retrieval-Augmented Generation) — Retrieval-Augmented Generation — instead of relying only on its training memory, the model first retrieves relevant documents from a knowledge base and answers…

core

API — Application Programming Interface — the doorway that lets software talk to a model over the internet. Instead of a chat window, you send a structured request…
Deep Learning — A branch of machine learning that uses neural networks with many stacked layers to learn extremely complex patterns. Each layer transforms the data a little —…
Generative AI — AI that creates new content — images, video, music, text — instead of just analyzing existing data. Generative models learn patterns from enormous datasets,…
Inference — The moment a trained model actually runs — taking your input and producing an output. Training happens once, at enormous cost; inference happens every single…
LLM (Large Language Model) — Large Language Model — a transformer trained on huge amounts of text to predict the next token, which turns out to be enough to write, summarize, translate,…
Machine Learning — The field of building software that learns from data instead of following hand-written rules. A machine learning model is shown millions of examples, spots…
Multimodal — A model that understands or generates more than one type of data — text, images, audio, video — in a single system. Gemini can look at an image and describe…
Neural Network — A computing system loosely inspired by the brain: layers of simple units (neurons) connected by adjustable weights. During training, the network tweaks…
Open Weights — A model whose trained parameters are publicly downloadable, so anyone can run, study, or fine-tune it on their own hardware — unlike closed models accessible…
Parameters — The internal numbers (weights) a neural network adjusts during training — its learned knowledge, stored as billions of values. Model size is measured in…
Training Data — The collection of examples a model learns from — billions of images, texts, video clips, or songs, usually scraped or licensed at internet scale. Training data…
Transformer — The neural network architecture behind virtually all modern AI, introduced by Google researchers in the 2017 paper Attention Is All You Need. Its key trick,…

image

Aspect Ratio — The proportional relationship between an image's width and height, written like 16:9 or 1:1. Choosing it is a creative decision, not an afterthought: 1:1 for…
CFG Scale — Classifier-Free Guidance scale — a slider controlling how strictly a diffusion model follows your prompt versus doing its own thing. Low values (1-4) give the…
Character Consistency — Keeping the same character recognizable across many generated images or video shots — same face, hair, outfit, proportions. It is one of GenAI's hardest…
ControlNet — An add-on for diffusion models that gives you structural control: it locks composition to a guide input — a pose skeleton, depth map, edge sketch, or scribble…
Diffusion Model — The dominant architecture for AI image and video generation. It learns by adding noise to training images until they are pure static, then learning to reverse…
GAN — Generative Adversarial Network — an earlier generative architecture where two networks duel: a generator creates fakes and a discriminator tries to catch them,…
Image-to-Image — Generating a new image using an existing image as the starting point instead of pure noise. The model keeps the overall structure and composition while…
Inpainting — Regenerating only a selected (masked) region of an image while leaving the rest untouched. Paint over the area you want changed, describe the replacement, and…
Latent Space — A compressed, abstract map where a model represents concepts as coordinates — similar things sit close together, so kitten lives near cat and far from…
LoRA — Low-Rank Adaptation — a small add-on file that teaches an existing model a new style, character, or concept without retraining the whole thing. Instead of…
Negative Prompt — A second prompt listing what you do NOT want in the image — the model actively steers away from it. Classic entries: blurry, extra fingers, watermark, text,…
Outpainting — Extending an image beyond its original borders — the model imagines what exists outside the frame and paints it in, matching style, lighting, and perspective.…
Reference Image — An image you provide alongside a prompt to guide generation — for a face, a product, a color palette, or an overall style. It anchors the model to something…
Resolution — The pixel dimensions of an image or video, like 1024x1024 or 1920x1080 — more pixels means more detail and larger usable sizes. Generation models have native…
Sampling Steps — How many denoising iterations a diffusion model runs to turn random noise into your image. More steps means more refinement — up to a point: quality typically…
Seed — The number that initializes the random noise a diffusion model starts from. Same prompt, same settings, same seed = a pixel-identical image; change the seed…
Style Transfer — Applying the visual style of one image (or a named aesthetic) to different content — your photo redrawn as a watercolor, an anime frame, or a vintage film…
Text-to-Image — Generating an image directly from a written description. You type a prompt, the model (usually a diffusion model) turns it into pixels in seconds. It is the…
Upscaling — Increasing an image's resolution using AI that invents plausible detail — sharpening edges, refining textures, reconstructing faces — rather than just…
Watermark — A visible or invisible marker embedded in media to signal origin or ownership. In AI, invisible watermarks like Google's SynthID tag generated images and video…

video

B-Roll — Supplementary footage that supports your main content — the coffee close-up in a cafe review, the city timelapse under narration, the hands-typing shot in a…
Camera Movement — The language of how a shot moves: pan (rotate horizontally), tilt (vertically), dolly or push-in (move toward or away), tracking (follow a subject), crane…
Frame Rate (FPS) — How many individual frames a video shows per second (fps). 24fps is the cinematic standard with natural motion blur; 30fps suits web and social; 60fps looks…
Image-to-Video — Animating a still image into a video clip — the image becomes the first frame, and the model imagines the motion that follows, guided by your prompt. It is the…
Keyframe — A frame that defines a key moment in an animation — traditionally, lead animators drew keyframes and assistants filled the in-between frames. AI video borrows…
Lip Sync — Matching a character's mouth movements to spoken audio so they appear to actually say the words. AI lip-sync models analyze speech phonemes and animate the…
Motion Consistency — How stable objects, characters, and physics stay across a video's frames — the difference between smooth footage and AI weirdness like flickering textures,…
Render — The process of computing final video (or image) output from its sources — burning your edit, effects, and layers into a finished file. In AI pipelines you…
Storyboard — A shot-by-shot visual plan of a video — sketches or generated frames with notes on action, camera, and dialogue for each scene. In AI filmmaking, storyboarding…
Text-to-Video — Generating a video clip directly from a written description — the moving-picture sibling of text-to-image. Models like Sora 2 and Veo 3.1 produce clips…

audio

BPM (Beats Per Minute) — Beats per minute — the tempo of a piece of music. It anchors a track's energy: 60-80 BPM feels calm and cinematic, 90-110 sits in relaxed pop and hip-hop…
Sound Design — Crafting all the non-music, non-dialogue audio in a piece: ambience, effects, whooshes, footsteps, UI clicks, the low rumble that makes a scene feel real.…
Stem — An isolated layer of a song — vocals, drums, bass, or instruments — separated from the full mix. Stems are how producers remix and rebalance music: mute the…
Text-to-Music — Generating complete music — melody, instruments, vocals, structure — from a text description. Models like Suno turn 'dreamy synthwave with female vocals about…
Text-to-Speech (TTS) — Converting written text into spoken audio with an AI voice. Modern TTS is nearly indistinguishable from human speech, with controllable emotion, pacing, and…
Voice Cloning — Recreating a specific person's voice from sample recordings — sometimes from just seconds of audio — so new speech can be generated in that voice. Legitimate…

prompting

Chain of Thought — A prompting technique where the model reasons step by step before answering, instead of jumping to a conclusion. Adding 'think through this step by step' — or…
Context Window — The maximum amount of text (measured in tokens) a model can consider at once — its working memory. Everything counts against it: your instructions, examples,…
Few-Shot — Prompting with a handful of examples that demonstrate the pattern you want, before asking for a new one. Instead of describing your ideal product description…
Hallucination — When an AI model confidently states something false or generates details that do not exist — invented facts, fake citations, impossible object physics, or…
Iteration — The core loop of AI creation: generate, evaluate, adjust one thing, regenerate. Nobody — not even experts — nails a complex image or video on the first try;…
Prompt — The instruction you give an AI model — the text (and sometimes images) describing what you want. In creative AI, the prompt is your primary tool: it carries…
Prompt Engineering — The craft of writing and refining prompts to get reliably great results from AI models. It covers structure (subject, then details, then style, then technical…
System Prompt — A hidden instruction layer that defines an AI's behavior before any user input arrives — its role, rules, tone, and boundaries. When a chatbot acts like a…
Temperature — A setting (usually 0-2) that controls how random a language model's word choices are. Low temperature (0-0.3) makes output focused and deterministic — ideal…
Token — The chunk of text a language model actually reads and writes — roughly three quarters of an English word on average; generating might split into gener + ating.…
Zero-Shot — Asking a model to perform a task with no examples — just the instruction. 'Write a haiku about GPUs' is zero-shot: the model relies purely on what it learned…