AI Glossary: 50 AI Terms Explained in Plain English

New AI tools drop all the time, and so do new terms. One week everyone's talking about context windows, the next it's agentic AI and vibe coding.

This AI glossary covers 50 terms you'll run into when working with image, video, music and text generation tools.

Terms are grouped by category rather than alphabetically, so related concepts sit together. If you're looking for a specific term, use your browser's search (Ctrl+F or Cmd+F). Each definition explains what the term means, why it matters and where you might encounter it.

Artificial intelligence (AI)

Software that performs tasks that normally require human judgment, such as understanding language, recognizing images or making decisions. In 2026, most AI tools you interact with use a subset of AI called machine learning.

Machine learning (ML)

A method where software learns patterns from data instead of following hard-coded rules. You don't program it to recognize a cat. Instead, you show it thousands of cat photos, and it figures out the pattern.

Deep learning

A type of machine learning that uses neural networks with many layers. It's the approach behind image generators, voice cloning tools and large language models (LLMs). "Deep" refers to the number of layers, not the depth of understanding.

Neural network

A system of connected nodes loosely inspired by the human brain. Each node processes a small piece of information and passes the result forward. Stacking thousands of these produces models that can generate images, translate languages or write code.

Training

The process of feeding a model data so it learns patterns. An image model is trained on millions of images. A language model is trained on text from books, websites and code. Training happens once (or periodically). After that, the model runs on what it learned.

Inference

When a trained model generates output from your input. Every time you type a prompt and get a response, that's inference. Training teaches the model. Inference uses what it learned.

Parameters

The internal values a model learned during training. A model with 70 billion parameters has 70 billion tuned values shaping its behavior. More parameters generally means more capability, but also more computation to run.

Large language model (LLM)

A neural network trained on massive amounts of text that can generate, summarize, translate and reason with language. ChatGPT, Claude, Gemini and DeepSeek are all LLMs. The "large" refers to the number of parameters.

Context window

The amount of text a model can read and remember in a single conversation, measured in tokens. A 1-million-token context window can hold roughly 750,000 words at once. Once you exceed the window, the model starts forgetting earlier content. This is the LLM context window limit that affects how much code, text or conversation history a model can work with at once.

Token

The smallest unit a language model processes. Roughly, one token equals about three-quarters of a word. When a model has a "128K context window," that's 128,000 tokens, not words. On SYNTX.AI, tokens are also the currency you use to pay for generations across all models.

Prompt

The text you give to an AI model to tell it what you want. Can be as simple as "write a poem" or as detailed as a full creative brief with style, format and constraints.

Prompt engineering

The practice of writing prompts that get consistently good results from an AI model. This means being specific about what you want, providing examples, setting constraints and structuring your request so the model understands your intent. Prompt engineering applies to text, image, video and music AI models.

System prompt

Hidden instructions that set a model's behavior before the user's conversation begins. When a chatbot always responds in a formal tone or refuses certain requests, that's the system prompt at work.

Fine-tuning

Taking a pre-trained model and training it further on a specific dataset to specialize its behavior. A general language model fine-tuned on legal documents becomes better at legal writing. Fine-tuning is how models get customized for specific industries or tasks.

Retrieval-Augmented Generation (RAG)

A technique that connects a language model to an external knowledge source so it can look up facts before answering. Instead of relying only on what it learned during training, the model searches a database, a set of documents or the web, then uses what it finds to generate a response. This reduces hallucinations and keeps answers current.

AI hallucination

When an AI model generates information that sounds plausible but is factually wrong. A language model might cite a study that doesn't exist, or an image model might add fingers to a hand that should only have five. Hallucinations happen because the model is predicting what comes next based on patterns, not checking facts.

AI agent

An AI system that can plan, take actions, use tools and adjust its approach based on results, not just answer a single question. An AI agent might browse the web, write code, run tests and open a pull request, all from one instruction. Claude Code and ChatGPT's Codex are two examples of AI agents.

Multimodal

A model that processes more than one type of input or output, such as text, images, audio, video. A multimodal large language model can read a photo and describe it, or take a text prompt and generate an image. GPT-5.6, Gemini and FLUX 3 are all multimodal.

AI-native

A product or workflow designed around AI from the start, not one that added AI as an afterthought. What does AI native mean in practice? A traditional photo editor that bolted on an AI filter isn't AI native. A platform like SYNTX.AI that's built entirely around AI model access is.

Temperature

A setting that controls how random or predictable a model's output is. Low temperature produces consistent, safe responses. High temperature produces more creative, surprising (and sometimes less accurate) output. You'll see this setting in text models and some image generators.

Agentic AI

Agentic AI is the umbrella term for AI agents, autonomous coding tools and multi-step workflows. Basically, any AI that operates with a degree of autonomy, making decisions and taking actions across multiple steps without waiting for human input at each one.

Diffusion model

The architecture behind most AI image generators. It starts with random noise and gradually removes it, guided by your prompt, until a coherent image forms. Stable Diffusion, FLUX and Nano Banana all use variations of this approach.

Text-to-image

Generating an image from a written description. You type "a cat sitting on a red chair," and the model produces that image.

Image-to-image

Using an existing image as a starting point and modifying it based on a text prompt. You upload a photo and describe a change: "make it sunset lighting" or "change the background to a beach."

Negative prompt

A prompt that tells the model what to exclude from the generated image. If your portraits keep coming out blurry, adding "blurry, out of focus, low quality" as a negative prompt pushes the model away from those results. Not every model supports negative prompts, but most open-source ones like Stable Diffusion and FLUX do.

Seed

A number that determines the starting noise pattern for a diffusion model. Using the same seed with the same prompt produces the same image. This is useful when you want to recreate a result or make small adjustments without starting from scratch.

Classifier-Free Guidance (CFG) scale

A setting that controls how strictly the model follows your prompt. Higher CFG means the output matches your words more closely but can look oversaturated or artificial. Lower CFG gives the model more creative freedom but may drift from what you asked for.

Inpainting

Editing a specific region of an image while leaving the rest untouched. You select an area, describe what should replace it, and the model fills it in. Inpainting is useful for fixing hands, removing objects or changing a background without regenerating the entire image.

Outpainting

Extending an image beyond its original borders. The model generates new content that matches the style and composition of the existing image, effectively making the canvas larger.

Upscaling

Increasing the resolution of an image using AI. Unlike traditional resizing, AI upscaling generates new detail as it enlarges. Generative upscalers like Magnific invent detail for a sharper look. Accurate upscalers like Topaz preserve the original exactly.

Low-Rank Adaptation (LoRA)

A lightweight method for customizing a model without retraining it from scratch. A LoRA trained on a specific art style or character can be loaded on top of a base model like Stable Diffusion or FLUX to shift its output in that direction, using a fraction of the compute that full fine-tuning requires.

ControlNet

A system that adds spatial control to image generation. Instead of relying on the prompt alone, ControlNet uses a reference like a depth map, a pose skeleton or an edge outline to guide the composition. It tells the model where things should be, not just what they should look like.

Aspect ratio

The proportional relationship between an image's width and height. Common ratios: 1:1 (square, Instagram), 16:9 (widescreen, YouTube), 9:16 (vertical, Reels and TikTok). Setting the right aspect ratio before generating saves you from cropping or regenerating later.

Checkpoint

A saved version of a model's weights at a specific point in training. In the Stable Diffusion ecosystem, "checkpoints" are the full model files you download and swap to change the output style. Different checkpoints produce different aesthetics from the same prompt.

Variational Autoencoder (VAE)

A component in image generation pipelines that compresses and decompresses images. It translates between the pixel space you see and the latent space the model works in. A better VAE produces sharper colors and fewer washed-out results.

Text-to-video

Generating a video clip from a written description. You describe a scene, and the model produces a moving clip with visuals and, on newer models, synchronized audio.

Image-to-video

Animating a still image. You upload a photo and describe the motion you want, and the model turns the still into a moving clip while keeping the subject consistent.

Video-to-video

Transforming existing footage using AI. You upload a video and describe changes: restyle the scene, swap a background, change the lighting. The model applies the edits while preserving the original subject.

Temporal consistency

How well a video model keeps faces, objects and lighting stable across frames. Poor temporal consistency means a character's face changes between frames or lighting flickers.

Frame rate (fps)

The number of individual frames displayed per second of video. 24fps is standard cinema. 30fps is standard web video. 60fps is smooth, high-quality motion. Higher frame rates need more compute to generate.

Keyframe

A specific frame in a video that defines a starting or ending point for a transition. Some AI video models let you set keyframes to control how a scene changes over time, giving you more direction over the output than a single text prompt allows.

Element binding

A feature in models like Kling 3.0 that locks reference images (a face, a product, a setting) to a generation so they stay consistent across shots and camera moves.

Text-to-song

Generating a complete song with vocals, lyrics and instrumentation from a text prompt. You describe a genre, mood and vocal style, and the model produces a finished track. Suno is the most widely used text-to-song tool.

Stem separation

Splitting a finished audio track into its individual components, like vocals, drums, bass, guitar and keys. This lets you remix, adjust levels or remove a vocal without access to the original recording session.

Voice cloning

Creating a digital replica of a specific voice from a short recording sample. You upload audio of someone speaking or singing, and the model generates new speech or vocals in that voice. ElevenLabs is the leading tool for AI voice generation and cloning.

TTS (Text-to-Speech)

Converting written text into spoken audio. Modern TTS models like ElevenLabs v3 produce speech with natural pacing, emotion and multilingual support, far beyond the robotic output of earlier systems.

Vibe coding

Building software by describing what you want in plain language and letting an AI model generate the code. You direct the architecture and review the output rather than writing every line yourself. AI vibe coding became mainstream in 2025 and is now used by a significant share of startups and solo developers.

Copilot

An AI coding assistant that works inside your editor, suggesting code completions, writing functions from comments and answering questions about your codebase. GitHub Copilot was the first widely adopted version. Claude Code and Gemini Code Assist are popular alternatives.

SWE-bench

A benchmark that tests whether an AI model can fix real bugs in real open-source codebases. A score of 80% means the model successfully resolved 80 out of 100 real-world software issues. It's the closest thing to a standardized test for AI coding ability.

Application Programming Interface (API)

A way for software to communicate with an AI model directly, without using a chat interface. Developers use APIs to build AI-powered products, automate workflows and integrate models into existing tools.

Open source / open weight

An AI model whose code or trained weights are publicly available for anyone to download, inspect, modify and run. FLUX Schnell, Stable Diffusion and DeepSeek are open-weight models. "Open source" technically means the code is open. "Open weight" means the trained model file is available but the training code may not be.

Latency

The time between sending a request to a model and receiving the output. Low latency (under 200 milliseconds) is critical for real-time voice agents and chatbots. For image and video generation, latency matters less since you're waiting for a finished file.

Batch processing

Generating multiple outputs from a single request. Instead of running one image at a time, batch processing lets you generate 4, 10 or 50 variations in one pass. Saves time when you need volume.

Rate limiting

A cap on how many requests you can send to an AI model in a given time period. The reason your free tier slows down or blocks you after a certain number of generations.

This AI glossary covers the terms you're most likely to encounter when working with creative AI tools. We'll keep it updated as the vocabulary changes, so bookmark this page.

If you want to explore any of the models referenced here, they're all available on SYNTX.AI, under one subscription plan. Every user gets welcome tokens on signup, so you can test the platform risk-free.