The AI Video Glossary 2026: Every Term, Explained
Every industry runs on vocabulary, and AI video now has more of it than most people learn in a year of casual use. Terms are the currency of this field: they’re how you brief collaborators, how you read model documentation, how you debug a render, and how you tell a vendor what rights you need. Miss the vocabulary and you can’t even ask the right question.
This glossary is organized by theme — the seven shot elements, prompt terms, model terms, workflow terms, industry terms and copyright terms — with short definitions you can quote, cite and forward. Confusable pairs get their own section at the end, because half of all prompt arguments are just two words talking past each other.
Key takeaways: - The core vocabulary is the seven shot elements: subject, motion, camera, composition, lighting, style, timing — the shared skeleton of every usable video prompt. - Model dialect is the most important 2026 concept most creators can’t define: structure travels across models, vocabulary doesn’t. - Workflow terms (extraction, baseline render, one-variable change) are what separate deliberate iteration from random re-rolling. - Copyright terms are not lawyer trivia — likeness, license and royalty-free decide what you can actually ship. - Most confusion comes from near-synonyms: prompt vs script, reference vs seed, upscale vs enhance, dialect vs style.
The seven shot elements
| Term | Definition |
|---|---|
| Subject | Who or what the shot is about: one identity, one anchor detail, stated once. Subject drift (identity changing mid-shot) is the most common render failure. |
| Motion | What moves and why: physical verbs with cause and effect, not mood adjectives. Motion is the element viewers read as “alive” or “fake.” |
| Camera | What the camera does: one move (pan, tilt, push-in, orbit), its speed and its framing. Two camera moves in one clause produce camera confusion. |
| Composition | Where things sit in the frame: shot size, subject placement, depth and negative space. Composition is the cheapest tool for visual surprise. |
| Lighting | Light source, direction and quality (hard/soft, warm/cool). One named source beats three adjectives; mixed lighting sources produce muddy color. |
| Style | The visual register: medium, texture, period, grade. An anchored style clause is what stops a render from looking like everyone else’s. |
| Timing | Duration, shot count and pacing — where the beats fall. Timing is designed, not hoped for; it schedules the hook and the payoff. |
The full methodology behind the seven elements: the video prompt engineering guide.
Prompt terms
| Term | Definition |
|---|---|
| Prompt | The instruction text that tells a generative model what to produce. In video, a usable prompt describes all seven shot elements, not just the subject. |
| Prompt engineering | The practice of designing, structuring and iterating prompts as a system rather than guessing magic words. It is decision-making expressed in text. |
| Negative prompt | A list of things the model should avoid (artifacts, styles, objects). It’s a targeted tool for recurring unwanted elements, not a default layer — see best negative prompts for AI video. |
| Seed | The number that initializes a generation, making results reproducible on the same model and settings. A seed controls randomness; it carries no content of its own. |
| Reference clip | An existing video used as structural or visual input — to study, to extract elements from, or to guide generation. What you may legally reuse from it depends on rights, not technology. |
| Identity anchor | A fixed descriptor (face, outfit, prop) that keeps a subject consistent across shots and renders. One anchor works; three competing ones cause drift. |
| Style register | The named aesthetic band a video lives in (documentary handheld, cel animation, 80s tape). A specific register outperforms generic “cinematic.” |
| Shot size | How much of the subject the frame contains: close-up, medium, wide. Stating one shot size per shot fixes most framing complaints. |
| Aspect ratio | The frame’s width-to-height proportion (16:9, 9:16, 1:1). It’s a platform decision made before rendering, not a fix after. |
| Token | The unit models read text in. Prompt length costs tokens; clarity per clause beats token volume almost everywhere. |
Model terms
| Term | Definition |
|---|---|
| Text-to-video (T2V) | Generating video directly from a text prompt. Maximum freedom, maximum unpredictability. |
| Image-to-video (I2V) | Animating a starting image with motion instructions. I2V is the standard fix for subject consistency problems. |
| Video-to-video | Transforming an existing clip (restyle, re-animate, extend) while keeping its structure. This is where input rights matter most. |
| Model dialect | The phrasing conventions a model family responds to best — action-first for Kling, sensory material for Hailuo, lean frames for Seedance. Structure transfers across models; dialect vocabulary must be rewritten — more in why the same prompt differs across Runway, Kling and Sora. |
| Fine-tuning | Training a model further on specific data to bias it toward a style, subject or motion pattern. Heavyweight compared to prompt-level control. |
| LoRA | A lightweight fine-tune adapter that teaches a narrow concept (a face, a style) without retraining the whole model. Common for character consistency. |
| Checkpoint | A saved version of model weights. “Which checkpoint” can matter as much as which model brand. |
| Inference | The act of running the model to produce output — i.e., rendering. Every inference costs time or credits, which is why iteration budgeting matters. |
Workflow terms
| Term | Definition |
|---|---|
| Extraction | Pulling the seven shot elements out of an existing video into structured prompt form — reverse engineering at the prompt level. |
| Reverse engineering (video) | Reconstructing how a video achieves its effect — its structure — so you can rebuild it with new content. Steal the skeleton, not the pixels. |
| Seven-element breakdown | The output of extraction: subject, motion, camera, composition, lighting, style, timing written as separate clauses. |
| Baseline render | The first render of a prompt, unchanged. Never fix a prompt you haven’t seen fail. |
| Iteration loop | The cycle render → diagnose → change → render again. Quality compounds through loops, not through one perfect prompt. |
| One-variable change | Altering exactly one element between renders so improvements are attributable. Two changes per render is noise, not process. |
| Continuity | Consistency of identity, wardrobe, light and physics across shots. Continuity is designed at prompt time, not repaired in post. |
| Post-mortem | A one-line note after publishing: what worked, where viewers left, what to change next. It’s the cheapest learning tool in the stack. |
Industry terms
| Term | Definition |
|---|---|
| Generative video | Video produced by models that synthesize frames rather than capture them. The category includes T2V, I2V and V2V systems. |
| Synthetic media | Any AI-generated or AI-altered media — images, video, voice, music. Disclosure rules for synthetic media are tightening platform by platform. |
| Deepfake | Synthetic video depicting a real person doing or saying something they didn’t. The term spans harmless parody to prohibited abuse; intent and consent decide the category. |
| Digital avatar | A synthetic on-screen presenter or character, often driven by text or a performance reference. Avatars built on real people require consent for that person’s likeness. |
| Virtual influencer | A fully synthetic recurring persona with its own audience. Its legal comfort comes precisely from having no real human referent. |
| Hook | The first-second element that earns the next three seconds of attention. Hooks are designed before prompts, not discovered after renders. |
| Retention | The share of viewers still watching at each moment of a video. Retention curves diagnose hooks and rhythm far better than view counts do. |
| B-roll | Supporting footage that covers cuts and illustrates narration. Generative B-roll is the fastest-growing commercial use case for AI video. |
Copyright and rights terms
| Term | Definition |
|---|---|
| Copyright | The legal right over an original fixed work. Whether AI output qualifies is jurisdiction-specific and still developing — treat it as a question per project. |
| License | Written permission to use someone else’s work under stated conditions (scope, media, territory, duration). A license is what turns “not mine” into “usable.” |
| Likeness | A person’s recognizable identity — face, voice, mannerisms. Likeness claims are enforced faster than copyright claims on most platforms. |
| Right of publicity | The right controlling commercial use of a person’s identity. Stronger in some jurisdictions than others, but present enough to plan around. |
| Fair use / fair dealing | Limited-use defenses that allow some unlicensed use for purposes like commentary and parody. Narrow, fact-specific, and never a blanket permission. |
| Transformative use | Use that adds new meaning or purpose rather than substituting for the original. The heart of the inspiration-vs-duplication line. |
| Public domain | Works free of copyright restrictions by age, waiver or law. “On the internet” and “public domain” are unrelated concepts. |
| Royalty-free | Licensed for use without per-use fees — not “free of rights.” Royalty-free still has terms, and those terms bind your client work. |
Commonly confused pairs
| Pair | The difference |
|---|---|
| Prompt vs script | A script describes what happens in a story; a prompt describes what a model renders in a shot. A script is for humans and scenes; a prompt is for one generation. |
| Reference vs seed | A reference carries content and structure to imitate or study; a seed carries only randomness state. One is about what, the other about reproducibility. |
| Upscale vs enhance | Upscaling increases pixel count (sharpness can follow); enhancing improves perceived quality — detail, color, artifacts. A 4K upscale of a broken render is a bigger broken render. |
| Dialect vs style | Style is what the video looks like (your aesthetic); dialect is how the model wants to be told (your phrasing). Style lives in the content; dialect lives in the vocabulary. |
| Consent vs credit | Consent is permission from a rights holder; credit is attribution. Credit without consent is still unauthorized use. |
| Copyright vs license | Copyright is ownership of the work; a license is permission to use it under conditions. You can make videos with zero owned copyrights and a stack of valid licenses. |
FAQ
What are the seven elements of an AI video prompt? Subject, motion, camera, composition, lighting, style and timing. They’re the standard breakdown used to extract structure from reference videos and to write prompts that leave the model as few decisions as possible.
What is a model dialect in AI video? The phrasing conventions a model family responds to — action-first, sensory, lean-frame and so on. The seven-element structure transfers between models; the dialect vocabulary has to be rewritten per model, which is why identical prompts render differently across platforms.
What is the difference between a prompt and a script? A script tells a human crew the story and the scenes; a prompt tells a model what one shot looks like. Long scripts are normal; long undifferentiated prompts are usually the problem.
Is a prompt the same thing as a reference video? No. A reference video is input material — something to study, extract from, or guide generation. A prompt is output-facing text you write (or derive from that reference) to drive the model.
Which glossary terms matter most for copyright? Likeness, license, royalty-free and transformative use. They decide whether your inputs are clean, whether real people can appear, and what you can promise a client about commercial use.
Reverse-engineer the videos you love: SocialToPrompt.
Share Article