This AI video glossary explains the 40 terms you are most likely to meet when you commission, make or approve video produced with generative AI. The terms are grouped by stage, from models and prompts to editing, sound, advertising and the rules, and each definition is one to three sentences of plain English.
Definitions were checked on 20 September 2026. The words are stable. The tools behind them are not, so see the best AI video generators in 2026 for the current names.
Models and generation
- Generative AI. Software that produces new text, images, sound or video from patterns learned from very large amounts of existing material. It predicts a plausible result. It does not look anything up or film anything.
- Diffusion model. The type of model behind most current image and video generators. It is trained to remove noise from pictures, so it can start from pure random noise and refine it, step by step, into a frame that matches the prompt.
- Text-to-video. Making a clip from a written description alone. It is fast and good for exploring ideas, but the model decides almost everything about how the result looks.
- Image-to-video. Animating a still image that you supply or approve first. Because the look is fixed before anything moves, this is the method studios rely on, as we explain in how to make Instagram Reels with AI.
- Video-to-video. Changing existing footage with a prompt or a reference: a new style, a different background, another outfit. The movement of the original clip is kept.
- Seed. The number that sets the random starting point of a generation. Reusing a seed with the same prompt, settings and model version reproduces the result, which lets you change one thing at a time.
- Temporal consistency. How well a face, a product or a background stays the same from one frame to the next. Poor consistency shows up as flicker, drifting labels and features that slowly change.
- Artefact. A visible error made by the model, such as an extra finger, melting text or an object that passes through another.
Prompts and control
- Prompt. The instruction given to a model. A good video prompt reads like a line from a shot list: subject, action, setting, light, lens and camera movement.
- Negative prompt. A list of things you do not want, such as text, extra people or blur. Not every tool supports it.
- Reference image. A picture given to the model to show what a person, product or style should look like. Several references from different angles work better than one.
- Keyframe. An image that fixes what the clip must look like at a given moment. Many tools accept a first and a last frame, and some accept more: Luma’s Ray3.2, for example, takes up to 16 in one clip.
- Character consistency. Keeping the same person or mascot recognisable across many shots and videos. Our guide to consistent AI characters covers the methods.
- Camera control. Telling the model how the camera moves: push in, orbit, tilt, handheld. Some tools offer presets and others read it from the prompt.
- LoRA. Short for low-rank adaptation. It is a small add-on trained on a small set of images so that a model can reproduce a specific product, face or style, and it is mostly used with models you can run yourself.
Production and editing
- Storyboard. A sequence of frames that shows a video shot by shot before it is made.
- Look development. The stage where still images fix the style, the people and the product before any motion is generated. Changes here take minutes, which is why it is the best place to give feedback.
- Upscaling. Raising the resolution of a clip, for example from 720p to 4K, with a model that adds plausible detail. It sharpens footage but cannot repair a wrong label.
- Inpainting. Regenerating one selected area of an image or a video while leaving the rest alone. It is how a wrong hand or a stray logo gets fixed without starting the shot again.
- Compositing. Layering separately made elements into one shot, such as a real photograph of your product placed into a generated scene. It is the most reliable way to keep product videos accurate.
- Colour grade. Adjusting colour and contrast across every shot so that a film looks like one piece. Clips from different models look different, so the grade matters more than usual.
Sound and voice
- Native audio. Sound generated together with the picture by the same model, including speech, effects and ambience. Veo 3.1, Kling 3.0 and Seedance 2.5 all offer it.
- Text-to-speech (TTS). A synthetic voice reading a script. Good results still take directing through punctuation and retakes.
- Voice cloning. Building a synthetic voice from recordings of a real person. It needs that person’s written consent and a clear agreement on what the voice may be used for.
- Lip-sync. Matching the mouth movements of a person on screen to an audio track. It is what makes an avatar or a dubbed presenter look as if they are really speaking.
- AI dubbing. Translating the speech in a video into another language, often in the original speaker’s voice and with matched lips. We cover the practical side in bilingual video in English and Russian.
- Sound design. The footsteps, fabric, room tone and other effects that make pictures feel real. Generated footage often arrives silent or thin, so this step matters. See sound design for AI video.
Advertising and platforms
- Hook. The first one to three seconds of a short video, which decide whether people keep watching. We list formats worth trying in AI video ad hooks to test.
- UGC-style ad. An advert made to look like a customer’s own phone video. When the person is generated it is often called AI UGC, which we explain in AI UGC ads explained.
- AI avatar or AI presenter. A digital person who delivers a script to camera. The avatar can be invented or built from a real person with their consent. See our AI presenters and avatars service.
- 9:16. The vertical aspect ratio used by Instagram Reels, TikTok and YouTube Shorts, normally exported at 1080 by 1920 pixels.
- Safe zone. The part of a vertical frame that the app’s own interface does not cover. Captions, logos and prices belong inside it, because the username, buttons and caption text sit on top of the edges. The margins in pixels are in our reference on Reels safe zones and export settings.
- Creative testing. Running several versions of an advert, usually with different hooks, to find out which one performs. Cheap variations are a main reason brands use AI for paid social ads.
Rights and rules
- Deepfake. Video or audio that shows a real person doing or saying something they did not. Article 50 of the EU AI Act, which applies from 2 August 2026, requires people who publish deepfakes to disclose that the content is artificial.
- Likeness and consent. A person’s likeness is their recognisable face, voice or manner. Using it in a video or an avatar needs written permission that states the media, the territory and the term.
- AI label. A notice on a post or advert saying that the content was made or changed with AI. TikTok asks creators to label realistic AI-generated content, Meta shows an “AI info” label, and YouTube asks uploaders to disclose realistic altered or synthetic content. We set out when an AI label is required on each platform in a separate guide.
- Content Credentials. Tamper-evident data attached to a file that records where it came from and how it was edited. The standard comes from the C2PA, which compares it to a nutrition label for digital content. Platforms read it to apply AI labels automatically.
- Invisible watermark. A signal hidden in the pixels or the audio that marks a file as AI-generated and is designed to survive ordinary editing. Google’s SynthID, embedded in everything Veo and Gemini Omni produce, is the best-known example.
- Usage rights. The terms that say where, how long and in which media you may use a video. Paid advertising, extra countries and longer terms are often priced separately, so agree them before production. Start by stating where the video will run, which is line three of our video brief template.
- Copyright in AI output. Whether anyone owns a generated video depends on the country and on how much a human shaped it. The US Copyright Office concluded in January 2025 that material generated wholly by AI is not protected and that prompts alone are not enough. The UK’s older rule on computer-generated works is under review. Our explainer on who owns AI-generated video covers what a brand should get in writing. This is general information, not legal advice.
How to use this glossary
You do not need these words to commission a good video. They help when you read a quote, where look development, upscaling and usage rights become lines with prices, and when you give feedback, where “the label drifts in shot four” gets a faster fix than “something looks off”.
At Zero Shot, an AI video studio in London, we try to keep jargon out of our own replies. If a term in a proposal is not on this list, ask the studio to explain it. If you have a project in mind, send a short brief and we will reply in plain English with a fixed price.
Questions people ask
What is the difference between text-to-video and image-to-video?
Text-to-video creates a clip from a written prompt alone, so the model decides how everything looks. Image-to-video starts from a still image that you supply or approve, so the look is fixed before anything moves. Image-to-video is the more controllable of the two.
What does a seed do in AI video?
A seed is the number that sets the random starting point of a generation. The same seed with the same prompt, settings and model version gives the same or a very similar result, which makes it possible to change one thing at a time.
What is the difference between a deepfake and an AI avatar?
A deepfake shows a real person doing or saying something they did not, usually without permission. An AI avatar is a digital presenter, either invented or built from a real person who has agreed to it in writing. Consent is the dividing line.
Do AI videos have to be labelled?
Often, yes. TikTok, Meta and YouTube have rules on disclosing realistic AI-generated content, and the transparency duties in Article 50 of the EU AI Act apply from 2 August 2026. Check each platform's current help page. This is general information, not legal advice.