In short: This glossary explains the AI terms that come up in briefs, quotes and production meetings. Each entry says what the term means and, where it matters on a shoot, what to ask. The entries are in alphabetical order, and each one is short enough to read in a minute.
Agent
An AI agent is a system that works through a series of steps toward a goal, such as answering enquiries, booking a slot or sorting files, often by calling other tools. A brief for an agent should say what it may do without approval and what it must never do.
Content Credentials
Content Credentials are a signed record of where a file came from and what was done to it, using the C2PA specification. They can note AI use and edits, and they travel with the file when the tools support them. The labelling AI content guide explains what they can and cannot show.
Context window
The context window is how much text, and in some tools how many images, a model can read at one time. Anything outside it is not seen when the model answers. Long briefs and long scripts may need to be split for this reason.
Deepfake
A deepfake is a realistic picture, video or audio of a real person doing or saying something they did not. The word is used loosely, so ask for the specific technique when a brief or a quote uses it. The voice and likeness consent checklist sets out what to agree first.
Diffusion model
A diffusion model makes an image or a video by starting from noise and refining it step by step until it matches the prompt. Many image and video generators use this approach. Two runs of the same prompt give different results, which is why the seed and the references matter.
Fine-tuning
Fine-tuning trains an existing model further on a smaller set of examples, so it follows a house style or a subject more closely. It needs the examples and the right to use them. Ask who holds the trained model afterwards.
Generative AI
Generative AI is a family of models that make new text, images, audio or video from an input, rather than only analysing existing data. On a production it covers everything from a draft script to a generated background. A quote should name the tool and the stage where it is used.
Hallucination
A hallucination is an output that sounds confident and is wrong, such as an invented fact, quote or product feature. Language models produce them, and generated images can contain them too, such as text that looks like writing but says nothing. A person should check every factual claim before publication.
Image-to-video
Image-to-video makes a short moving clip from a still picture by animating the camera or the subject. It can suit product motion, but the movement can change the shape of the item. Check each frame at full size.
Inpainting
Inpainting replaces a chosen part of an image, such as a sign or a background, while the rest of the picture stays fixed. It is used for clean-up and small changes. The mask tells the tool which area to change.
Large language model
A large language model, or LLM, is a model trained on large amounts of text to predict and generate language. It can write, summarise and answer questions, and it can be wrong with confidence. Many of the chat tools in use run on one.
Likeness
Likeness is how a person looks or sounds. It matters because a generated face or voice can stand for a real person even when no name is used. Get consent before using a generated likeness of anyone real.
Lip sync
Lip sync matches a mouth to speech in a video, either by re-animating a face to a new audio track or by matching a dubbed track. Results vary with the camera angle, the language and the speed of the speech. Sync errors show first on close shots, so check those at full size.
Negative prompt
A negative prompt lists what the picture or video should avoid, such as extra fingers or stray text. It works only in tools that support it, and it can be ignored. For brand work, put the must-not-appear items in the brief as well.
Outpainting
Outpainting extends a picture beyond its original edges, for example to make a square image wider. The added area is generated, so it may not match the light or the perspective of the original. Check the join between the two.
Prompt
A prompt is the instruction given to a generative model, usually in text, sometimes with an image. A good prompt names the subject, the setting, the camera, the light and what must stay the same. Keep the prompts with the project, so the work can be repeated.
Provenance
Provenance is the recorded history of a file: where it came from, who made it and what was done to it. Content Credentials are one way to record it. Provenance does not show that the content is true.
Retrieval-augmented generation
Retrieval-augmented generation, or RAG, lets a model search a set of documents before it answers, so it draws on your own sources as well as its training. It is common in assistants built for a business's own content. The answers are only as good as the documents and the search behind them.
Seed
A seed is a number that sets the starting noise for a generated picture or video. The same seed, prompt and settings can reproduce a result closely, which helps with small changes. Tools differ in how far they honour a seed.
Synthetic voice
A synthetic voice is speech generated by a model rather than recorded from a person in a booth. It can read a script without a voice actor in the room. Label it where a listener would expect to know, and check it against the script word by word.
Text-to-video
Text-to-video makes a moving clip from a written prompt with no source footage. It is quick for mood and concept work, but it can get the product, the people and the physical detail wrong. Test a short clip before commissioning a full piece.
Token
A token is a small piece of text, often part of a word, that a language model reads and writes. Tools price and limit their use by tokens, so a long brief uses more of them than a short one.
Training data
Training data is the material a model learned from. Before you rely on a tool's output for commercial work, ask what the tool was trained on and whether the makers had the right to use it.
Upscaling
Upscaling makes a small picture or video larger by adding detail with a model. It can make a file usable on a bigger screen, but the detail it adds was not in the original. Do not treat an upscaled image as evidence of what a real object looks like.
Voice clone
A voice clone is a synthetic voice made from recordings of one person. Making one needs that person's consent and a written agreement on how the voice can be used. Label the output as synthetic wherever a listener would expect to know.
Watermark
A watermark is a mark, visible or invisible, that identifies a file or its source. An invisible watermark can help a platform find a record of a file, but it can be removed or damaged. Use it alongside Content Credentials, not in place of them.
Related guides
- AI video for small businesses: what's real, what's risky
- When to use AI and when to shoot it for real
- AI in the production workflow, stage by stage
- Who is responsible for an AI-made ad?
- Voice and likeness: a consent and contract checklist
- All guides and tools
The picture at the top of this page is an illustration, not a photograph of a client or a real job.
Questions we get asked
What is the difference between a prompt and a brief?
A prompt is the instruction given to a single generation. A brief is the wider document covering the goal, the audience, the references, the approver and the delivery format.
What does hallucination mean for a product video?
A generated product can show features or details the real item does not have. Check the output against real photographs at full size before you approve it.
Which AI terms matter most in a commissioning brief?
Likeness, training data, provenance and synthetic voice. Each one raises a question about consent, rights or disclosure that the brief should answer before the work starts.
granary.digital · 01243 884993 · studio@granary.digital · The Old Grainstore, Adsdean, Chichester, West Sussex, PO18 9DN · Prices exclude VAT unless marked inc VAT