Image-to-video, lip sync, LoRA, Seed, C2PA... When you start with the tools d AI video, image or audio, the vocabulary arrives before the results. The interfaces are often in English, the tutorials too, and the same word changes the meaning of a tool The other. This glossary brings together the 60 terms most commonly found in these tools, with a short definition, a concrete example and, when useful, the English term you will see in the menus.
📌 Essentials
- 60 terms classified in 6 families : basics, prompts and settings, video, image, audio, then avatars and straight.
- Text to video part of a text; image-to-video from an animated image. The second gives more control over the rendering.
- Lip sync = sync the lips of a face with an audio track. It is the brick of dubbing AI And avatars.
- C2PA and invisible watermarks serve to prove that a content has been generated by AI. Since 2 August 2026, the AI Act requires the labelling of deepfakes.
- Definitions are written to be read alone: you can jump directly to the term you are interested in.
Contents: L AI in simple terms • Basics • Prompts and settings • Video AI • Image AI • Audio and Voice • Avatars, deepfakes and rights • FAQ
What is it?artificial intelligence In simple terms?
L artificial intelligence, it is a program that learns from examples instead of following handwritten rules. We don't explain to him what a cat is: we show him millions of pictures of cats, and he ends up spotting alone what distinguishes them. L AI generative goes a step further. She doesn't just recognize a cat, she makes a new one, in image, video or sound, from a simple description.
All tools on this site (generators of videos, voices, images) are based on this principle. And that's why they share a common vocabulary. Once these words are understood, we pass a tool the other much faster. To go further on the overall operation, our file AI video unfolds the technical pipeline from end to end.
What is the vocabulary of the artificial intelligence ? The 14 basic terms
These concepts come back everywhere, in advertisements of new models as well as in the conditions of use. No need to master the math behind: knowing what they mean is enough.
1. Artificial intelligence (AI) : a set of techniques that allow a machine to perform tasks that are associated with human intelligence (recognize an image, understand a sentence, produce a text). In France, the term is defined in the Official Journal in the vocabulary list published by the French Enrichment Commission.
2. AI general : branch AI creating new content (text, image, video, sound, code) from a set. ChatGPT, Midjourney, Kling or ElevenLabs are AI general.
3. Machine learning : method where the program improves its results by analyzing data, without coding each rule. It's the base of almost all the AI current.
4. Deep learning : machine learning form that piles many layers of artificial neurons. It was she who made possible reliable voice recognition and the generation of realistic images.
5. Network of neurons : calculation structure inspired by the brain. Thousands to billions of small units connected between them transmit values and adjust their connections during learning.
6. Model : the « brain » training that produces results. When a tool Announcement « Kling 2.6 » or « Veo 3 »he talks about the version of his model. The same site can offer several models of choice.
7. Training data (dataset) : the mass of examples (images, videos, recordings, texts) on which the model has learned. Their origin raises questions of copyright still debated in the courts.
8. Training : the long and expensive phase, where the model adjusts its parameters by comparing its productions with the examples. It is done once, at the publisher, before the posting.
9. Inference The moment the already trained model works for you. Every generation of video is an inference, and it is it that credits make you pay.
10. Parameters : internal settings of the model, learned during training. We're talking about models with billions of parameters. More is not always better: architecture and data quality count as much.
11. LLM (large language model) : model driven on huge amounts of text to understand and produce language. GPT, Claude, Gemini or Mistral are LLM. In tools video, they are often used to rewrite your prompt or generate a script.
12. Multimodal A model that deals with several types of content at the same time, such as text and input images, video and output sound.
13. Dissemination model : a generation technique which starts with a random noise (a sort of TV snow) and « damaged » step by step until you get an image or video consistent with the prompt. Most generators current images and videos use a variant.
14. Open model (open source, open weights) Model whose weights are published and downloadable, such as Stable Diffusion, Stream in some versions or Wan. We can run it on his own machine. « Open weights » is more precise than« open source », as training code and data are not always provided.
Prompts, credits and settings: 8 terms to drive tools
It's the vocabulary of interfaces. We meet him in every generator, sometimes with slightly different names for the same setting.
15. Prompt (invited) : the written instructions given to the AI. For a video, a good quick describes the subject, action, decor, light and camera movement. « A cat » gives a random result; « a red cat jumps on a window edge at sunset, close-up, slow camera » gives an intention.
16. Prompt negative : list of what we don't want to see (« blur, text, deformed hands »). All tools This is not proposed, and recent models need it less.
17. Seed (grain) : number which sets the random starting point of a generation. Even prompt + even seen + same model = very close result. Practical to redo a variant without losing everything.
18. Guidance (CFG scale, « creativity ») : cursor that sets how fast the model should stick to the prompt. Too high, the image becomes rigid or saturated; too low, the model takes liberties.
19. Appropriations : the internal currency of the tools d AI. Each generation uses it according to duration, resolution, quality mode and options (e.g. audio). Free plans give a small stock, often renewed every day or month.
20. Fine tuning : slightly retrain an existing model on a small set of specific data, so that it adopts a precise style, product or face.
21. Hallucination : when AI produced with assurance something false or inconsistent. In text, a invented date; video, a hand with six fingers or an object that crosses a wall.
22. Wait queue (tail) : the delay before your generation starts. On free plans generators video, it can range from a few minutes to several hours during rush hours, and subscribers pass in front.
💡 Tip: Note the exact seed and prompt of each successful generation in a simple text file. On the day you have to decline the video in vertical format or change a detail, you go back from there instead of burning your credits to find the same rendering.
AI Video: 16 terms to understand the generators
The heart of the subject. These terms appear in all the sheets of new models and in the menus of generators.
23. Text-to-video (text to video) : generate a video only from a written prompt. The model invents everything: characters, decor, framing.
24. Image-to-video (image to video) Animating an existing image (photo, illustration, generated image). The model retains composition and adds movement. This is the method that most creators use to keep the rendering control.
25. Video to video : transform an existing video, for example changing the style (animated drawing, painting) or replacing an element, while keeping the original movements.
26. Start and end image (start frame, end frame, keyframes) The model invents the transition between the two. Recent models accept several key images to guide an entire plan.
27. Motion control : to reproduce the movement of a reference video (a dance, a gesture) on another character, e.g. that of a photo. Kling AI popularized this function.
28. Motion brush : an area of the image is painted and the path to be followed is drawn. Useful for moving a single element (hair, car) without touching the rest.
29. Camera control : settings that require a virtual camera movement (travelling, panoramic, zoom, orbit), in addition to or in place of the description in the prompt.
30. Multi-shot (multiple planes) : ability to generate at once a sequence of several planes that sequence (wide plane, then close-up), with the same characters.
31. Time coherence The fact that objects, textures and light remain stable from one image to the next. When it is missing, we see glitters or forms that « melt ».
32. Character consistency : keep the same face, the same outfit and the same silhouette of one plane or video to another. It's the number one challenge to tell a story. with AI.
33. Images per second (fps) : number of images displayed each second. The cinema shoots at 24 fps, most web videos at 24, 25 or 30 fps.
34. Resolution (720p, 1080p, 4K) : the size of the image in pixels. 1080p corresponds to 1920 x 1080 pixels in horizontal format, 4K at 3840 x 2160. Free plans are often limited in 720p.
35. Aspect ratio (format) : ratio between width and height. 16:9 for YouTube, 9:16 for TikTok, Reels and Shorts, 1:1 for some posts.
36. Extension of video (extend) : extend a clip already generated by a few seconds by leaving its last image, to exceed the maximum duration of one generation.
37. Native Audio : generation of sound (dialogues, noise, atmosphere) at the same time as the image, by the same model, instead of adding it after the fact. Veo 3 and Kling 2.6 have generalized this approach since 2025.
38. Upscaling : increase the resolution of an image or video by reconstructing the details by AI, for example from 720p to 4K. The result does not invent a real definition, it guesses it.
What difference between text-to-video and image-to-video?
In text-to-video, the model starts from scratch: it alone decides what the character, the decor and the light look like. It's quick to explore ideas, but the result varies a lot from one try to another. In an image-to-video, you first fix the image (a product photo, a portrait, a generated and retouched illustration), then the model anime. You control the composition, the model manages the movement. In practice, most creators who publish AI videos We work in image-to-video: we generate or choose the perfect image, then we create it. We detail the method in our guide image-to-video.
Image AI : 7 terms to know
39. Text-to-image : create an image from a prompt. Midjourney, GPT Image, Flux, Nano Banana or Imagen work like this.
40. Image-to-image : from an existing image and transform it according to a prompt (change the style, lighting, season) keeping its structure.
41. Inpainting : change only a selected area of an image (remove a passerby, change a jacket color); I AI reconstruct the rest so that the retouching is based in the image.
42. Outpainting : to enlarge an image beyond its edges by inventing what is around. Useful for passing a square visual in 16:9 format.
43. LoRA : small adjustment module (a few tens to hundreds of megabytes) that is grafted to a d-image or video model to teach it a style, a character or a product, without retraining the whole model.
44. Style reference (style reference) : image provided to l tool for him to take up his aesthetics (colors, grain, type of illustration) without copying its contents.
45. Removal (background removal) : automatically isolate a subject from its background. Today it is one of the tasks that AI Do the best, in one click.
Audio and voice: 8 terms for voiceover, dubbing and music
46. Voice synthesis (TTS, text-to-speech) : turn a written text into a spoken voice. The best French voices of 2026, like those ofElevenLabsmanage intonation and connections in a very natural way. Comparison in our article AI text to speech.
47. Voice cloning (voice cloning) : create a synthetic copy of a real voice from an audio sample, from a few seconds to several hours depending on the desired quality. To be used only with one's own voice or with the written consent of the person.
48. Speech-to-speech : replace the stamp of a recording with another voice by keeping the tone, rhythm and emotions of the original take.
49. Transcript (speech-to-text) Convert a word to written text. This is the basis of automatic subtitles and video editing by the text.
50. Lip sync (lab synchronization) : to match the movement of the lips of a face with an audio track, whether real or generated.
51. Dubbing AI : to translate a video into another language by keeping the voice of the speaker (by cloning) and, often, by recaling his lips to the new language. Our guide for translate video with AI compares tools.
52. Separation of tracks (stems) : isolating the voice, battery, bass or other instruments of an already mixed piece. Isolated tracks are called stems.
53. Generative music : music composed by a AI from a prompt (genus, atmosphere, duration), with or without words. The issue of user fees depends on the licence of each tool.
What is the lip sync in AI ?
Lip sync AI Analyzes the sounds of an audio track (the phonemes) and distorts the area of the mouth of a video or photo so that it articulates them. It is used in three cases: to make speak an avatar or a photo, to double a video in another language without a gap between voice and lips, and to correct a missed sentence without turning the scene. Quality is judged on labial consonants (« p », « b », « m ») : if the lips do not close, the illusion falls. It is also the technical brick of the deepfakes, hence the detailed rules of transparency below.
Avatars, deepfakes and rights: 7 essential terms
These words are not technical, but ignoring them can be expensive. They concern the right to image, copyright and new European obligations.
54. Avatar AI : Digital character that presents face camera text that you provide. It can be chosen from a library or created from your own video. Synthesia is one of the tools the most used in company for training.
55. Talking photo (talking photo) : a single animated facial photo for it to pronounce a text, thanks to the lip sync. See our guide for Animate a photo with AI.
56. Face Swap (face exchange) : replace one face with another in a photo or video. Often playful use, but framed as soon as it is a real person. Our folder AI swap face details the tools and limits.
57. Deepfake (hyperstructure) : image, audio or video content generated or modified by AI that makes you think that a real person has said or done something. « Hyperstructure » is the term used in the French version of the European Regulation on the AI ; the official recommendation published in the Official Journal in 2020 is « videotox » (or « video infox »). The legal framework in France is summarized in our article on deepfakes and the law.
58. C2PA (Credentials) : open standard worn by the Coalition for Content Provenance and Authenticity (Adobe, Microsoft, Google, OpenAI and others). It adds to the signed metadata file that indicates how it was created or modified, including by a AI.
59. Invisible watermark : mark hidden in pixels or sound, invisible to the eye, which detects that a content comes from a tool d AI. SynthID, by Google DeepMind, is the best known example. Unlike C2PA metadata, it is partly resistant to recoding and recoding.
60. AI Act and Commercial Licence : the AI Act is the European Regulation on AI. Article 50, which has been applicable since 2 August 2026, requires the reporting of deepfakes, as well as texts generated by AI published to inform the public on matters of public interest. The commercial licence is the right granted by tool According to your subscription, use what you generate in a paid context (pub, monetized channel, customer). Free plans often exclude.
⚠️ Attention: the fact that tool let you generate a video doesn't mean you can broadcast it. Check two things before publication: the license of your plan (commercial use or not) and the consent of any recognizable real person (face or voice).
The table glossary: the 12 most popular terms
For those who just want the quick equivalent, here are the most sought-after terms, with their translation and the place where they are crossed.
| Term | In French | In one sentence | Where we meet him |
|---|---|---|---|
| Prompt | Invite, order | The text that describes what we want | All generators |
| Text to video | Text to video | Video generated from single text | Kling, Veo, Seedance, PixVerse |
| Image-to-video | Image to video | An Animated Fixed Image | Same tools, dedicated tab |
| Seed | Seed | Number that makes a result reproducible | Advanced settings |
| Upscaling | Scale | Increase resolution by AI | Video and photo editors |
| Inpainting | Targeted retouch | Edit an image area | Generators images |
| LoRA | Adjustment module | Learning a style or character to the model | Open models, tools pro |
| TTS | Voice synthesis | A text read by a synthesis voice | Tools voice, video editors |
| Voice clone | Voice cloning | Synthetic copy of a real voice | Tools voice, dubbing |
| Lip sync | Lab synchronization | Lips tied to the voice | Avatars, dubbing, talking pictures |
| Deepfake | Hyperstructure | False credible content of a real person | News, legal framework |
| C2PA | Content certificates | Metadata that trace the origin of a file | Adobe, cameras, networks |
For more technical definitions, on the automatic learning side, glossary of Google on AI general goes much further, and the CNIL publishes its own glossary focus on personal data. Ours remains voluntarily practical: it covers what we encounter using the tools Not what we're studying to build them.
🎯 Where to start
If you only had to remember five words to start: quick, image-to-video, credits, lip sync and commercial license. With them, you can describe what you want, choose the right method, estimate what it costs, understand how you make a face speak and know if you have the right to publish the result. The rest comes with practice.
FAQ: vocabulary AI video, image and audio
What a difference between AI and AI Generative?
L AI In a broad sense, all the techniques that allow a machine to learn from data, for example to recognize a face or filter spam. L AI generative is the part that produces new content: text, image, video, voice or music.
What does that mean? « model » in a tool d AI Video?
The model is the driven engine that makes the video. The same site can offer several (e.g. different versions of Kling), with different qualities, durations and credit costs.
How to know if a video was generated by AI ?
Look for a mention or label on the platform, check the C2PA metadata (Content Credentials) when they are present, and observe the details: hands, text in the image, reflections, consistency from one plane to another. No clue alone is infallible.
What is the official French term for deepfake?
The Official Journal has recommended since 2020 « videotox » or « video infox », limited to video. « Hyperstructure », proposed in Quebec, is the term of the European regulation on AI in French and also covers image and sound; in use, it is synonymous with « deepfake ».
Do we need to speak English to use the tools d AI Video?
No. Most tools understand the prompts in French, even if some models get better results in English. Knowing the English terms of the menus, those of this glossary, is enough.
