AI video, defined.
29 terms you run into when generating video with AI — what they mean, and why they matter once a film is minutes long instead of seconds. Definitions are general; where Cinovix takes a specific position, it says so.
Last updated: September 2, 2026
Glossary
The basics.
- AI video generator
- Software that produces moving footage from a text prompt or a still image instead of from a camera. Most generators output short silent clips; a smaller group produces finished video with narration, music and editing already applied.
- Text-to-video (t2v)
- Generating video directly from a written description, with no input image. It gives the model the most freedom and the least control: two t2v runs of the same prompt can return two completely different-looking subjects.
- Image-to-video (i2v)
- Generating video that starts from a supplied still image, with the prompt describing only what should move. Because the first frame is fixed, i2v is far more controllable than text-to-video, which is why consistency-critical pipelines animate images rather than generating video from text alone.
- Prompt
- The written instruction given to a generative model. In video there are usually two per shot: an image prompt describing what the frame contains, and a motion prompt describing what changes during the clip.
- Seed
- The number that initialises a model's randomness. Reusing a seed with an identical prompt reproduces an identical result, which makes seeds the standard tool for making generative output repeatable.
Glossary
Consistency.
- Anchor image
- The still frame a scene's video is generated from. Because every frame of the clip is conditioned on it, the anchor decides what the scene looks like far more strongly than the prompt does.
- Master reference image
- One canonical image of a film's subject, generated before anything else and used as the visual authority for the whole production. Anchoring every scene to a single master reference is the standard fix for characters changing appearance between shots.
- Identity drift (morphing)
- The gradual change of a face, body or outfit across a generated sequence — the character in minute eight no longer matches the one in minute one. Drift accumulates with generation length, which is the core reason most AI video tools cap clips at a few seconds.
- Temporal consistency
- Whether things that should stay the same actually stay the same from frame to frame and shot to shot — faces, clothing, lighting, room geometry. It is the single hardest problem in long-form AI video and the main axis on which long-form tools are judged.
- Continuity supervision
- A review pass that reads all planned scenes as one set and flags contradictions — a coat that appears in scene 2 and vanishes in scene 5, rain that stops without a cut — before rendering starts. It is the generative equivalent of a script supervisor on a film set.
- Hallucination
- Content the model invents that was never asked for: an extra hand, a person in an empty room, text that reads as gibberish. In video it is more damaging than in text because a single hallucinated frame breaks the illusion of a shot.
Glossary
Structure.
- Storyboard
- The scene-by-scene plan of a film: what happens, where, for how long, in what order. In an AI pipeline the storyboard is what turns one prompt into a list of individually renderable shots.
- Chapter planning
- Grouping a long film's scenes into a handful of acts, each with its own setup, tension and payoff, before scenes are written. Without it, sixty generated scenes read as sixty disconnected clips rather than one story.
- Scene, shot and clip
- A shot is one continuous camera take; a scene is the action happening in one place and time, which may be several shots; a clip is one output file from a generator. In AI video one scene often has to be assembled from several generated clips.
- Sub-clip
- A segment a scene is split into when it is longer than one generation can produce. Sub-clips of the same scene are kept anchored to the same reference image so the joins are invisible.
- Keyframe
- A frame used as a fixed point that generation must match — typically the first frame, and sometimes also the last. Supplying both ends lets a model generate a controlled transition between two known images.
Glossary
Rendering.
- Frame rate (fps)
- How many frames make up one second of video. It is the reason generated clip length is quantised: a model that produces a fixed number of frames at a fixed frame rate can only output specific durations.
- Frame interpolation
- Synthesising in-between frames to raise a clip's effective frame rate after generation. It makes motion look smoother without the cost of generating more frames directly.
- Render budget cap
- A hard spending ceiling attached to a single job, enforced while it runs rather than reported afterwards. Because AI video cost scales with duration, a cap is the difference between a priced product and an open-ended bill.
- Parallel rendering
- Generating many scenes at the same time instead of one after another. It is why a film with sixty scenes can finish in tens of minutes rather than hours; the limit is the provider's concurrency, not the film's length.
Glossary
Finishing.
- Colour grading
- Adjusting colour, contrast and tone across a whole film so every shot belongs to the same world. It is what separates footage that looks generated from footage that looks shot, and it has to be applied consistently or the cut falls apart.
- Film grain
- Fine texture added over the image to imitate photochemical film stock. In generated video it also serves a practical purpose: it masks the unnaturally clean surfaces that make AI footage read as synthetic.
- Text-to-speech (TTS)
- Generating spoken narration from written text. In video production the hard part is not the voice but the timing: narration has to land inside the shots it describes, which means the cut and the speech have to be planned together.
- Sidechain ducking
- Automatically lowering the music whenever narration is present, then bringing it back up. It is the standard mix technique for keeping a score audible without ever burying the voice.
- Loudness normalisation (LUFS)
- Levelling a finished mix to a target perceived loudness rather than to peak level. Platforms normalise uploads to their own target, so delivering at a sensible LUFS value prevents the platform from turning your film down for you.
- CRF (constant rate factor)
- The quality dial in common video encoders: lower numbers mean higher quality and larger files. It sets quality rather than bitrate, so a complex scene is automatically given more data than a simple one.
- H.264 and H.265
- The two dominant video codecs. H.264 plays essentially everywhere and is the safe web choice; H.265 delivers similar quality in a much smaller file but has narrower playback support, which is why films are often exported in both.
Glossary
Formats.
- Long-form AI video
- AI-generated video measured in minutes rather than seconds. It is a different problem from clip generation: it requires narrative planning, consistency across dozens of shots, and a real audio mix, none of which a single-shot model provides.
- Faceless video
- Narrated video where no presenter appears on camera — the visuals carry the story while a voiceover does the talking. It is the dominant format for anonymous YouTube channels, and it suits generated video because it sidesteps the hardest consistency problem: human faces.
See these ideas in practice in how Cinovix works, or read why long-form AI video is hard.
Free early access