AI Cinema Glossary: 24 Terms Explained for Beginners

Film writing about AI is full of borrowed technical words, and they are often used loosely. This glossary defines the ones you will meet in articles about AI cinema and AI-made footage, roughly in the order a single generated shot is made. The definitions are deliberately plain, and each one ends with the part that matters on screen.

The model and how it learns

Model

A model is a system that turns an input — a prompt, an image, a frame — into an output such as a generated shot, a label or a next word. In film writing, “model” usually means the whole generative tool, not one component inside it.

Parameters and weights

The learned numbers inside a model. Training adjusts them; inference uses them without changing them. When a model is called “large”, the count of parameters is usually what is meant.

Training

The process of adjusting a model’s weights on data, usually by comparing outputs with targets and nudging the weights to reduce error. Training is expensive and mostly happens before you ever touch the tool.

Inference

Running a trained model to get an output. Generation is inference: no learning happens, and the model’s knowledge is frozen at the point training ended.

Prompt

The instruction text — and sometimes image — that conditions a generation. A prompt steers the output, but it is not code. Identical prompts can produce different results across models and versions.

Token

A unit of text that a model processes, often a word or part of a word. Token counts matter because context limits and costs are measured in tokens.

Multimodal model

A model that handles more than one kind of input or output, such as text plus images or audio. Most current video tools are multimodal in at least this sense.

How a generated shot is built

Latent space

A compressed numerical representation of images or video. Models work in this space because it is smaller and smoother than raw pixels, and points near each other tend to look similar. A “latent walk” means moving between points to produce gradual change.

Variational autoencoder (VAE)

A pair of networks that compress an image into a latent representation and reconstruct it. In many image tools the VAE is the part that decodes a latent result back into pixels.

Diffusion model

A generative model trained to reverse a gradual noising process: it learns to remove noise step by step until an image or clip appears. The approach was set out in the paper that introduced denoising diffusion probabilistic models.

Sampling steps

The number of denoising steps a model takes. More steps can add detail, but they also add time and do not guarantee a better result.

Seed

The random number that starts the noise. Reusing a seed with the same model, settings and software can reproduce an image closely. Change the model version or platform and the result may drift.

Guidance scale

A setting that controls how strongly the model follows the prompt, often implemented as classifier-free guidance. Low values give looser, more surprising outputs; very high values tend to produce brittle or oversaturated results.

Checkpoint

A saved file of model weights, usually bundling the components needed to run a particular model. Different checkpoints of the same family can look and behave quite differently.

LoRA

Short for Low-Rank Adaptation: a small set of trainable matrices added to a frozen base model. It teaches a style, character or concept with far fewer trainable parameters than full fine-tuning.

Fine-tuning

Further training of a pre-trained model on new data. LoRA is one efficient way to do it, and “fine-tune” is often used loosely for any adaptation.

ControlNet

An add-on that conditions a diffusion model on spatial information such as a pose, a depth map or an edge image, so the output follows a specific structure instead of inventing one.

Inpainting and outpainting

Filling or replacing part of an image (inpainting), or extending it beyond its original frame (outpainting). Both are common in shot clean-up and set extension.

Making motion

Frame interpolation

Synthesising new frames between existing ones to raise the frame rate or slow footage down. Older tools used optical flow; newer ones use learned models. Interpolation can smooth motion, but it cannot recover what was never shot.

Temporal consistency

How well a sequence stays coherent from frame to frame: faces, clothing, lighting and physics. Poor temporal consistency shows up as flicker, morphing or objects that change between shots.

The output and the evidence

Hallucination

A plausible but false output: an object that appears from nowhere, a hand with too many fingers, text that almost spells something. The term describes confident nonsense, not a technical fault with a single cause.

Deepfake

Synthetic or manipulated video or audio, usually of a person’s face or voice, made with deep learning. The word is often used loosely for any convincing fake, but the useful definition names a specific technique and a specific harm.

AI upscaling

Increasing resolution with a learned model. It can sharpen real detail, but it may also invent detail that was never in the source, which is why restoration ethics care about it.

C2PA and Content Credentials

C2PA is an open technical standard for recording the origin and edit history of digital content. Content Credentials is the consumer-facing name for that metadata. It is a signed claim about provenance, not proof that the content is true.

Why this vocabulary matters

These terms are not neutral. A review that says a shot was “made with AI” tells you almost nothing: hand-drawn animation, a diffusion model and a deepfake all fit the phrase. Knowing the pipeline lets you ask sharper questions — which model, trained on what, run by whom, edited how. Those are the questions that turn a vague reaction into criticism.

For a viewing path through the films that ask the bigger questions, start with the AI-themed movies watch guide. For the practical side of checking synthetic footage, see how to spot AI-generated footage. And if you want the production mechanics rather than the vocabulary, read how AI video models work.