Glossary
The vocabulary of creative AI.
Plain definitions of the terms this work runs on, written for the people doing it rather than the people publishing papers. If a term here shows up in a job posting, we say what it means on set as well as in the model.
- Multimodal
- A model that works across more than one kind of input or output: text, image, audio, video. The word describes the model's range, not a job title.
- It is also why this society is called what it is. Film, design, image, sound, and story are separate crafts that now share one set of tools.
- Diffusion
- The generation method behind most image and video models. It starts from noise and removes it step by step until an image consistent with the prompt remains.
- Practical consequence: more steps is not automatically better, and the last few steps decide fine detail, which is where most quality settings actually matter.
- Latent space
- The compressed internal representation a model works in rather than raw pixels. Editing in latent space is why a small prompt change can restructure an entire image instead of nudging it.
- Seed
- The number that fixes a generation's randomness. Same seed, same prompt, same settings gives the same result.
- Seed discipline is the difference between a shot you can iterate on and a shot you can only re-roll. Record it like a slate.
- LoRA
- Low-Rank Adaptation. A small add-on trained on a handful of images that teaches a base model one specific character, style, or object without retraining the model.
- The standard answer to character consistency, and the reason a lead can survive a dozen shots.
- ControlNet
- A conditioning layer that constrains a generation to a structure you supply: a pose, a depth map, an edge outline. It is how a shot gets composed rather than rolled for.
- Inpainting
- Regenerating a masked region of an existing image while leaving the rest untouched. The repair pass most finished work depends on, and the one that separates a portfolio piece from a first draft.
- Context window
- How much text, image, or video a model can consider at once. A long context lets a model hold a whole script; a short one forces you to feed it a scene at a time.
- Eval
- A repeatable test that measures what a model actually does on a defined task, scored against a fixed rubric.
- Distinct from a demo, which is a selected result. Our benchmark publishes the harness with the numbers so anyone can rerun it.
- Open-weight model
- A model whose trained parameters are published, so it can be run and modified on your own hardware.
- Not the same as open source, which would also require the training data and the code. The distinction matters when a client asks where the footage was processed.
- Temporal consistency
- Whether a subject stays the same from frame to frame and shot to shot. The dominant failure mode in generated video, and the thing audiences notice first even when they cannot name it.
- Previsualisation
- Locking look, character, and framing in stills before generating any video. Video inherits every inconsistency the stills carry, so this pass decides the film.