CONTROL
Every lever between a prompt and a predictable frame: conditioning, references, guidance, negatives and the trained identity.
CFG Scale
A dial controlling how literally a diffusion model obeys the prompt, trading obedience against image quality.
Camera Control
Directing a generative video model with the vocabulary of a camera department (dolly, truck, crane, rack focus) instead of describing what the shot should feel like.
Camera Path
The defined route a virtual camera takes through a shot, specified as a movement.
Character Consistency
Keeping the same face and body recognisably identical, wardrobe included, across every shot and every session.
ControlNet
A conditioning layer that constrains a diffusion model with structural input (a pose skeleton, a depth map, an edge trace) so composition survives a change of style.
Denoising Strength
How far an image-to-image generation is allowed to move from its input, from a light retouch to a complete reimagining.
Depth Map
A greyscale image encoding how far each pixel is from the camera, used to hold a composition steady while everything else changes.
Drift
The slow, cumulative change in a subject across a set of generations (a face, a product, a colour) where no single frame looks wrong but the run does.
First-Last Frame
Handing a video model a start image and an end image and asking it to generate the motion between them.
Lip Sync
Matching a generated or existing face’s mouth movement to an audio track, and the point at which a synthetic presenter becomes convincing or does not.
LoRA
A small trained adapter that teaches a large model one style, character or object without retraining the model itself.
Master Plate
The single locked reference frame of a subject that every later generation is built from, so the subject cannot wander across a set.
Motion Brush
A control that lets you paint which regions of a still image should move and in which direction, instead of describing motion in words.
Negative Prompt
A list of things to steer away from, applied as a counterweight to the main prompt.
Pose Estimation
Extracting a figure’s joint positions as a stick-figure skeleton, then conditioning generation on it so the body does what you asked.
Prompt Engineering
Writing model instructions with the precision of a brief: what, in what order, with what constraints and what it must not do.
Prompt Weighting
Syntax for telling a model which words in a prompt carry the weight, so nobody has to repeat them and hope.
Reference Image
A picture supplied alongside the prompt to steer composition, style, colour or subject.
Reference Strength
How heavily a model is told to weigh a supplied reference image against the text prompt when the two disagree.
Regional Prompting
Applying different prompts to different areas of one frame, so elements stop bleeding into each other.
Rotoscoping
Isolating a moving subject from its background frame by frame, still the most reliable way to protect a real element inside a generated one.
Sampler
The algorithm that decides how a diffusion model steps from noise to image, trading speed against fidelity and consistency.
Seed
The number that fixes a model’s starting randomness, making a generation reproducible.
Segmentation Mask
A map marking which pixels belong to which object, used to change one element of a frame without disturbing the rest.
Set Specification
A short written document fixing the light direction, colour temperature, surfaces and lens language for a whole set, written before anything is generated.
Silhouette Overlay
A control gate that lays a finished render over its master plate at around forty per cent opacity to check the product outline has not drifted.
Style Transfer
Applying the visual character of one image or corpus to the content of another.
Trained Identity
A face trained once from a sheet of stills, which then carries across sessions, models and formats without re-uploading a reference.
Enough theory.
Bring us the launch and we will show you which of these terms is about to cost you a week.